CONTRIBUTIONS TO FINGERPRINT RECOGNITION

By

Joshua James Engelsma

A DISSERTATION

Submitted to

Michigan State University

in partial fulﬁllment of the requirements

for the degree of

Computer Science – Doctor of Philosophy

2021

ABSTRACT

CONTRIBUTIONS TO FINGERPRINT RECOGNITION

By

Joshua James Engelsma

From the early days of the mid to late nineteenth century when scientiﬁc research ﬁrst began

to focus on ﬁngerprints, to the present day ﬁngerprint recognition systems we ﬁnd deployed on

our day to day devices, the science of ﬁngerprint recognition has come a long way. In spite of this

progress, there remains challenging problems to be solved. This thesis highlights a few of these

problems, and proposes solutions to address them.

One area of further research that must be conducted on ﬁngerprint recognition systems is that

of robust, operational evaluations. In chapter two of this thesis, we show how the current practices

of using calibration patterns to evaluate ﬁngerprint readers are limited. We then propose a realistic

fake ﬁnger called the Universal Target. The Universal Target is a realistic, 3D, fake ﬁnger (or

phantom) which can be imaged by all major types of ﬁngerprint sensing technologies. We show the

entire manufacturing (molding and casting) process for fabricating the Universal Targets. Then,

we show a series of evaluations which demonstrate how the Universal Targets can be used to

operationally evaluate current commercial ﬁngerprint readers. Our Universal Target is a signiﬁcant

step forward in enabling more realistic, standardized evaluations of ﬁngerprint readers.

In our third chapter, we shift gears from improving the evaluation standards of ﬁngerprint

readers to instead focus on the security of ﬁngerprint readers. In particular, we turn our atten-

tion towards detecting fake ﬁngerprint (spoof) attacks. To do so, we open source a ﬁngerprint

reader (built from low-cost ubiquitous components), called RaspiReader. RaspiReader is a high-

resolution ﬁngerprint reader customized with both direct-view imaging and FTIR imaging in order

to better detect ﬁngerprint spoofs. We show through a number of experiments that RaspiReader en-

ables state-of-the-art ﬁngerprint spoof detection accuracy. We also demonstrate that RaspiReader

enables better generalization to what are known as “unseen attacks” (those attacks which were

not seen during training of the spoof detector). Finally, we show that ﬁngerprints captured by

RaspiReader are completely compatible with images captured by legacy ﬁngerprint readers for

matching.

In chapter four, we move on to propose a major improvement to the ﬁngerprint feature extrac-

tion and matching sub-modules of ﬁngerprint recognition systems. In particular, we propose a deep

network, called DeepPrint, to extract a 200 byte ﬁxed-length ﬁngerprint representation. While

prevailing ﬁngerprint matchers primarily utilize minutiae points and expensive graph matching

algorithms for comparison, two DeepPrint representations can be compared with only 192 mul-

tiplications and 191 additions. This is extremely useful for large scale search where potentially

billions of pairwise ﬁngerprint comparisons must be made. The DeepPrint representation also en-

ables practical encrypted matching using a fully homomorphic encryption scheme. This enables

better protection of the ﬁngerprint templates which are stored in the database. While discriminative

ﬁxed-length representations are available for both face and iris recognition, such a representation

has eluded ﬁngerprint recognition. This chapter aims to ﬁll that void.

Finally, we conclude our thesis by working to extend ﬁngerprint recognition to all ages. While

current ﬁngerprint recognition systems are being used by billions of teenagers and adults around

the world, the youngest people among us remain disenfranchised. In particular, modern day ﬁn-

gerprint recognition systems do not work well on infants and young children. In this penultimate

chapter, we aim to rectify this major shortcoming. To that end, we prototype a high-resolution

(1900 ppi) infant ﬁngerprint reader. Then, we track and ﬁngerprint 315 infants (under the age of 3

months at enrollment) at the Dayalbagh Children’s Hospital in Agra India over the course of 1 year

(4 different sessions). To match the infant ﬁngerprints, we develop our own high-resolution infant

ﬁngerprint matcher. Our experimental results demonstrate signiﬁcant promise for the extension of

ﬁngerprint recognition to all ages. This work has the potential for major global good as all young

infants and children could be given a veriﬁable digital identity for better vaccination tracking as a

child and for government beneﬁts and assistance as an adult.

In summary, this thesis makes major contributions to the entire end-to-end ﬁngerprint recogni-

tion system and extends its use case to all ages.

Copyright by
JOSHUA JAMES ENGELSMA
2021

To my family, my friends, and my loving wife Kate

v

ACKNOWLEDGMENTS

“Thy word is a lamp unto my feet, and a light unto my path.”

Psalm 119:105

As I sit down to write this thesis and anticipate the end of my time as a PhD student, I sense a

feeling of nostalgia already sweeping over me. Perhaps it is too cliche to say, but it nevertheless

true, that the past ﬁve years have ﬂown by. First and foremost, I am grateful to God for giving me

the health, mind, and strength to complete this thesis. I am also incredibly grateful for all of the

friends, colleagues, and mentors who have instructed me, guided me, and learned along side of me

during this journey. Although there are too many to adequately address in this conﬁned space, I

want to call out a few individuals who have been particularly special to me throughout the PhD.

I want to begin by thanking my advisor Dr. Jain. When I arrived in East Lansing ﬁve years

ago, I was an eager but raw PhD student that lacked any signiﬁcant research training. In spite

of this, Dr. Jain took a chance on me in bringing me into his lab and patiently teaching me how

to rigorously conduct and disseminate scientiﬁc research. When I look back to what I was as a

researcher when I started to where I am today, I can see remarkable growth. I attribute this to the

guidance and example of Dr. Jain. Dr. Jain has told me several times that “being a PhD advisor is

like polishing a cloudy diamond”. Thank you Dr. Jain for your hard work in polishing this cloudy

diamond. Thank you also for your friendship and timely counsel on matters outside of research

throughout the PhD.

Next, I want to thank my wife and best friend Kate. Kate has selﬂessly moved away from

family and friends to facilitate my education and continues to tirelessly support me and our two

children, Vivian and Abigail, in so many ways. During the most difﬁcult times of the PhD, her

support and encouragement kept me moving forward. She deserves the credit for this PhD as

much as I do. Kate, you, Vivian and Abigail bring me unspeakable joy in life.

Thank you to all of my family. I am especially thankful to my parents for raising me, providing

for me, and encouraging me to chase big dreams. Thank you all for the numerous one hour trips

vi

from Grand Rapids to Lansing for family fellowship. Thank you Charlie and James for meeting

me out here on occasion for much needed study reprieve via bike rides and video games.

I am grateful for all of my fellow PRIPies, especially the members of my cohort (Debayan,

Yichun, and Sixue). I want to especially thank Deb for his close friendship throughout the PhD.

I have many fond memories of playing pickup soccer together, attending Hindi musical events,

debating about current events over a beer, and traveling to India for research on ﬁve separate oc-

casions. Thank you Sixue and Yichun for all of our conversations in the lab, your encouragements

throughout, and for your willingness to collaborate with me on coursework and research. Thank

you also to Steven Grosz, Vishesh, Divyansh, Inci, Tarang, Cori, and Sunpreet, current and former

PRIPies with whom I had the pleasure of studying with. Thank you also to non-PRIPies Yaojie,

Amin, Joel, Renu, Cunjian, Steven Hoffman, and Adam Terwilliger for your friendship. A special

thanks to Kai Cao, a PRIP post-doc whose mentorship has been invaluable to me and who has

become a close friend. Finally, thanks to all of those involved on the various IARPA ODIN GCTs.

I have many fun memories of our travels to the Baltimore area together.

Thank you to my committee, Dr. Xiaoming Liu, Dr. Arun Ross, and Dr. Mi Zhang for

your feedback and comments on this thesis. A special thanks to Dr. Liu and Dr. Ross for your

feedback throughout the IARPA ODIN project and your general life advice. Thank you Dr. Vishnu

Boddeti for your collaboration with me on encrypted matching. Thank you Brenda Hodge, Steven

Smith, and Amy King for your administrative assistance and your patience with my tardy travel

reimbursement forms. A special thanks to Christopher Perry, for support on the IARPA ODIN

project and for being a great friend, Brian Wright for his assistance with the ECE department

3D printer, and Dr. Anjoo Bhatnagar and Dr. Prem Sudhish for their facilitating our infant data

collections in India. Finally, I would like the thank NIST (especially Nick Paulter), and IARPA for

their support of the research contained in this thesis. There are of course many others who have

impacted me along the way and I thank all of you.

vii

TABLE OF CONTENTS

LIST OF TABLES .

.

LIST OF FIGURES .

.

.

.

.

.

.

LIST OF ALGORITHMS .

.

.

.

.

.

.

.

.

.

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

xi

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. xiv

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .xxiv

Chapter 1

.

.

.

.

.

.

Introduction .

.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
1.1 History of Fingerprint Recognition . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Major Applications .
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 Pipeline of Fingerprint Recognition Systems . . . . . . . . . . . . . . . . . . . . .

1
2
5
7
1.3.1
Fingerprint Readers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3.2
Feature Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.3.3 Template Database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.3.4 Matching .
. 17
Fingerprint Reader Evaluations . . . . . . . . . . . . . . . . . . . . . . . . 17
Fingerprint Presentation Attack Detection . . . . . . . . . . . . . . . . . . 18
Fixed-Length Fingerprint Representations . . . . . . . . . . . . . . . . . . 20
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
Infant Fingerprints
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. 22

.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

.
1.4 Remaining Challenges

1.4.1
1.4.2
1.4.3
1.4.4

1.5 Contributions

.
.

.

.

.

.

.

.

.

Chapter 2

.

.

.

.

. .

2.1

2.3 Casting .

2.2.1 Mold Fabrication .
2.2.2

Universal 3D Wearable Fingerprint Targets: Advancing Fingerprint
Reader Evaluations
.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
Introduction .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.1.1
3D Fingerprint Targets . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.1.2
Fingerprint Reader Interoperability . . . . . . . . . . . . . . . . . . . . . . 29
2.1.3 Universal 3D Fingerprint Targets . . . . . . . . . . . . . . . . . . . . . . . 30
2.2 Mold & Scaffold Fabrication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
Scaffolding Fabrication . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.3.1 Material Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.3.2 Material Fabrication and Casting Procedure . . . . . . . . . . . . . . . . . 39
2.3.3 Material Characterization . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.4 Target Fidelity and Reproducibility . . . . . . . . . . . . . . . . . . . . . . . . . . 41
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
.
. . . . . . . . . . . . . . . . 51
. . . . . . . . . . . . . . . . 54
. . . . . . . . . . . . . . . . . . . . . 55

2.5.1 Evaluating Readers with Calibration Patterns
2.5.2 Evaluating Readers with Fingerprint Patterns
2.5.3 Reader Interoperability Evaluations

2.4.1
.
2.4.2 Reproducibility .
.

2.5 Experiments .

Fidelity .

. .

.

.

.

.

.

.

.

.

.

.

.

.

.

.

viii

2.6 Summary . .
. .
2.7 Acknowledgment

.
.

.
.

.
.

.
.

.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

Chapter 3

.

.

.

.

.

. .

3.4.1

Introduction .

RaspiReader: Open Source Fingerprint Reader . . . . . . . . . . . . . . . 59
3.1
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.2 RaspiReader Construction and Calibration . . . . . . . . . . . . . . . . . . . . . . 68
3.2.1 Camera Placement
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.2.2 Case Fabrication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
Image Acquisition Hardware and Software . . . . . . . . . . . . . . . . . . 70
3.2.3
3.2.4
Fingerprint Image Processing . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.3 Live and Spoof Fingerprint Database Construction . . . . . . . . . . . . . . . . . . 74
. . . . . . . . . . . . . . . . . . . . . . 76
3.4 Spoof Detection Experiments and Results
Spoof Detection Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
. . . . . . . . . . . . . . . 77
3.4.1.1
3.4.1.2 CLBP Features From RaspiReader Images
. . . . . . . . . . . . 78
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.4.1.3 MobileNet
Spoof Detection Results
. . . . . . . . . . . . . . . . . . . . . . . . . . . 81
3.4.2.1 Known-Material Scenarios . . . . . . . . . . . . . . . . . . . . . 81
. . . . . . . . . . . . . . . . . . . . . 83
3.4.2.2 Cross-Material Scenarios
3.5
. 87
3.6 Computational Resources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
3.7 Summary .
. 89
.
.
. 89
3.8 Acknowledgment

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Interoperability of RaspiReader . . . . . . . . . . . . . . . . . . . . . . . . . . .

LBP Features From COT SA Images

.
. .

3.4.2

. .

.
.

.

Chapter 4

.
.
.
.
.

.
.
.
.
.
.
. .
. .

4.4 DeepPrint Matching .

4.1
4.2 Prior Work .
4.3 DeepPrint
.

Learning a Fixed Length Fingerprint Representation . . . . . . . . . . . . 91
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
.
.
. .
Introduction .
.
. 95
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
. .
. .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.3.1 Overview .
4.3.2 Alignment .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
4.3.3 Minutiae Map Domain Knowledge . . . . . . . . . . . . . . . . . . . . . . 103
4.3.4 Multi-Task Architecture
. . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.3.5 Template Compression . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
Fusion of DeepPrint Score with Minutiae Score . . . . . . . . . . . . . . . 109
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
4.5.1
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
4.5.2 Two-stage DeepPrint Search . . . . . . . . . . . . . . . . . . . . . . . . . 112
4.6 Secure DeepPrint Matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
4.7 Datasets . .
4.7.1 NIST SD4 & NIST SD14 . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
4.7.2
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
.
Search (1:N Comparison) . . . . . . . . . . . . . . . . . . . . . . . . . . . 115

4.8 COTS Matchers .
.
4.9 Benchmark Evaluations .

.
Faster Search .

4.5 DeepPrint Search .

FVC 2004 DB1 A .

4.9.1

4.4.1

. .

. .

.
.

.

.

.

.

.

.

.

.

.

.

ix

.

. .

4.9.2.1
4.9.2.2

4.9.2 Authentication .

4.10 Large Scale Search .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. 117
Fusion with Minutiae-Matchers . . . . . . . . . . . . . . . . . . 118
Secure Authentication . . . . . . . . . . . . . . . . . . . . . .
. 119
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
. 120
. 120
. 121
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
4.11 Ablation Study . .
4.12 Interpretability .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
. .
4.13 Computational Resources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
4.14 Summary . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125

4.10.1 DeepPrint Search .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.10.2 Minutiae Re-ranking . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.10.3 Product Quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. .

.
.

.
.

.
.

.

.

.

.

.

Chapter 5

.

.

.

.

.

.

.

.

. .

. .

5.1

Introduction .
5.1.1

5.2 Related Work .
5.3 High-Resolution Fingerprint Reader
5.4 Longitudinal Fingerprint Dataset
5.5

. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Infant Fingerprint Recognition . . . . . . . . . . . . . . . . . . . . . . . . 126
. 127
Fingerprints for Infant-ID . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
. 131
. . . . . . . . . . . . . . . . . . . . . . . . . 134
. . . . . . . . . . . . . . . . . . . . . . . . . . . 136
Infant Fingerprint Matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. 139
5.5.1 Minutiae Matcher . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
5.5.2 Minutiae Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
5.5.2.1 Manual Minutiae Markup for Training . . . . . . . . . . . . . . 143
5.5.2.2 Minutiae Aging . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
5.5.2.3 Minutiae Match Score . . . . . . . . . . . . . . . . . . . . . . . 147
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
5.5.3 Texture Matcher .
5.5.4 Latent Fingerprint Matcher . . . . . . . . . . . . . . . . . . . . . . . . . . 148
Enhancement . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
5.5.4.1
5.5.4.2
Image Aging . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
Final Match Score . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
. 150
. . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
5.6.1 Experimental Protocol
Infant Authentication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
5.6.2
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
Infant Search .
5.6.3
5.6.4 Ablations .
.
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
5.6.5 Longitudinal Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
. 160

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

5.6 Experimental Results .

5.7 Summary .

5.5.5

. .

. .

.
.

.

.

.

.

.

.

Chapter 6

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
.
. 162
6.1 Contributions
6.2 Suggestions for Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164

. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Summary .
.

.
.

.
.

.
.

.
.

.

BIBLIOGRAPHY .

.

.

.

.

.

.

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165

x

LIST OF TABLES

Table 2.1 Properties of the Human Finger [42, 51, 52]

. . . . . . . . . . . . . . . . . . 27

Table 2.2 Properties of published 3D Printed Targets [4, 5, 9]

. . . . . . . . . . . . . . 29

Table 2.3 Properties of our 3D Casted Targets . . . . . . . . . . . . . . . . . . . . . . 30

Table 2.4 Average point-to-point ridge distances observed on universal ﬁngerprint tar-
gets, measured using the Keyence Optical Microscope at 50X and 100X magniﬁca-
tion. The expected point-to-point ridge distance is 0.508 mm. (standard deviation
is recorded in parenthesis).
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. 44

Table 2.5 Universal Fingerprint Target Similarity Scores (SD4 ﬁngerprint image vs.

corresponding target image). Proposed Targets.

. . . . . . . . . . . . . . . . . . . 46

Table 2.6

3D Printed Target Similarity Scores (SD4 ﬁngerprint image vs. correspond-

ing target image). Targets from [4–6].

. . . . . . . . . . . . . . . . . . . . . . . . 47

Table 2.7 Universal Fingerprint Target Genuine Similarity Scores (SD4 Fingerprint
Image vs. Corresponding Target Image) Mean and (Standard Deviation) of Scores
for 10 Impressions are Reported . . . . . . . . . . . . . . . . . . . . . . . . . .

. 48

Table 2.8 Speciﬁcations of the Fingerprint Readers Used in Our Experiments . . . . . . 49

Table 2.9 Mean (µ) and std. deviation (σ) of center-to-center ridge spacings (in pixels)
on images acquired from 3 universal ﬁngerprint targets. The expected ridge spacing
is 9.8 pixels. .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

.

.

.

.

.

.

Table 2.10 Mean (µ) and std. deviation (σ) of center-to-center ridge spacings (in pixels)
on images acquired from 3D printed targets. The expected ridge spacing is 8.28
pixels.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

. .

. .

. .

.

.

.

.

Table 2.11 Mean (µ) and std. deviation (σ) of center-to-center ridge spacings (in pixels)
on images acquired from 6 universal ﬁngerprint targets. Expected ridge spacing (in
pixels) for each target is reported in parenthesis

. . . . . . . . . . . . . . . . . . . 53

Table 2.12 Mean (µ) and std. deviation (σ) of center-to-center ridge spacings (in pixels)
on images acquired from 3D printed ﬁngerprint targets. Expected ridge spacing (in
pixels) for each target is reported in parenthesis

. . . . . . . . . . . . . . . . . . . 55

xi

Table 2.13 Genuine and Imposter Score1 Statistics and Matching Performance Measures
when Comparing Fingerprint Images Acquired from Different Types of Fingerprint
Readers. Mean of Genuine Scores (µG), Mean of Imposter Scores (µI), True
Accept Rate (TAR) and False Accept Rate (FAR2) are Reported.

. . . . . . . . . . 56

Table 3.1 Primary Components Used to Construct RaspiReader. Total Cost is $175.20 . 67

Table 3.2 Summary of Spoof Fingerprints Collected . . . . . . . . . . . . . . . . . . . 75

Table 3.3 Summary of Live Finger Data Collected . . . . . . . . . . . . . . . . . . . . 75

Table 3.4 Textural Features and Known Testing Materials . . . . . . . . . . . . . . .

. 82

Table 3.5 MobileNet and Known Testing Materials

. . . . . . . . . . . . . . . . . . . 83

Table 3.6 Textural Features and Cross-Material Testing . . . . . . . . . . . . . . . . . 84

Table 3.7 MobileNet and Cross-Material Testing . . . . . . . . . . . . . . . . . . . . . 85

Table 3.8 Lumidigm Spoof Detection Accuracy . . . . . . . . . . . . . . . . . . . . . 85

Table 3.9 Fingerprint Matching Results . . . . . . . . . . . . . . . . . . . . . . . . .

. 88

Table 4.1 Comparison of variable length minutiae representation with ﬁxed-length

DeepPrint representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

Table 4.2 Published Studies on Fixed-Length Fingerprint Representations

. . . . . . . 96

Table 4.3 Localization Network Architecture . . . . . . . . . . . . . . . . . . . . . . . 102

Table 4.4 Effect of Compression on Accuracy . . . . . . . . . . . . . . . . . . . . . . 108

Table 4.5 Benchmarking DeepPrint Search Accuracy against Fixed-Length Represen-

tations in the Literature and COTS . . . . . . . . . . . . . . . . . . . . . . . . . . 116

Table 4.6 Authentication Accuracy (FVC 2004 DB1 A) . . . . . . . . . . . . . . . . . 117

Table 4.7 Authentication Accuracy (Rolled-Fingerprints)

. . . . . . . . . . . . . . . . 118

Table 4.8 Encrypted Authentication using DeepPrint Representation . . . . . . . . . . 119

Table 4.9 DeepPrint + Minutiae Re-ranking Search Accuracy (1.1 million background) 121

Table 4.10 DeepPrint + PQ: Search Accuracy (1.1 million background)

. . . . . . . . . 121

xii

Table 4.11 DeepPrint Representation Comparison . . . . . . . . . . . . . . . . . . . . . 122

Table 4.12 DeepPrint Ablation Study . . . . . . . . . . . . . . . . . . . . . . . . . . . 122

Table 5.1 Related work on child and infant ﬁngerprint recognition.

. . . . . . . . . . . 130

Table 5.2

Infant Longitudinal Fingerprint Dataset Statistics . . . . . . . . . . . . . . . 138

Table 5.3 Minutiae Extraction Network . . . . . . . . . . . . . . . . . . . . . . . . . . 142

Table 5.4

Infant Authentication Accuracy3 (0 − 3 months at enrollment with 3 month

time lapse between enrollment and authentication) . . . . . . . . . . . . . . . . . . 151

Table 5.5

Infant Search Accuracy3 (0−3 months at enrollment with 3 month time lapse
between enrollment and search) . . . . . . . . . . . . . . . . . . . . . . . . . . .
Table 5.6 Ablated Infant Authentication Accuracy4 (0 − 3 months at enrollment with

. 152

3 month time lapse between enrollment and authentication) . . . . . . . . . . . . . 154

Table 5.7 Ablated Veriﬁnger Performance . . . . . . . . . . . . . . . . . . . . . . . . 155

Table 5.8 Ablated COTS Latent Matcher (LM) Performance

. . . . . . . . . . . . . . 156

Table 5.9 Ablated DeepPrint Performance . . . . . . . . . . . . . . . . . . . . . . . . 156

Table 5.10 Ablated Fingerprint Reader Authentication Results . . . . . . . . . . . . . . 158

Table 5.11 Longitudinal Search Results . . . . . . . . . . . . . . . . . . . . . . . . . . 159

Table 5.12 Longitudinal Authentication Results . . . . . . . . . . . . . . . . . . . . . . 159

xiii

LIST OF FIGURES

Figure 1.1 One of the earliest recorded uses of ﬁngerprints include chinese clay seals

which were used to sign business transactions. Image retrieved from [63].

. . . . .

Figure 1.2 Example ﬁngerprints of each of the ﬁve major ﬁngerprint classes deﬁned by

Sir Edward Henry. Images retrieved from the NIST SD4 database [129].

. . . . . .

Figure 1.3 Various applications of ﬁngerprint recognition.

(a) An example of a la-
tent ﬁngerprint left behind on a dollar bill, which could be subsequently used to
search a database of known criminals; (b) a woman has her ﬁngerprints taken at
US Customs (OBIM system) prior to entry into the country; (c) An Indian citizen
is authenticated by the Aadhaar system; (d) a mobile phone is unlocked, bypass-
ing the need for a password or key-code for access. Images retrieved from Google
Images.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. .

. .

. .

.

.

.

Figure 1.4 Enrollment Phase. A ﬁngerprint is captured by the reader and transferred to
the feature extractor where minutiae and other salient, discriminative features are
extracted and packed into a template. The extracted template is then stored in the
database. .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

.

.

.

.

.

.

.

.

.

Figure 1.5 Fingerprint Authentication Schematic. During authentication, a 1:1 match
is conducted between a newly extracted template, and a template already stored in
the database. In this scenario, we are answering the question, “Is this person a
match to the speciﬁed enrollment template?” . . . . . . . . . . . . . . . . . . . . .

Figure 1.6 Fingerprint Search Schematic. During search, N matches are conducted
between a newly extracted template (probe or query), and N templates already
stored in the database. The matcher returns a ranked list of the candidates most
similar to the query. In this scenario, we are answering the question, “Who is this
person?” .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

.

.

.

.

.

.

.

.

2

3

5

9

9

9

Figure 1.7 Fingerprints captured using ink and card stock paper. The ﬁngerprints in
the top row are rolled ﬁngerprints, whereas the ﬁngerprints in the bottom row are
slap/plain ﬁngerprints. Image reproduced from [88]. . . . . . . . . . . . . . . . . . 10

Figure 1.8 Examples of different

readers
(Idemia https://www.idemia.com/, HIDGlobal https://www.hidglobal.com/). Im-
ages retrieved from Google Images.

types of optical-based ﬁngerprint

. . . . . . . . . . . . . . . . . . . . . . . . . 11

Figure 1.9 Examples of different types of solid-state readers (both capacitive) and an
in-display, mobile-phone, ultrasound reader. Images retrieved from Google Images.

12

xiv

Figure 1.10 The most popular ﬁngerprint representation consists of (a) global level-1
features (ridge ﬂow, core, and delta) and (b) local level-2 features, called minutiae
points, together with their descriptors (e.g., texture in local minutiae neighbor-
hoods). The ﬁngerprint image illustrated here is a rolled impression from the NIST
SD4 database [129]. The number of minutiae in NIST4 rolled ﬁngerprint images
range all the way from 12 to 196. . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

Figure 1.11 Examples of Level-2 features (two types of minutiae) and Level-3 features
. . . . . . . . . . . . . . . . . . . . . . . . .

(sweat pores and incipient ridges).

. 15

Figure 1.12 Example of minutiae match between two ﬁngerprint impressions of the
same ﬁnger. This example highlights the difﬁculty of minutiae matching given
a poor quality enrollment image (left) which has many missing minutiae. Despite
this difﬁculty, a COTS minutiae matcher is able to correctly match these two ﬁnger-
prints with a score of 150, well above the score threshold of 69 @ F AR = 0.01%.
Fingerprints retrieved from the FVC 2004 DB1 A database [109] . . . . . . . . . . 16

Figure 1.13 Examples of existing ﬁngerprint reader calibration targets. These targets
are useful for white-box testing ﬁngerprint readers, ensuring that they meet certain
quantitative imaging thresholds, however, they are very dissimilar from human ﬁn-
gers. As such they are not useful for realistic operational evaluations of ﬁngerprint
readers.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

. .

. .

. .

.

.

.

Figure 1.14 Examples of ﬁngerprint spoofs made from different materials. The material
variety demonstrates why it is difﬁcult to develop a spoof detector which general-
. . . . . . . . . . . . . . . . . . . . . . . . .
izes well across all material types.

. 19

Figure 1.15 Failures of a state-of-the-art COTS minutiae-based matcher (minutiae an-
notated with COTS). The genuine pair (two impressions from the same ﬁnger) in
(a) was falsely rejected at 0.1% FAR (score of 9) due to heavy non-linear distortion
and moist ﬁngers. The imposter pair (impressions from two different ﬁngers) in
(b) was falsely accepted at 0.1% FAR (score of 38) due to the similar minutiae dis-
tribution in these two ﬁngerprint images (the score threshold for COTS A @ FAR
= 0.1% is 34). These slap ﬁngerprint impressions come from public domain FVC
2004 DB1 A database [109]. The number of minutiae in FVC 2004 DB1 A images
range from 11 to 87. .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

.

.

Figure 1.16 Examples of low quality infant ﬁngerprints. These examples demonstrate
the difﬁculty in using automated ﬁngerprint recognition systems for infant recogni-
tion. Note the small inter-ridge spacing, debris, motion blur, and moisture through-
out the different impressions. These images were captured when the infant’s were
2 months old.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

.

.

.

.

.

.

xv

Figure 2.1 A Universal 3D Fingerprint Target fabricated in (a) can be imaged by a va-
riety of popular ﬁngerprint readers (contact-optical, contactless-optical, and capac-
itive) shown in (b). The sensed images of the 3D ﬁngerprint target in (a) are shown
in (c). This demonstrates that our targets are appropriate for ﬁngerprint reader
interoperability evaluation studies. Similarity scores for each sensed ﬁngerprint
image (with the 2D mapped target image) are displayed below each ﬁngerprint im-
age in (c). Veriﬁnger 6.3 SDK was used for generating similarity scores. The score
threshold at 0.01 % FAR is 33.
. . . . . . . . . . . . . . . . . . . . . . . . . . .

. 26

Figure 2.2 Example phantom of a human hand [116] used in the medical domain.

. . . 27

Figure 2.3 High ﬁdelity, wearable, 3D ﬁngerprint targets.

(a) 3D ﬁngerprint target
printed using TangoBlackPlus FLX980 [5], (b) 3D ﬁngerprint target printed using
TangoPlus FLX 930 [4], (c) 3D ﬁngerprint target printed using TangoBlackPlus
FLX980 and then sputter coated with 30 nm titanium + 300 nm of gold [6], (d) our
casted 3D ﬁngerprint target using a mixture of PDMS (Polydimethylsiloxane) and
Pantone 488C color pigment [157] [40], and (e) our casted universal 3D ﬁngerprint
target using a mixture of conductive PDMS, silicone thinner, and Pantone 488C
color pigment [157] [154] [158]. 3D targets in (a), (b), and (c) were printed on a
high resolution 3D printer (Stratasys Objet350 Connex). . . . . . . . . . . . . . .

. 28

Figure 2.4 System block diagram of the proposed molding and casting process for
making 3D targets.
(a) A 3D negative mold (of a 2D ﬁngerprint image) and a
supporting scaffolding system (necessary for making the ﬁngerprint target wear-
able) are electronically fabricated; (b) 3D electronic models are manufactured by
3D printing and chemical cleaning; (c) conductive silicone, silicone thinner, and
human colored dye are mechanically mixed to produce a casting material with
similar conductive, mechanical, and optical properties to the human skin; (d) the
material fabricated in (c) is cast into the mold and scaffolding system; (d) vacuum
degassing ensures that air bubbles are removed from the casted material; (e) wear-
able ﬁngerprint targets are extracted 72 hours after pouring the casting material; (f)
the wearable, 3D ﬁngerprint target is used for ﬁngerprint reader evaluations.

. . . . 31

Figure 2.5 Process ﬂow for fabricating electronic 3D ﬁngerprint mold, M . . . . . . . 32

Figure 2.6

(a) High ﬁdelity 3D printed ﬁngerprint mold M. (b) View of ﬁngerprint
engraving on M at 20X magniﬁcation. The magniﬁed image in (b) shows that
all the friction ridge patterns are clearly present in the mold M. These friction
ridge patterns are inverted, since negative molds are necessary to produce positive
ﬁngerprint targets (Fig 7 (c)).

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

Figure 2.7

3D wearable Universal Fingerprint Target (a) front view, (b) rear view, and

(c) view of the Universal Fingerprint Target ridges at 20X magniﬁcation. . . . . . . 36

xvi

Figure 2.8 Fabricating scaffolding F using the dimensions of the mold, M. (a) scaf-
folding framework F is electronically modeled; (b) the electronic scaffolding sys-
tem is physically generated in acrylonitrile butadiene styrene (ABS) using a high
resolution 3D printer. Using F in conjunction with M, 3D wearable ﬁngerprint
targets T are repeatably produced.

. . . . . . . . . . . . . . . . . . . . . . . . . . 37

Figure 2.9 Fingerprint impressions captured from targets lacking proper mechanical
characteristics. Notice (a) the presence of aberrations resulting from excessive
elasticity in the target and (b) partial impression due to excessive hardness of the
target.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

. .

. .

. .

.

.

.

.

Figure 2.10 Comparison of the Universal Fingerprint Target material spectrogram to a
range of spectrograms obtained by NIST from 51 human subjects. We plotted the
range using estimated data points from the ﬁgure in [30].
. . . . . . . . . . . . .

. 41

Figure 2.11 Images of the universal ﬁngerprint target (mapped with circular sine grat-
ings) captured using a Keyence optical microscope [98]. Point-to-point ridge dis-
tances are measured. (a) Image at 50X magniﬁcation and annotated with 20 point-
to-point distances. (b) Image at 100X magniﬁcation and annotated with 10 point-
to-point distances. (c) 3-D image generated by the microscope which qualitatively
illustrates the uniformity in ridge height of the circular gratings on the universal ﬁn-
gerprint target. The granular texture in (a), (b), and (c) is evidence of the aluminum
coated silver particles mixed into the universal ﬁngerprint target which allows the
target to be imaged by capacitive ﬁngerprint readers.

. . . . . . . . . . . . . . . . 42

Figure 2.12 Comparing the source ﬁngerprint image to the image of the correspond-
ing universal ﬁngerprint target. (a) NIST SD4 S0083 rolled ﬁngerprint image is
compared to (b) a universal ﬁngerprint target image; (b) is fabricated using (a)
and imaged using an Appendix F certiﬁed, optical, 500 ppi ﬁngerprint reader. A
similarity score of 608 is computed between (a) and (b) using Veriﬁnger 6.3 SDK
(threshold is 33 at FAR=0.01 %). The minutia points in correspondence between
(a) and (b) are shown.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

.

.

. 45

Figure 2.13 Example ﬁngerprint impressions from 6 universal ﬁngerprint targets (one

per column) on 3 types of ﬁngerprint readers.

. . . . . . . . . . . . . . . . . . . . 50

Figure 3.1 Prototype of RaspiReader: two ﬁngerprint images (b, (i)) and (b, (ii)) of
the input ﬁnger (a) are captured. The raw direct image (b, (i)) and the raw, high
contrast FTIR image (b, (ii)) both contain useful information for spoof detection.
Following the use of (b, (ii)) for spoof detection, image calibration and processing
are performed on the raw FTIR image to output a high quality, 500 ppi ﬁngerprint
for matching (b, (iii)). The dimensions of the RaspiReader shown in (a) are 100
mm x 100 mm x 105 mm (about the size of a 4 inch cube).

. . . . . . . . . . . . . 60

xvii

Figure 3.2 Fingerprint images acquired using the RaspiReader.

Images in (a) were
collected from a live ﬁnger.
Images in (b) were collected from a spoof ﬁnger.
Using features extracted from both raw image outputs ((i), direct) and ((ii), FTIR)
of the RaspiReader, our spoof detectors are better able to discriminate between live
ﬁngers and spoof ﬁngers. The raw FTIR image output of the RaspiReader (ii) can
be post processed (after spoof detection) to output images suitable for ﬁngerprint
matching. Images in (c) were acquired from the same live ﬁnger (a) and spoof
ﬁnger (b) on a commercial off-the-shelf (COTS) 500 ppi optical reader. The close
similarity between the two images in (c) qualitatively illustrates why current spoof
detectors are limited by the low information content, processed ﬁngerprint images
(c, (iii)) output by COTS readers.

. . . . . . . . . . . . . . . . . . . . . . . . . . 61

Figure 3.3 Example spoof ﬁngers and live ﬁngers in our database. (a) Spoof ﬁngers
and (b) live ﬁngers used to acquire both spoof ﬁngerprint impressions and live
ﬁngerprint impressions for conducting the experiments reported in this thesis. The
spoofs in (a) and the live ﬁngers in (b) are not in 1-to-1 correspondence.

. . . . . . 62

Figure 3.4 Schematic illustrating RaspiReader functionality.

Incoming white light
from three LEDs enters the prism. Camera 2 receives light rays reﬂected from
the ﬁngerprint ridges only (light rays are not reﬂected back from the ﬁngerprint
valleys due to total internal reﬂection (TIR)). This image from Camera 2, with
high contrast between ridges and valleys can be used for both spoof detection and
ﬁngerprint matching. Camera 1 receives light rays reﬂected from both the ridges
and valleys. This image from Camera 1 provides complementary information for
spoof detection.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

.

.

.

.

.

Figure 3.5 Electronic CAD model of the RaspiReader case. The camera and LED
mounts are positioned at the necessary angles and distance to the glass prism, mak-
ing the reproduction of RaspiReader as simple as 3D printing the open-sourced
STL ﬁles.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

. .

. .

. .

.

.

Figure 3.6 Processing a RaspiReader raw FTIR ﬁngerprint image into a 500 ppi ﬁnger-
print image compatible for matching with existing COTS ﬁngerprint readers. (a)
The RGB FTIR image is ﬁrst converted to grayscale. (b) Histogram equalization is
performed to enhance the contrast between the ﬁngerprint ridges and valleys. (c)
The ﬁngerprint is negated so that the ridges appear dark, and the valleys appear
white. (d), (f) Calibration (estimated using the checkerboard calibration pattern
in (e)) is applied to frontalize the ﬁngerprint image to the image plane and down
sample (by averaging neighborhood pixels) to 500 ppi in both the x and y axis.

. . 71

xviii

Figure 3.7 Acquiring Image Transformation Parameters. A 2D printed checkerboard
pattern (a) is imaged by the RaspiReader (b). Corresponding points between the
frontalized checkerboard pattern (a) and the distorted checkerboard pattern (b) are
deﬁned so that perspective transformation parameters can be estimated to map (b)
into (c). These transformation parameters are subsequently used to frontalize ﬁn-
gerprint images acquired by RaspiReader for the purpose of ﬁngerprint matching.
The checkerboard imaged in (b) is also used to acquire the native resolution of
RaspiReader in order to scale matching images to 500 ppi in both the x and y axis
as shown in (c).

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

.

.

.

.

.

Figure 3.8 Native resolution (ppi) in (a) x-axis and (b) y-axis over the raw FTIR
RaspiReader image. As is normal, native resolution changes across the image be-
cause the right side of the image is closer to the camera than the left side. . . . . .

. 74

Figure 3.9 Failure to Capture.

Several spoofs are unable to be imaged by the
RaspiReader due to their dissimilarity in color. In particular, because spoofs in
(a) and (b) are black, all light rays will be absorbed preventing light rays from re-
ﬂecting back to the FTIR imaging sensor. In (c), the dark blue color again prevents
enough light from reﬂecting back to the camera. (a) and (b) are both ecoﬂex spoofs
coated with two different conductive coatings. (c) is a blue crayola model magic
spoof attack.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

.

.

.

.

.

.

Figure 4.1 Flow diagram of DeepPrint: (i) a query ﬁngerprint is aligned via a Local-
ization Network which has been trained end-to-end with the Base-Network and
Feature Extraction Networks (no reference points are needed for alignment); (ii)
the aligned ﬁngerprint proceeds to the Base-Network which is followed by two
branches; (iii) the ﬁrst branch extracts a 96-dimensional texture-based representa-
tion; (iv) the second branch extracts a 96-dimensional minutiae-based representa-
tion, guided by a side-task of minutiae detection (via a minutiae map which does
not have to be extracted during testing); (v) the texture-based representation and
minutiae-based representation are concatenated into a 192-dimensional represen-
tation of 768 bytes (192 features and 4 bytes per ﬂoat). The 768 byte template
is compressed into a 200 byte ﬁxed-length representation by truncating ﬂoating
point value features into integer value features, and saving the scaling and shifting
values (8 bytes) used to truncate from ﬂoating point values to integers. The 200
byte DeepPrint representations can be used both for authentication and large-scale
ﬁngerprint search. The minutiae-map can be used to further improve system ac-
curacy and interpretability by re-ranking candidates retrieved by the ﬁxed-length
representation. .

.

.

.

.

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93

xix

Figure 4.2 Fingerprint

impressions from one subject

training
dataset [188]. Impressions were captured longitudinally, resulting in the variability
across impressions (contrast and intensity from environmental conditions; distor-
tion and alignment from user placement). Importantly, training with longitudinal
data enables learning compact representations which are invariant to the typical
noise observed across ﬁngerprint impressions over time, a necessity in any ﬁnger-
print recognition system.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

in the DeepPrint

.

. 98

Figure 4.3 Unaligned ﬁngerprint images from NIST SD4 (top row) and corresponding

DeepPrint aligned ﬁngerprint images (bottom row).

. . . . . . . . . . . . . . . . . 101

Figure 4.4 Minutiae Map Extraction. The minutiae locations and orientations of an
input ﬁngerprint (a) are encoded as a 6-channel minutiae map (b). The “hot spots”
in each channel indicate the spatial location of the minutiae points. The kth channel
of the hot spots indicate the contributions of each minutiae to the kπ/3 orientation. 103

Figure 4.5 The custom multi-task minutiae branch of DeepPrint. The dimensions
inside each box represent the input dimensions, kernel size, and stride length, re-
spectively.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

. .

. .

. .

.

.

Figure 4.6 Examples of poor quality ﬁngerprint images from benchmark datasets. Row
1: Rolled ﬁngerprint impressions from NIST SD4. Row 2: Slap ﬁngerprint images
from FVC 2004 DB1 A. Rolled ﬁngerprints are often heavily smudged, making
them challenging to accurately recognize. FVC 2004 DB1 A also has several dis-
tinct challenges such as small overlapping ﬁngerprint area between two ﬁngerprint
images, heavy non-linear distortions, and extreme ﬁnger conditions (wet or dry).
Minutiae annotated with COTS A.
. . . . . . . . . . . . . . . . . . . . . . . . .

. 114

Figure 4.7 Closed-Set Identiﬁcation Accuracy of DeepPrint (with and without Product
Quantization (PQ)) on NIST SD4 and NIST SD14 (last 2,700 pairs) supplemented
with a gallery of 1.1 Million. Rank-1 Identiﬁcation accuracies are 95.15% and
94.44%, respectively. Search time is only 160 milliseconds. After adding product
quantization, the search time is reduced to 51 milliseconds and the Rank-1 accura-
cies only drop to 94.8% and 94.2%, respectively.

. . . . . . . . . . . . . . . . . . 120

Figure 4.8

Illustration of DeepPrint interpretability. The ﬁrst row shows three exam-
ple ﬁngerprints from NIST SD4 which act as inputs to DeepPrint. The second
row shows which pixels the texture branch is focusing on as it extracts its feature
representation. Singularity points are overlaid to show that the texture branch ﬁx-
ates primarily on regions surrounding the singularity points. The last row shows
pixels which the minutiae branch focuses on as it extracts its feature representa-
tion. We overlay minutiae to show how the minutiae branch focuses primarily on
regions surrounding minutiae points. Thus, each branch of DeepPrint extracts com-
plementary features which comprise more accurate and interpretable ﬁxed-length
ﬁngerprint representations than previously reported in the literature.

. . . . . . . . 124

xx

Figure 5.1 Face images (top row) and corresponding left thumb ﬁngerprints (bottom
row) of six different infants under 3 months of age. Face images were captured
by a Xiaomi MI A1 smartphone camera and ﬁngerprint images were captured by
our 1,900 ppi RaspiReader [49, 82] at the Saran Ashram Hospital, a charitable
organization in Dayalbagh, Agra, India.

. . . . . . . . . . . . . . . . . . . . . . . 126

Figure 5.2 Face images (top row) and corresponding left thumb ﬁngerprints (bottom
row) of an infant, Meena Kumari, acquired on (a) December 16, 2018 (Meena was
3 months old), (b) December 18, 2018 (3 months, 2 days old), (c) March 5, 2019
(6 months old), and (d) September 17, 2019 (12 months old) at Saran Ashram
Hospital, Dayalbagh, India. Note that as Meena ages, ﬁngerprint details emerge
such as visible pores. This level of detail is enabled by our 1,900 ppi reader.

. . . . 129

Figure 5.3 Overview of the Infant-Prints system. . . . . . . . . . . . . . . . . . . . . . 133
Figure 5.4 Prototype of the 1,900 ppi compact (1” × 2” × 3”), ergonomic ﬁngerprint
reader. An infant’s ﬁnger is placed on the glass prism with the operator applying
slight pressure on the ﬁnger. The capture time is 500 milliseconds. The prototype
can be assembled in less than 2 hours. See the video at: https://www.youtube.com/
watch?v=f8tYbE9Cwd0.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

.

Figure 5.5 An infant’s ﬁngerprints are acquired via (a) a 500 ppi commercial reader
(Digital Persona U.are.U 4500) and (c) our custom 1,900 RaspiReader. The cap-
tured ﬁngerprint images of the right thumb from the commercial reader and the
Infant-Prints reader for a 13 day old infant are shown in (b) and (d), respectively.
Manually annotated minutiae are shown in red circles (location) with a tail (orien-
tation). Blue arrows denote pores on the ridges.
. . . . . . . . . . . . . . . . . .

. 135

Figure 5.6

(a) Prototype of our 1,900 ppi contactless ﬁngerprint reader. During capture,
an infant’s ﬁnger is placed on top of a small, contactless, rectangular opening (an-
notated in red) on the reader (the size of this opening can be adjusted with different
sized slots). A camera captures the infant’s ﬁngerprint from behind the rectangular
opening. Examples of a processed (segmented, contrast enhanced), contactless in-
fant thumb-print (2 months old) is shown in (b) and the same infant’s thumb-print
acquired via contact-based ﬁngerprint reader in (c).

. . . . . . . . . . . . . . . . . 137

Figure 5.7

Infant ﬁngerprint collection at Saran Ashram hospital, Dayalbagh, India.
Pediatrician, Dr. Anjoo Bhatnagar, explaining longitudinal ﬁngerprint study to
the mothers while the authors are acquiring an infant’s ﬁngerprints in her clinic.
Parents also sign a consent form approved by the Institutional Review Board (IRB)
of our organizations.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139

.

.

.

xxi

Figure 5.8 Overview of the minutiae extraction algorithm. An input ﬁngerprint of any
size (n× m) is passed to the minutiae extraction network (Table 5.3). The network
outputs a n × m × 12 minutiae map H which encodes the minutiae locations and
orientations of the input ﬁngerprint. Finally, the minutiae map is converted to a
minutiae set {(x1, y1, θ1), ..., (xN , yN , θN )} of N minutiae.

. . . . . . . . . . . . . 141

Figure 5.9 An example infant ﬁngerprint patch (a) and the corresponding minutiae
map (b). Note, we only show 3 channels of the 12 channel minutiae map here for
illustrative purposes (red channel is the ﬁrst channel, green is the ﬁfth channel, and
blue is the ninth channel). Given the full 12 channels of the minutiae map in (b),
we can compute the minutiae locations (x, y) and orientations θ of the 1,900 ppi
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
ﬁngerprint patch in (a).

.

.

. 141

Figure 5.10 View of the manual minutiae markup/editing software used to markup
minutiae locations on a subset of infant ﬁngerprint images. These markups were
later used as ground truth to train our high resolution infant minutiae extractor. The
ﬁngerprint on the left (blue annotations) is coarsely annotated with Veriﬁnger v11
SDK to help speed up the annotation process. The ﬁngerprint on the right (red
annotations) shows the manually edited minutiae.

. . . . . . . . . . . . . . . . . . 143

Figure 5.11 Top row: Veriﬁnger minutiae detections; Bottom row: Minutiae detec-
tions from our high-resolution minutiae extractor. Manually marked minutiae are
annotated in red. Note that Veriﬁnger detects many of the true minutiae, but also
extracts a signiﬁcant number of spurious minutiae. Our proposed minutiae extrac-
tor has slightly lower detection accuracy (of true minutiae) than Veriﬁnger, how-
ever, it extracts signiﬁcantly fewer spurious minutiae. We further compare the two
approaches quantitatively in our experimental results.

. . . . . . . . . . . . . . . . 145

Figure 5.12 Effects of aging. (a) Acquired 3 month old enrollment image (orange) is
overlaid on a 1 year old probe image (blue). (b) An aged 3 month old enrollment
image (orange) is overlaid on a 1 year old probe image (blue). (c) 3 month old
enrollment minutiae set (green) is overlaid on a 1 year old probe minutiae set (red).
(d) An aged 3 month old enrollment minutiae set (green) is overlaid on a 1 year old
probe minutiae set (red). Following aging (b, d), the enrollment image and probe
image (and corresponding minutiae sets) overlap better.

. . . . . . . . . . . . . . . 146

Figure 5.13 Overview of the Infant-Prints texture matcher. We modify DeepPrint [48]
to accept 1,900 ppi high resolution infant ﬁngerprint images. The network is pre-
trained on adult ﬁngerprint images and then ﬁne-tuned (red layers) with the infant
dataset collected in [84].

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147

Figure 5.14 Infant ﬁngerprint (a) before enhancement and (b) after enhancement. Look-
ing inside the small window (red square) we can see that the enhanced infant ﬁn-
gerprint (b) has noticeably improved sharpness and clarity throughout the friction
ridge pattern when compared to (a).

. . . . . . . . . . . . . . . . . . . . . . . . . 149

xxii

Figure 5.15 Flipping a False Reject case to a True Accept by using our high-resolution
minutiae extractor.
(a) Minutiae are both extracted and matched using Veriﬁn-
ger. The signiﬁcant number of spurious minutiae extracted by Veriﬁnger render
it impossible for Veriﬁnger to establish minutiae correspondences. (b) Minutiae
are extracted using our high-resolution minutiae extractor and subsequently fed
into Veriﬁnger. Because our minutiae extractor is much more resistant to spurious
minutiae (on infant ﬁngerprints) than Veriﬁnger’s minutiae extractor, the Veriﬁnger
matcher is able to establish enough true minutiae correpondences to ﬂip this False
Reject to a True Accept. Quantitatively speaking, the Veriﬁnger match score is
improved from 23 to 48. .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157

Figure 5.16 Score Histograms comparing the contact-based RaspiReader with the con-
tactless RaspiReader (single ﬁnger performance). Using a contact-based reader
shows much better score separation than the contactless reader (TAR=72.9% vs.
TAR=35.6% @ FAR=1.0%).
. . . . . . . . . . . . . . . . . . . . . . . . . . . .

. 159

Figure 5.17 Example Infant-Prints failure cases. (a, b) Example of a False Accept due
to the similar friction ridge patterns, and the moisture in the enrollment image (a).
(c, d) Example of a False Reject due to the motion blur of the uncooperative infant
(d). These images highlight several of the challenges in infant ﬁngerprint recognition.160

xxiii

LIST OF ALGORITHMS

Algorithm 1 Extraction of Color Local Binary Patterns . . . . . . . . . . . . . . . . . . 79

Algorithm 2 Extract DeepPrint Representation . . . . . . . . . . . . . . . . . . . . . . 99

xxiv

Chapter 1

Introduction

When many of us hear the words “ﬁngerprint” or “ﬁngerprint recognition”, our minds will im-

mediately wander to a number of science ﬁction or crime solving television shows such as Person

of Interest1, NCIS2, or Forensic Files3, where a ﬁngerprint left behind at a crime scene is used to

identity a suspected criminal, or, in the case of Person of Interest, is used to spoof a ﬁngerprint

recognition system with a 3D printed ﬁngerprint in order to gain access to a secure facility. The

prevalence of ﬁngerprints and ﬁngerprint recognition in modern day entertainment shows is a tes-

tament to the ever increasing ubiquity of ﬁngerprint recognition in our society. Indeed, ﬁngerprint

recognition systems are now widely deployed across a plethora of different applications includ-

ing government services and facility access, smartphone unlock, forensics, customs and border

control, and national ID [111].

In this chapter, we explain how ﬁngerprint recognition came to be so pervasive in our day to

day lives. We begin by discussing the early history of ﬁngerprints and their progression (through

major scientiﬁc advances) towards the many applications, both in law enforcement as well as con-

sumer applications, we ﬁnd them in today. We then describe the pipeline of a standard ﬁngerprint

recognition system along with all of its constituent modules. Finally, we list some of the challeng-

1https://www.imdb.com/title/tt1839578/
2https://www.imdb.com/title/tt0364845/
3https://www.imdb.com/title/tt0247882/

1

Figure 1.1 One of the earliest recorded uses of ﬁngerprints include chinese clay seals which were
used to sign business transactions. Image retrieved from [63].

ing problems in state-of-the-art ﬁngerprint recognition systems and give an overview of how this

thesis aims to address these limitations.

1.1 History of Fingerprint Recognition

Human interest in ﬁngerprints dates back thousands of years. In fact, early examples of ﬁngerprint

patterns have been found in ancient Babylonian tablets dated back to 1955-1913 BC and later on in

ancient Chinese clay seals dated 600-700 AD [111]. However, it was not until centuries later that

ﬁngerprint recognition came to be studied with systematic scientiﬁc rigor for person recognition.

In 1684, Nehemiah Grew published the ﬁrst scientiﬁc study on the ridges, furrows, and pore

structure of ﬁngerprints [111]. Following this ﬁrst scientiﬁc paper, the late 19th century saw several

prominent scientists make major contributions to the ﬁeld of ﬁngerprint recognition. Sir William

Herschel ﬁrst proposed that the ridge structure of the ﬁngerprint remained unaltered over time

by examining his ﬁngerprint pattern in 1860 and then again in 1890 [35]. Dr. Henry Faulds

observed that not only were ﬁngerprints permanent, but they also grew back into the exact same

pattern when the outer skin of the ﬁngerprint was removed [35]. In other words, ﬁngerprints were

a permanent physical characteristic of an individual that remained with them throughout their

2

(a) Whorl

(b) Right Loop

(c) Left Loop

(d) Arch

(e) Tented Arch

Figure 1.2 Example ﬁngerprints of each of the ﬁve major ﬁngerprint classes deﬁned by Sir Edward
Henry. Images retrieved from the NIST SD4 database [129].

lifetime. This characteristic of ﬁngerprint permanence remains a central tenant to modern day

ﬁngerprint recognition systems [92, 111].

Other notable pioneers include Sir Edward Henry, and Sir Francis Galton. Henry is credited

with devising the Henry Classiﬁcation system which categorizes ﬁngerprints into one of several

classes (Figure 1.2). This classiﬁcation system was adopted by the New Scotland Yard in 1901

for criminal identiﬁcation [35]. Sir Francis Galton, a polymath and cousin of Charles Darwin,

wrote the landmark book “Finger Prints” in 1892. In it, he wrote down what he observed made

ﬁngerprints unique and thus able to identify an individual [35]. In particular, Galton made note

of ﬁngerprint minutiae, the endings and bifurcations throughout the ﬁngerprint ridge structure,

which are still the most common feature representation used in automated ﬁngerprint identiﬁcation

systems today [60].

Following the seminal studies of ﬁngerprints in the late nineteenth century, a paper detailing

the ﬁrst algorithmic approach to ﬁngerprint recognition was proposed by Mitchell Trauring in

Nature, 1963 [170]. In the 60 years since this ﬁrst paper on automated ﬁngerprint recognition was

published, signiﬁcant progress has been achieved in the area of automated ﬁngerprint identiﬁcation

(AFIS) systems. Today, standardized evaluations such as NIST FpVTE 2012 [178] and the more

recent FVC-ongoing [39,107], show ﬁngerprint authentication accuracies (single ﬁnger) as high as

T AR = 99.98% @ F AR = 0.01% (FVC-ongoing standard 1-to-1 database) and T AR = 99.4%

@ F AR = 0.01% (FVC-ongoing hard 1-to-1 database). Fusing the scores from multiple ﬁngers

can even further boost the recognition performance which is why ﬁngerprints from all 10 ﬁngers

3

are captured by law enforcement agencies and national ID systems. The cause of remaining failure

cases can be attributed to noisy and heavily distorted ﬁngerprint images [178].

In addition to becoming nearly perfect with respect to recognition accuracy, modern day AFIS

are much more well understood thanks to rigorous statistical studies on the central tenants of ﬁn-

gerprint recognition over the last 20 years. In particular, it had long been assumed to be true that

ﬁngerprints were both (i) unique (to a speciﬁc ﬁnger and a speciﬁc person), and (ii) permanent

(they do not change over time) without a strong statistical backing. This changed for the better in

2002 when Pankanti, Prabhakar, and Jain published their study: “On the individuality of ﬁnger-

prints” [136]. In it, the authors developed statistical models that showed that the probability of two
ﬁngerprint patterns, each with 36 minutiae would have a probability of 5.47 × 10−59 of sharing
all 36 minutiae points. In other words, the probability of two identical ﬁngerprints was incredibly

miniscule. Following this statistical study on the individuality or uniqueness of ﬁngerprints, in

2015, Yoon and Jain published another study in the Proceedings of the National Academy of Sci-

ences (PNAS) providing strong backing to the premise that ﬁngerprints are permanent, or remain

the same over time [188]. In their study, Yoon and Jain analyzed automated ﬁngerprint match

scores from 15,597 subjects over time lapses of 5-12 years. They found that although the match

scores for a given subject did drop over time, the overall recognition accuracy remained stable for

the maximum time interval in the dataset of 12 years. This lends strong statistical evidence to the

premise that ﬁngerprints are indeed a permanent physical trait.

From the early days of manual ﬁngerprint comparison to the modern day highly accurate AFIS,

the study of ﬁngerprints have come a long way. Not surprisingly, this has led to a proliferation of

ﬁngerprint recognition systems into a number of different applications globally. In the next section,

we discuss some of the more well known of these ﬁngerprint recognition applications.

4

(a) Forensics

(b) Border Security

(c) National ID

(d) Mobile Unlock

Figure 1.3 Various applications of ﬁngerprint recognition. (a) An example of a latent ﬁngerprint
left behind on a dollar bill, which could be subsequently used to search a database of known
criminals; (b) a woman has her ﬁngerprints taken at US Customs (OBIM system) prior to entry
into the country; (c) An Indian citizen is authenticated by the Aadhaar system; (d) a mobile phone
is unlocked, bypassing the need for a password or key-code for access. Images retrieved from
Google Images.

1.2 Major Applications

Thanks to the high performance (accuracy and speed) of modern day AFIS [39], and the strong

statistical backing of the foundational premises (uniqueness and permanence) upon which AFIS

are built, ﬁngerprint recognition has exploded into a myriad of different applications throughout

our world today. Some of the more noteworthy or well known of these applications are enumerated

below and are shown in Figure 1.3.

• Forensics: Already in the mid to late nineteenth century, Faulds, Herschel, Henry, and
Bertillon were manually examining ﬁngerprints to identify repeat criminals [35]. In 1924,

the FBI formally started the Identiﬁcation Division of the FBI to collect and store inked

ten-print cards from criminals. Later on in 1999, this was rolled into the FBI’s Integrated

Automated Fingerprint Identiﬁcation System (IAFIS) where ﬁngerprints (tenprints) were

digitized, stored, and automatically compared. Finally, in 2011, the FBI’s Next Genera-

tion Identiﬁcation (NGI) system was established to improve upon the outdated IAFIS sys-

tem [56]. In particular, it enabled faster and more accurate ﬁngerprint recognition capabil-

ities and utilized additional biometric modalities (face and iris). Today, NGI continues to

maintain a database of 78 million criminal ﬁngerprint records, and 58 million civilian ﬁnger-

5

print records. In August of 2020, the NGI system averaged 18,000 suspected criminal latent

ﬁngerprint queries per day [57].

• Border Security:

The Ofﬁce of Biometric Identity Management (OBIM, formerly US-VISIT) program man-

ages the largest biometric repository in the United States. In 2020, the program was projected

to process 152 million query records against a continually growing gallery size of 280 mil-

lion records. The OBIM vision statement is to “lead the use of biometric identity for a safer

world, enhanced individual privacy, and improved quality of life” [37]. One of the primary

ways OBIM accomplishes this vision statement is by preventing criminals and dangerous

individuals from entering the United States. Of the 88 million query transactions recorded

in 2014, OBIM successfully ﬂagged 2.7 million queries to individuals on the United States

watchlist4.

• National ID:

The world’s largest biometric recognition system is India’s Aadhaar (meaning foundation

in Hindi). Aadhaar uses all 10 ﬁngerprints, 2 irises, and face image of an Indian citizen

to deduplicate and then link the citizen to a 12-digit unique identiﬁer. As of September

of 2020, over 1.2 billion Indian citizens have been enrolled into Aadhaar5. This system

is successfully used to provide beneﬁts to the marginalized segment of the population and

to facilitate ﬁnancial transactions. One limitation of Aadhaar is that it starts enrollment at

the age of 5 years. Unfortunately, this leaves many of India’s more vulnerable citizens (its

infants and children) at risk.

• Mobile Unlock and Payments

Perhaps the most prevalent use of ﬁngerprint recognition in today’s world is that of mobile

unlock and payments. As of August 2020, there are estimated to be 5.15 billion smartphone

4https://bit.ly/3cCSEgL
5https://uidai.gov.in/aadhaar dashboard/

6

users globally6. Of these over 5 billion smartphones, it has been reported that in 2018, 60%

of them would be equipped with ﬁngerprint recognition technology for unlock and payment

services7. Furthermore, this percentage is on a year over year upward trend.

Due to the convenience ﬁngerprint recognition technology has afforded smartphones, major

payments companies (Visa8 and Master Card9) are now integrating ﬁngerprint recognition

technologies directly into credit cards via a concept referred to as “Match on Card”. Match

on Card would enable a user to enroll their ﬁngerprint template to a chip on their credit

card which could then be used to perform a ﬁnancial transaction in lieu of a pin number.

The ﬁngerprint template data would never leave the credit card (so as to keep the ﬁngerprint

template secure) as it would be directly matched to the query ﬁngerprint on the chip on the

card.

The applications and statistics enumerated above indicate that it is entirely plausible that over

half of the world’s population are now using ﬁngerprint recognition in their day to day lives. The

data also suggests that these numbers will continue to grow. In the next section, we dive into the

pipeline of modern day automated ﬁngerprint recognition systems to better understand (from a

technical point of view) how ﬁngerprint recognition systems have become so prevalent in our day

to day lives.

1.3 Pipeline of Fingerprint Recognition Systems

Automated ﬁngerprint recognition systems operate in one of three major modes: (i) enrollment,

(ii) authentication, or (iii) search. Each of these modes of operation is supported by multiple sub-

modules within the recognition system including ﬁngerprint (1) sensing, (2) feature extraction, and

(3) matching. In the following section, we describe each of these modes of operation, and also the

individual sub-modules that support such functionality.

6https://bit.ly/2HyRojs
7https://bit.ly/3cATAlH
8https://vi.sa/3mUT3A8
9https://bit.ly/3j7lJn9

7

• Enrollment: During the enrollment stage (Figure 1.4), a ﬁngerprint is captured by the ﬁn-
gerprint reader and transferred to the feature extractor. The feature extractor then extracts

salient, discriminative features (e.g. minutiae points) and packs them into a template along

with the user’s meta-data. Finally, the template is saved in the enrollment database. Ide-

ally, this template should be encrypted prior to its storage in the template database to protect

against the event that a hacker breaks into the enrollment database. After enrollment, the

ﬁngerprint recognition system can operate in one of two modes. It can be utilized for au-

thentication (1:1 matching) or search (1:N matching) [111].

• Authentication: Fingerprint authentication refers to a 1 to 1 or (1:1) matching application.
In such a scenario, the user presents their ﬁngerprint to the reader along with a claimed iden-

tity (e.g. a PIN, or the identity is implicitly known via ownership of a device). Subsequently,

a probe feature set (or template10) is extracted from the newly presented ﬁngerprint, and an

enrollment template for the claimed identity is retrieved from the database. Finally, these

two templates are compared to make the binary decision of match or no match (Figure 1.5).

Examples of ﬁngerprint authentication include access control, and India’s Aadhaar, where

authentication is made based upon a 12-digit Aadhaar number for government beneﬁts and

assistance [111].

• Search:

Fingerprint search refers to a 1 to N or (1 : N) matching application (where

N is the number of all the users in the enrollment database).

In this scenario, a user’s

ﬁngerprint is again captured by the ﬁngerprint reader. This query ﬁngerprint is then passed

to the feature extractor to extract a salient, discriminative template. This query template

is then matched to each of the N templates already enrolled in the gallery. The matcher

returns a rank order list of possible candidates which are most similar to the query or probe

representation (Figure 1.6). A notable example of ﬁngerprint search is when a ﬁngerprint

left behind at a crime scene (or a latent ﬁngerprint) is searched against a criminal database

10We refer to the terms feature set, template, representation interchangeably in this thesis.

8

in an effort to identity a suspect [111]. Another example is the de-duplication done prior to

enrollment in a national ID system such as Aadhaar.

Figure 1.4 Enrollment Phase. A ﬁngerprint is captured by the reader and transferred to the feature
extractor where minutiae and other salient, discriminative features are extracted and packed into a
template. The extracted template is then stored in the database.

Figure 1.5 Fingerprint Authentication Schematic. During authentication, a 1:1 match is conducted
between a newly extracted template, and a template already stored in the database. In this scenario,
we are answering the question, “Is this person a match to the speciﬁed enrollment template?”

Figure 1.6 Fingerprint Search Schematic. During search, N matches are conducted between a
newly extracted template (probe or query), and N templates already stored in the database. The
matcher returns a ranked list of the candidates most similar to the query. In this scenario, we are
answering the question, “Who is this person?”

9

The functionality of each of the aforementioned ﬁngerprint recognition modes of operation is

enabled by a conﬁguration of multiple sub-modules (ﬁngerprint reader, feature extractor, matcher)

within the ﬁngerprint recognition system. Each of these sub-modules are described below.

1.3.1 Fingerprint Readers

Early applications of ﬁngerprint recognition in law enforcement required inking a user’s ﬁngers,

and having them press down on a sheet of card stock paper (Fig 1.7). The ﬁngerprints captured

could be rolled ﬁngerprints (captured by rolling the ﬁnger from one side to another) or slap/plain

ﬁngerprints captured by pressing the ﬁngers ﬂat against the card. These ﬁngerprints were then ﬁled

away and manually compared by an examiner. Since that time, a number of ﬁngerprint readers

have been developed which are signiﬁcantly more convenient than the old “ink on paper” capture

techniques.

Figure 1.7 Fingerprints captured using ink and card stock paper. The ﬁngerprints in the top row are
rolled ﬁngerprints, whereas the ﬁngerprints in the bottom row are slap/plain ﬁngerprints. Image
reproduced from [88].

Fingerprint readers or scanners use a variety of sensing technologies to capture and convert a

physical ﬁngerprint into a digital image. We make a distinction in this thesis between the terms

reader and sensor since a reader technically is utilizing one of a number of different sensors to cap-

10

ture the digital image. The major types of these sensing technologies include ultrasound, frustrated

total internal reﬂection (FTIR), direct-view imaging, capacitance, thermal, and pressure sensors.

Most all of these readers capture ﬁngerprints at 500 pixels per inch (ppi). More expensive readers

can capture at higher resolution (e.g. 1000 ppi) in order to capture ﬁner level features [111]. More

details on individual sensing technologies are enumerated below:

Figure 1.8 Examples of different types of optical-based ﬁngerprint readers (Idemia https://www.
idemia.com/, HIDGlobal https://www.hidglobal.com/). Images retrieved from Google Images.

• Optical: Of all the optical sensing technologies (Fig. 1.8), the most widely utilized is that
of frustrated total internal reﬂection (FTIR). FTIR sensing works by making use of both a

light source and a glass prism. In particular, by mounting a camera at an appropriate angle

to a glass prism, light from the ﬁngerprint ridges in contact with the glass prism are reﬂected

back to the camera, while light from the valleys scatter [111]. The result is a high contrast

ﬁngerprint image.

Other common optical ﬁngerprint readers use direct-view imaging where a light source il-

luminates the ﬁnger and light from both the ridges and valleys are reﬂected back towards

the camera [111]. The Lumidigm multi-spectral reader is built upon this sensing technol-

ogy. Many contactless optical readers (e.g. the Morpho Wave) also use direct-view imaging.

Direct-view ﬁngerprints have lower contrast than FTIR ﬁngerprints, but are less impacted by

moisture on the ﬁnger due to environmental humidity.

11

Figure 1.9 Examples of different types of solid-state readers (both capacitive) and an in-display,
mobile-phone, ultrasound reader. Images retrieved from Google Images.

• Solid-State:

Solid state sensing technology (Fig. 1.9) operates by using an array of mini-sensors which

measure differentials in capacitance, temperature, or pressure between the ridges and valleys.

Because of their small size and low cost, solid state sensors are the most commonly deployed

sensing technology on mobile devices [111].

• Ultrasound:

Ultrasound sensing (Fig. 1.9) works by employing acoustic waves towards the ﬁngertip on

the platen. Then, a receiver gathers the echoed responses and develops a depth proﬁle of the

ﬁngerprint [111]. One of the main advantages of ultrasound sensing is that it enables a sub-

surface ﬁngerprint image to be captured. This sub-surface ﬁngerprint could be a useful for

detecting fake ﬁngerprint attacks (otherwise commonly known as spoof attacks or presenta-

tion attacks). The sub-surface ﬁngerprint is also particularly useful for improving the image

quality of the elderly population which often has worn out and damaged ﬁngerprints on the

surface of the ﬁnger, but will still have a higher quality sub-surface ﬁngerprint. Up until

recently, commercial ultrasound readers held a very minor marketshare. However, recently

QualComm Inc. developed an in display ultrasound sensor for the mobile phone. This sensor

has been widely deployed in the Samsung smartphone series (Galaxy S10 onwards) 11.

11https://www.samsung.com/global/galaxy/what-is/ultrasonic-ﬁngerprint/

12

(a) Level-1 features

(b) Level-2 features

Figure 1.10 The most popular ﬁngerprint representation consists of (a) global level-1 features
(ridge ﬂow, core, and delta) and (b) local level-2 features, called minutiae points, together with
their descriptors (e.g., texture in local minutiae neighborhoods). The ﬁngerprint image illustrated
here is a rolled impression from the NIST SD4 database [129]. The number of minutiae in NIST4
rolled ﬁngerprint images range all the way from 12 to 196.

Two chapters of this thesis will focus heavily on the ﬁngerprint reader component of ﬁnger-

print recognition systems. In particular, (i) the second chapter of this thesis involves the proper

operational evaluation of the aforementioned ﬁngerprint readers, particularly evaluating the inter-

operability of ﬁngerprint recognition systems when two different types of sensing technologies

are used to capture the enrollment image and the probe/query image. (ii) Chapter 3 of this the-

sis focuses on securing the ﬁngerprint reader module by preventing fake ﬁngerprint attacks (more

commonly referred to as spoof attacks or presentation attacks).

1.3.2 Feature Extraction

After acquiring a digital format of the ﬁngerprint via a ﬁngerprint reader employing one of the

aforementioned sensing technologies, the next step in the ﬁngerprint recognition pipeline is to

13

extract salient and discriminative features which will comprise the ﬁngerprint template. Fingerprint

feature sets are usually separated into one of three levels of features (Fig. 1.10).

• Level-1: Level-1 features are coarse or more global level features (Fig. 1.10). These features
include the orientation ﬁeld (or ridge-ﬂow), and ridge spacing statistics of the ﬁngerprint im-

age. Additionally, major landmarks called singularities (further divided into core points and

deltas) are included. Finally, the ﬁngerprint type (earlier introduced as the Henry Classiﬁ-

cation System in Figure 1.2) can be considered as Level-1 features. While Level-1 features

can be useful for aligning ﬁngerprints, quickly indexing large galleries for a candidate list, or

classifying them in accordance with the Henry Classiﬁcation System, they are not discrimi-

native enough for recognition standalone [111].

• Level-2:

Level-2 features are more local features than the global Level-1 features (Fig. 1.10). These

features are known as minutiae points. A minutiae point can be one of two types. Either it

is a ridge bifurcation (a point at which a running ridge splits into two), or an ending (a point

at which a running ridge terminates) (Fig. 1.11). Furthermore, each minutiae is comprised
of a spatial location in the ﬁngerprint {x, y} and a ridge orientation (θ). The collection
of all minutiae in a ﬁngerprint image comprise the most widely standardized and utilized

ﬁngerprint template. Minutiae points are best extracted at a ﬁngerprint image resolution 500

pixels per inch (ppi) [111]. They can be extracted using a variety of different techniques

including the more recent deep network based approaches [125, 168].

• Level-3:

Level-3 features are the ﬁnest level of ﬁngerprint features (Fig. 1.11). They include fea-

tures such as sweat pores, dots, and incipient ridges. In order to capture Level-3 features,

a more expensive 1000 ppi ﬁngerprint reader is required [86]. Due to the high-resolution

requirement, and computational cost of extraction, Level-3 features are typically not used

for ﬁngerprint recognition even though it has been shown that they can be used to further

14

Figure 1.11 Examples of Level-2 features (two types of minutiae) and Level-3 features (sweat
pores and incipient ridges).

improve the recognition performance [111]. Several potential use cases for Level-3 features

which can afford the computational complexity tradeoff in favor of higher accuracy include

high-resolution infant ﬁngerprint matching and latent ﬁngerprint matching.

More recently, deep networks have been employed to extract deep features from ﬁngerprint

images, including ﬁxed-length ﬁngerprint representations comprised of textural and minutiae re-

lated features [48]. Indeed, Chapter 4 of this thesis demonstrates the use of a deep network, called

DeepPrint to extract a ﬁxed-length representation from a ﬁngerprint image to enable faster large

scale search and more secure matching within the encrypted domain.

1.3.3 Template Database

A ﬁngerprint template is comprised of some combination of the aforementioned feature sets, along

with a user’s meta-data. The International Standards Organization (ISO) has deﬁned a standard

template under ISO/IEC 19794-2 (2005)12 (essentially a minutiae set). Typically, a commercial

vendor will support extraction of the ISO standard template (to enable compatibility and interoper-

ability with legacy template databases) and an additional proprietary template comprised of other

levels and types of features for improved accuracy and speed.

The collection of templates from all subjects comprise the template database. Due to the

plethora of personal identifying information (PII) in the template database, the template database

12https://www.iso.org/standard/52537.html

15

Figure 1.12 Example of minutiae match between two ﬁngerprint impressions of the same ﬁnger.
This example highlights the difﬁculty of minutiae matching given a poor quality enrollment image
(left) which has many missing minutiae. Despite this difﬁculty, a COTS minutiae matcher is able
to correctly match these two ﬁngerprints with a score of 150, well above the score threshold of 69
@ F AR = 0.01%. Fingerprints retrieved from the FVC 2004 DB1 A database [109]

.

must be adequately secured. Securing the template database in a manner that still allows the un-

derlying ﬁngerprint system to operate at high levels of accuracy remains a signiﬁcant research

challenge [87]. In Chapter 4 of this thesis, we show how a discriminative ﬁxed-length represen-

tation can be learned, secured, and matched in the encrypted domain using fully homomorphic

encryption.

1.3.4 Matching

The ﬁnal step in the ﬁngerprint recognition pipeline is that of matching, or comparing two tem-
plates. Typically the matcher will output a score s within some range (e.g. s ∈ [0, 1]). If the score
is above a threshold t, then the decision is a match, otherwise, it is a non-match. The threshold

t is selected in order to balance (according to the speciﬁc application, e.g. a mobile phone vs.

government facility access) the false accepts and false rejects of the matcher.

16

The most common approach to ﬁngerprint matching is that of minutiae matching. The goal

of minutiae matching is to align two minutiae sets (frequently of variable, unknown length) and

subsequently ﬁnding a maximum number of candidate corresponding minutiae (Fig. 1.12) [111].

Some of the challenges with minutiae matching include: (i) non-linear distortion between minutiae

sets and (ii) spurious or missing minutiae [72]. The extraction of ﬁxed-length representations

from ﬁngerprints, shown in Chapter 4, enables matching using simple distance metrics such as the

cosine distance [48] which operates at orders of magnitude faster speeds than minutiae matching

algorithms (a signiﬁcant beneﬁt for large scale ﬁngerprint search applications).

1.4 Remaining Challenges

Although ﬁngerprint recognition systems have come a long way with respect to automation, accu-

racy, statistical understanding, and prevalence throughout society from the early days of Galton,

Herschel, Henry, and other pioneers, current automated ﬁngerprint recognition systems still face

several limitations. The following section lists several of the limitations of prevailing ﬁngerprint

recognition systems. This list is by no means exhaustive, rather it is a speciﬁc list of problems

which this thesis aims to address.

1.4.1 Fingerprint Reader Evaluations

Current automated ﬁngerprint recognition systems make use of a plethora of different ﬁngerprint

readers, each employing different sensing technologies. Given that a particular recognition system

may make use of multiple types of ﬁngerprint readers, it is very important to have standardized

evaluations (especially interoperability evaluations, where the ﬁngerprint reader used during en-

rollment and authentication or search differ) for ﬁngerprint readers to ensure that the recognition

accuracy does not degrade as a result of poor ﬁngerprint sensing.

Existing standards for ﬁngerprint reader evaluations are primarily based upon the FBI PIV and

Appendix F standards [55]. The Appendix F standard is comparatively stringent, requires pristine

17

image capture, and is designed to facilitate evaluation of ﬁngerprint readers used in person search

scenarios (one to many comparisons). The PIV standard is a softer standard than Appendix F and

is designed to evaluate ﬁngerprint readers used in person veriﬁcation scenarios (one to one com-

parison). Both of these standards use imaging targets that are fabricated by projecting a calibration

pattern (e.g. sine gratings) onto a ﬂat surface (Fig. 1.13). These targets are useful for struc-

tural (white-box) testing of ﬁngerprint readers since they ensure that certain quantitative imaging

thresholds are met by the ﬁngerprint reader’s sensing component, however, these targets have lit-

tle resemblance to the human ﬁngers that the readers will be exposed to in an operational setting.

As such, controlled operational (black-box)13 evaluations of ﬁngerprint readers using the existing

standards and targets are limited at best.

(a) Single Finger
Reader Target

(b) 3D Contactless Reader Target

Figure 1.13 Examples of existing ﬁngerprint reader calibration targets. These targets are useful for
white-box testing ﬁngerprint readers, ensuring that they meet certain quantitative imaging thresh-
olds, however, they are very dissimilar from human ﬁngers. As such they are not useful for realistic
operational evaluations of ﬁngerprint readers.

1.4.2 Fingerprint Presentation Attack Detection

An outstanding security ﬂaw with ﬁngerprint recognition systems is that of successful spoof at-

tacks, or presentation attacks14. The most common type of presentation attack (referred to as

13White-box testing evaluates the internal sub-components of a system, whereas black-box testing focuses on testing

the end-to-end system using system inputs and outputs [13].

14In ISO standard IEC 30107-1:2016(E), presentation attacks are deﬁned as the “presentation to the biometric data

capture subsystem with the goal of interfering with the operation of the biometric system” [81].

18

Figure 1.14 Examples of ﬁngerprint spoofs made from different materials. The material variety
demonstrates why it is difﬁcult to develop a spoof detector which generalizes well across all mate-
rial types.

spooﬁng) occurs when a hacker intentionally assumes the identity of unsuspecting individuals,

called victims here, through stealing their ﬁngerprints, fabricating spoofs (fake ﬁngers made of

common household materials like gelatin or wood glue) (Fig. 1.14) with the stolen ﬁngerprints, and

maliciously attacking ﬁngerprint recognition systems with the spoofs into identifying the hacker

as the victim15 [113, 115, 187].

Over the last decade, a number of different approaches using either hardware or software have

been proposed to automatically detect and ﬂag ﬁngerprint spoof attacks at the ﬁngerprint reader

to thwart these attacks [113]. However, most of these approaches do not meet adequate levels of

accuracy for ﬁeld deployment. Furthermore, many of the more novel learning based approaches

fail when presented with spoofs fabricated from materials not seen during training of the spoof

detector. In fact, several studies have reported up to a three-fold increase in error when testing spoof

detectors on unknown material types [112, 167, 185]. As such, robust, generalizing ﬁngerprint

15Presentation attacks can also occur when (i) two individuals are in collusion or (ii) an individual obfuscates his or
her own ﬁngerprints to avoid recognition [113]. However, in this thesis our speciﬁc aim is to stop ﬁngerprint spooﬁng
presentation attacks.

19

Figure 1.15 Failures of a state-of-the-art COTS minutiae-based matcher (minutiae annotated with
COTS). The genuine pair (two impressions from the same ﬁnger) in (a) was falsely rejected at
0.1% FAR (score of 9) due to heavy non-linear distortion and moist ﬁngers. The imposter pair
(impressions from two different ﬁngers) in (b) was falsely accepted at 0.1% FAR (score of 38) due
to the similar minutiae distribution in these two ﬁngerprint images (the score threshold for COTS
A @ FAR = 0.1% is 34). These slap ﬁngerprint impressions come from public domain FVC 2004
DB1 A database [109]. The number of minutiae in FVC 2004 DB1 A images range from 11 to 87.

presentation attack detection remains a challenging, unsolved problem in ﬁngerprint recognition

systems.

1.4.3 Fixed-Length Fingerprint Representations

The most common type of ﬁngerprint representation is that of an unordered, variable length minu-

tiae set. Although AFIS based on minutiae representations (i.e. handcrafted features) have seen

tremendous success over the years, they have several limitations.

• Minutiae-based representations are of variable length, since the number of extracted minu-
tiae varies amongst different ﬁngerprint images even of the same ﬁnger (Figure 1.15). Varia-

tions in the number of minutiae originate from a user’s interaction with the ﬁngerprint reader

(placement position and applied pressure) and condition of the ﬁnger (dry, wet, cuts, bruises,

etc.). This variation in the number of minutiae causes two main problems: (i) pairwise ﬁn-

gerprint comparison is computationally demanding and varies with number of minutiae and

20

(ii) matching in the encrypted domain, a necessity for user privacy protection, is computa-

tionally expensive, and results in loss of accuracy [87].

• In the context of global population registration, ﬁngerprint recognition can be viewed as a
75 billion class problem (≈ 7.5 billion living persons around the globe, assuming nearly
all with 10 ﬁngers) with large intra-class variability and large inter-class similarity. This

necessitates extremely discriminative yet compact representations that are complementary

and at least as discriminative as the traditional minutiae-based representation. For example,

India’s civil registration system, Aadhaar, now has a database of over 1.2 billion residents

who are enrolled based on their 10 ﬁngerprints, 2 irises, and face image [171].

• Reliable minutiae extraction in low quality ﬁngerprints (due to noise, distortion, ﬁnger con-
dition) is problematic, causing false rejects in the recognition system. See also NIST ﬁnger-

print evaluation FpVTE 2012 [178].

Given these limitations with the prevailing minutiae representation, it is desirable to extract

discriminative ﬁxed-length representations from ﬁngerprints. Fixed-length representations can be

compared extremely quickly using simple distance metrics, and can be matched securely in the

encrypted domain using fully homomorphic encryption. Approaches in the literature attempting

to extract ﬁxed-length ﬁngerprint representations fail to match the accuracy of traditional minutiae

matchers [18, 89, 90, 159].

1.4.4 Infant Fingerprints

While ﬁngerprint recognition systems are now being used around the world in a number of appli-

cations by billions of teenagers and adults, infants and young children remain excluded from these

applications. Indeed, current automated ﬁngerprint recognition systems fail to work on infants (0-

12 months of age) and young children due to (i) small inter-ridge spacing (often less than 1 pixel

in-between two ridges), (ii) non-linear distortion from soft infant skin, (iii) uncooperative subjects

21

(a)

(b)

(c)

(d)

Figure 1.16 Examples of low quality infant ﬁngerprints. These examples demonstrate the difﬁ-
culty in using automated ﬁngerprint recognition systems for infant recognition. Note the small
inter-ridge spacing, debris, motion blur, and moisture throughout the different impressions. These
images were captured when the infant’s were 2 months old.

causing motion blur, (iv) wet or dry ﬁngers, and (v) debris, threads, or hairs from the infant’s cloth-

ing or the mother’s hair obfuscating the ﬁngerprints (Fig. 1.16). This is alarming since national

ID programs like India’s Aadhaar now use ﬁngerprints for providing government assistance and

beneﬁts distribution (Aadhaar starts enrollment at 5 years of age leaving younger ages excluded).

Developing a ﬁngerprint recognition system for infants could aid in establishing veriﬁable identity

for these infants which could in turn be used for a myriad of applications such as robust vaccination

tracking.

1.5 Contributions

This thesis aims to address each of the aforementioned limitations with state-of-the-art automated

ﬁngerprint recognition systems.

• Universal 3D Wearable Fingerprint Targets

To enable robust, standardized ﬁngerprint reader operational evaluations (especially inter-

operability), we present the fabrication of an interoperable 3D ﬁngerprint target through a

molding and casting process. We call our target the universal ﬁngerprint target. Like previ-

ous ﬁngerprint targets in [5], the universal ﬁngerprint targets share a 3D geometry similar to

22

a ﬁngerprint surface, have mechanical properties similar to human skin, and are mapped with

a ﬁngerprint image, either real or synthetic. However, unlike previous ﬁngerprint targets, the

universal ﬁngerprint targets are unique in that they incorporate the technically pertinent me-

chanical, optical, and electrical properties of the human skin within a single target, making

it possible for the universal ﬁngerprint targets to be imaged by all major ﬁngerprint sensing

technologies in use (capacitive, contact-optical, contactless-optical)16

• RaspiReader: Open Source Fingerprint Reader

To address the security vulnerability of presentation attacks in ﬁngerprint recognition sys-

tems, we open source a low cost, spoof resistant ﬁngerprint reader, called RaspiReader.

RaspiReader is an FTIR ﬁngerprint reader customized with two cameras for image acqui-

sition rather than a single camera. Use of two cameras enables robust ﬁngerprint spoof

detection, since we can extract features from two complementary, information rich images

instead of processed grayscale images output by traditional COTS optical ﬁngerprint read-

ers. We demonstrate in our experimental results that RaspiReader enables a signiﬁcant boost

in spoof detection performance in comparison to COTS optical readers. We also show that

RaspiReader is more generalizable to unseen materials than existing COTS readers.

• Learning a Fixed-Length Fingerprint Representation

To overcome the limitations of prevailing minutiae-based matchers, we design a deep net-

work embedded with ﬁngerprint domain knowledge, called DeepPrint, to learn a ﬁxed-

length representation of 200 bytes which discriminates between ﬁngerprint images from

different ﬁngers. While prevailing minutiae-matchers require expensive graph matching al-

gorithms for ﬁngerprint comparison, the 200 byte representations extracted by DeepPrint

can be compared using simple distance metrics such as the cosine similarity, requiring only
d multiplications and d − 1 additions, where d is the dimensionality of the representation
16We use the term universal to indicate that the targets can be imaged by all existing, major types of ﬁngerprint

readers (contact-optical, contactless-optical, capacitive, Ultrasound, and multi-spectral direct-view).

23

(for DeepPrint, d = 192)17. This fast comparison of DeepPrint representations is partic-

ularly useful for large-scale image search, when millions or even billions of comparisons

must be made. Another signiﬁcant advantage of this ﬁxed-length representation is that it can

be matched in the encrypted domain using fully homomorphic encryption [10, 14, 174, 175].

Finally, since DeepPrint is able to encode features that go beyond ﬁngerprint minutiae, it

is able to match poor quality ﬁngerprints when reliable minutiae extraction is not possible.

In short, DeepPrint enables faster, and more secure ﬁngerprint matching than the prevailing

minutiae representation with comparable levels of accuracy.

• Infant-ID: Fingerprints for Global Good

To extend the use of ﬁngerprint recognition for all ages, we have developed an end-to-end

infant ﬁngerprint recognition system. We have prototyped a high-resolution (1,900 ppi)

reader designed for infants. Using this reader, we have captured a longitudinal dataset (data

acquired over a time lapse of 1 year, in 4 sessions) of 315 infants from a rural clinic in

Agra, India. To match these high-resolution infant ﬁngerprints, we have developed a high-

resolution infant ﬁngerprint matcher. Finally, we have demonstrated that by using our infant

ﬁngerprint reader and matcher, we are able to enroll infants at 2-months of age, and recognize

them a full year later. This allows for our infant matcher to be used to alleviate infant

suffering around the world by providing every infant a digital and veriﬁable identity which

could be used for vaccination tracking, food distribution, and government assistance later in

life.

17The DeepPrint representation is originally 768 bytes (192 features and 4 bytes per ﬂoat value). We compress the
768 bytes to 200 by scaling the ﬂoats to integer values between [0,255] and saving the two compression parameters
with the features. This loss in precision (which saves signiﬁcant disk storage space) very minimally effects matching
accuracy.

24

Chapter 2

Universal 3D Wearable Fingerprint Targets:

Advancing Fingerprint Reader Evaluations

In this chapter, we propose a manufacturing process to create Universal 3D Wearable Fingerprint

Targets for repeatable and controlled ﬁngerprint reader evaluations [43]. The Universal Targets are

designed to be imaged across a variety of ﬁngerprint sensing technologies including capacitive,

contact-optical, and contactless-optical and as such, they are particularly well suited for ﬁngerprint

reader interoperability studies. Fingerprint reader interoperability refers to how robust ﬁngerprint

recognition systems are to variations in the images acquired by different types of ﬁngerprint read-

ers. To build universal 3D ﬁngerprint targets, we adopt a molding and casting framework consisting

of (i) digital mapping of ﬁngerprint images to a negative mold, (ii) CAD modeling a scaffolding

system to hold the negative mold, (iii) fabricating the mold and scaffolding system with a high

resolution 3D printer, (iv) producing or mixing a material with similar electrical, optical, and me-

chanical properties to that of the human ﬁnger, and (v) fabricating a 3D ﬁngerprint target using

controlled casting. Our experiments conducted with PIV and Appendix F certiﬁed optical (con-

tact and contactless) and capacitive ﬁngerprint readers demonstrate the usefulness of universal 3D

ﬁngerprint targets for controlled and repeatable ﬁngerprint reader evaluations and also ﬁngerprint

reader interoperability studies.

25

Figure 2.1 A Universal 3D Fingerprint Target fabricated in (a) can be imaged by a variety of
popular ﬁngerprint readers (contact-optical, contactless-optical, and capacitive) shown in (b). The
sensed images of the 3D ﬁngerprint target in (a) are shown in (c). This demonstrates that our targets
are appropriate for ﬁngerprint reader interoperability evaluation studies. Similarity scores for each
sensed ﬁngerprint image (with the 2D mapped target image) are displayed below each ﬁngerprint
image in (c). Veriﬁnger 6.3 SDK was used for generating similarity scores. The score threshold at
0.01 % FAR is 33.
2.1

Introduction

The current PIV and Appendix F evaluation standards for ﬁngerprint readers use simple calibration

patterns to ensure certain imaging quality thresholds are met. However, because these targets are

dissimilar (optically, electrically, and mechanically) from the real human ﬁngers the readers will

be imaging in an operational scenario, current evaluation standards are limited at best.

To address the challenges of robust operational evaluation inherent to imaging devices, the

medical imaging community has developed 3D targets (phantoms) as evaluation specimens. Phan-

toms are useful for evaluating a variety of medical imaging devices in areas such as radiography,

tomography, and ultrasonic imaging [116] [135]. Use of live subjects for repeated evaluation of

medical devices is impractical because of the health hazards and monetary costs involved. How-

26

Figure 2.2 Example phantom of a human hand [116] used in the medical domain.

Table 2.1 Properties of the Human Finger [42, 51, 52]

Shore A
Hardness

20-41

Tensile
Strength
(MPa)
5-30

Elongation
at Break

(%)

35-115

Electrical
Resistivity

(Ω-cm)

2.5 ∗ 102 – 8 ∗ 106

ever, realistic 3D phantoms (Fig 2.2) make accurate operational evaluation of these devices pos-

sible. Proper operational evaluation of ﬁngerprint readers can only be accomplished, in a similar

manner, by using 3D ﬁngerprint targets (phantoms) with similar characteristics to the human ﬁnger.

2.1.1 3D Fingerprint Targets

Some research has been conducted developing 3D targets towards achieving the aforementioned

goal. In 2011, Orandi et. al developed 3D cylindrical metal targets mapped with 2D calibration

patterns for contactless ﬁngerprint readers [134]. However, because these targets are rigid and

completely dissimilar in mechanical, optical, and capacitive properties to the human ﬁnger, they

can not be used by contact-based ﬁngerprint readers. More recently, in 2016, Arora et. al produced

high ﬁdelity 3D ﬁngerprint targets using a high resolution, state-of-the-art 3D printer [5] [4] [6].

27

Figure 2.3 High ﬁdelity, wearable, 3D ﬁngerprint targets. (a) 3D ﬁngerprint target printed using
TangoBlackPlus FLX980 [5], (b) 3D ﬁngerprint target printed using TangoPlus FLX 930 [4], (c)
3D ﬁngerprint target printed using TangoBlackPlus FLX980 and then sputter coated with 30 nm
titanium + 300 nm of gold [6], (d) our casted 3D ﬁngerprint target using a mixture of PDMS
(Polydimethylsiloxane) and Pantone 488C color pigment [157] [40], and (e) our casted universal
3D ﬁngerprint target using a mixture of conductive PDMS, silicone thinner, and Pantone 488C
color pigment [157] [154] [158]. 3D targets in (a), (b), and (c) were printed on a high resolution
3D printer (Stratasys Objet350 Connex).

These targets were a big step forward in the direction of realistic operational ﬁngerprint reader

evaluation because the targets employed a 3D geometry similar to the human ﬁnger, they were

fabricated using materials with similar mechanical properties as human skin, they were mapped

with real ﬁngerprint images, and they could be worn on a human ﬁnger. However, due to the

limited number of materials that can be used in 3D printers, the polymers used for printing (i) did

not have the same nominal electrical conductivity of human skin and (ii) did not have the spectral

reﬂectance of human skin. As a result, multiple types of targets (Figs. 2.3 (a), (b), (c)) were

fabricated for different types of ﬁngerprint readers (capacitive, contact-optical, and contactless-

optical) [5] [4] [6]. These individual targets worked for evaluating the type of reader for which

they were designed, however, they were not interoperable. That is, a target fabricated for one type

of ﬁngerprint reader (e.g. capacitive) would not work on a different type of ﬁngerprint reader (e.g.

optical). Because multiple types of targets were needed for evaluating different types of readers,

performing a standardized interoperability evaluation of ﬁngerprint reader technologies was not

possible with these 3D printed targets.

28

Table 2.2 Properties of published 3D Printed Targets [4, 5, 9]

Specimen
Material

Shore A
Hardness

Tensile
Strength
(MPa)

Elongation
at Break

(%)

Color

Electrical
Resistivity

(Ω-cm)

Cost
(USD)

TangoBlackPlus

FLX980

(Fig. 2.3 (a))

[5] [161]
TangoPlus
FLX930

(Fig. 2.3 (b))

[4] [161]

TangoBlackPlus

FLX980,

Ti-Au coating
(Fig. 2.3 (c))
[6] [161] [9]

26-28

0.8-1.5

170-220

Black

Insulator

$10.00

26-28

0.8-1.5

170-220

Translucent

Insulator

$10.00

26-28

0.8-1.5

170-220

Gold

2.4 ∗ 10−5

$12.00

2.1.2 Fingerprint Reader Interoperability

Past studies on ﬁngerprint reader interoperability have shown that when different ﬁngerprint read-

ers were used for enrollment and identiﬁcation (or veriﬁcation), some loss in recognition accuracy

ensued [144] [145] [120]. However, all of these studies were performed on data acquired from live

human subjects [91]. As such, variations (ﬁnger pressure and orientation; conditions of the ﬁnger,

e.g. wet or dry) between impressions on the different readers could account for some of the error

observed. We posit that in order to truly quantify the effects of interoperability, an interoperable

ﬁngerprint target would need to be mounted to a robot gripper and imaged on different readers at

the same pressure and orientation.

As noted in [119], continued advances in distributed computing have enabled less monolithic

ﬁngerprint recognition systems. This advent of larger, more distributed systems (e.g.

the Aad-

haar system) drastically increases the likelihood that the ﬁngerprint reader used to enroll a user’s

ﬁngerprint image at one location will not be the same reader (or model of reader) used later to

identify or verify the same individual at another location. Furthermore, even if the same reader

29

Table 2.3 Properties of our 3D Casted Targets

Specimen
Material

Shore A
Hardness

Tensile
Strength
(MPa)

Elongation
at Break

(%)

Color

Electrical
Resistivity

(Ω-cm)

Cost
(USD)

PDMS,

Pantone 488C
(Fig. 2.3 (d))

[40]

Conductive PDMS,

Thinner,

Pantone 488C
(Fig. 2.3 (e))

[157] [154] [158]

43

6.7

120

PMS 488C

Insulator

$0.86

38.5

2.0

80

Tan +
PMS 488C 9.8 ∗ 10−1 † $10.00

† Although the resistivity of the target differs from human skin, the resistivity value is sufﬁcient
for image capture by capacitive readers.

is used for both enrollment and identiﬁcation, advances in sensing technology could eventually

require replacement of the reader being used. As mentioned in [120], the cost to an institution

needing to re-enroll its entire database of users on a new reader could be monumental. Both of

these situations underscore the need to know and quantify ﬁngerprint reader interoperability. If

ﬁngerprint recognition systems are to continue to become more distributed, then the performance

change associated with interoperability must be objectively known and quantiﬁed. Doing so will

beneﬁt system users, reader manufacturers, system developers, and the institutions deploying the

system.

2.1.3 Universal 3D Fingerprint Targets

To enable robust, standardized ﬁngerprint reader interoperability evaluations, we present the fab-

rication of an interoperable 3D ﬁngerprint target (Fig. 2.1) through a molding and casting process

(Fig. 2.4). We call our target the universal ﬁngerprint target (Fig. 2.3 (e)). Like previous ﬁnger-

print targets in [5], the universal ﬁngerprint targets share a 3D geometry similar to a ﬁngerprint

surface, have mechanical properties similar to human skin, and are mapped with a ﬁngerprint im-

age, either real or synthetic. However, unlike previous ﬁngerprint targets, the universal ﬁngerprint

30

Figure 2.4 System block diagram of the proposed molding and casting process for making 3D tar-
gets. (a) A 3D negative mold (of a 2D ﬁngerprint image) and a supporting scaffolding system (nec-
essary for making the ﬁngerprint target wearable) are electronically fabricated; (b) 3D electronic
models are manufactured by 3D printing and chemical cleaning; (c) conductive silicone, silicone
thinner, and human colored dye are mechanically mixed to produce a casting material with similar
conductive, mechanical, and optical properties to the human skin; (d) the material fabricated in
(c) is cast into the mold and scaffolding system; (d) vacuum degassing ensures that air bubbles
are removed from the casted material; (e) wearable ﬁngerprint targets are extracted 72 hours after
pouring the casting material; (f) the wearable, 3D ﬁngerprint target is used for ﬁngerprint reader
evaluations.

targets are unique in that they incorporate the technically pertinent mechanical, optical, and electri-

cal properties of the human skin within a single target (Tables 2.1, 2.2, and 2.3), making it possible

for the universal ﬁngerprint targets to be imaged by all major ﬁngerprint sensing technologies in

use (capacitive, contact-optical, contactless-optical)1. The universal ﬁngerprint targets enable and

facilitate, for the ﬁrst time, a standardized assessment of ﬁngerprint reader interoperability. The

universal ﬁngerprint targets also enable controlled data collection useful for ﬁngerprint distortion

modeling.

More concisely, the contributions of this chapter are:

• A controlled, repeatable process for creating ﬁngerprint target molds, and fabricating high
quality ﬁnger castings. Unlike previous works [5], this casting fabrication process is not

1We use the term universal to indicate that the targets can be imaged by all existing, major types of ﬁngerprint
readers (contact-optical, contactless-optical, and capacitive). Since the submission of this manuscript, we have also
veriﬁed that the targets can be imaged by Optical Coherence Tomography (OCT), ultrasound, and multispectral ﬁnger-
print readers. If new sensing technologies emerge requiring additional properties in casted targets, our ﬂexible casting
process allows for concocting a new material with the necessary properties. So, our process can be easily extended to
manufacture targets for new ﬁngerprint sensing technologies that may emerge.

31

Figure 2.5 Process ﬂow for fabricating electronic 3D ﬁngerprint mold, M

restricted to a small number of materials. Additionally, it is not cost prohibitive as it is based

on a potentially high-throughout casting process.

• Fabricating high ﬁdelity universal 3D ﬁngerprint targets with similar mechanical, optical,
and electrical properties to the human skin. Previous targets did not simultaneously possess

both the optical and electrical properties of human skin within a single target.

• Fingerprint image capture, using the same 3D target, from optical readers (contact and con-
tactless) and capacitive readers. Our universal ﬁngerprint targets enable standardized inter-

operability data collection for the ﬁrst time ever.

• Experimental evaluations, using the universal 3D ﬁngerprint targets and three different
types of commercial off-the-shelf (COTS) ﬁngerprint readers2 (contact-optical, contactless-

optical, and capacitive). Our results quantify the loss in ﬁngerprint recognition accuracy

when different readers are used for enrollment and identiﬁcation (or veriﬁcation). These

ﬁndings validate the use of our universal 3D ﬁngerprint target for further ﬁngerprint reader

interoperability studies.

2Because of our Non-Disclosure Agreement with the vendors, we cannot provide the make and model of the readers

used in our experiments.

32

2.2 Mold & Scaffold Fabrication

To fabricate a ﬁngerprint target T , we begin by electronically modeling (and subsequently manu-

facturing) a ﬁngerprint mold M and scaffolding framework F .

2.2.1 Mold Fabrication

First, a negative3 ﬁngerprint mold is electronically designed (Fig. 2.5), 3D printed, and chemically

cleaned. This process is further broken down in the following steps.

i) Inner Mold Surface - Using techniques similar to [5], a 2D ﬁngerprint image is mapped onto

a smooth 3D ﬁnger surface mesh S in a manner that retains the topology inherent to the 2D image

(Fig. 2.5 (a)). More formally, let S be a mesh of triangular faces F =

f1, f2, f3,

..., fn

(cid:21)

,

(cid:20)

(cid:20)

(cid:21)

(cid:20)

..., vc

(cid:21)

and 3-dimensional vertices V =

v1, v2, v3,

. Each face in F is explicitly deﬁned

as an ordered list of 3 vertices from V, e.g. f1 =

vi, vj, vk

. Additionally, every face in F

contains a normal vector which is implicitly encoded by the order of the 3 vertices used to deﬁne

the face. In particular, the direction of the normal vector is determined by taking the cross product

of the vectors formed with respect to the order of the face’s three vertices. For example, the normal
vector for face f1 is f1,normal = a × b, where a is a vector having tail at vi and head at vj, while b
is a vector having tail at vj and head at vk.

Because the end goal of the electronic modeling of M is to produce a negative mold, the

mapped surface S must be inverted by ﬂipping all the faces of S (Fig. 2.5 (b)). For every face,

this ﬂipping is attained by reversing the order of its three vertices - and consequently the implicitly

(cid:20)

(cid:21)

encoded direction of its normal vector. For example, by changing f1 =

to ˆf1 =
, the normal vector ˆf1,normal computed by a × b is reversed in direction, since a is

vk, vj, vi
now a vector having tail at vk and head at vj, while b is a vector having tail at vj and head at vi.

vi, vj, vk

ii) Outer Mold Surface - After iteratively inverting all n

f1, f2, f3,

..., fn

faces, the

next step in generating mold M is to imprint the ﬁngerprint surface S inside of an open ended

3In molding and casting, positive sculptures are produced from their negative mold.

33

(cid:20)

(cid:21)

(cid:21)

(cid:20)

cylindrical surface C (Fig. 2.5 (c)). Surface C acts as the exterior of the ﬁnal mold M. As such,

dimensions for C are determined empirically so as to provide strength and durability to the mold

and to prevent usage of excess material. Our experiments show that setting the height of C to
Cheight = 1.25 ∗ Sheight balances the need for structural support and minimizes material cost for
casted targets (here Sheight is the height of the ﬁngerprint surface S). The diameter of the mold

(Cdia) is ﬁxed at 34 mm. While Cdia could have been dynamically chosen based upon the diameter

of S (Sdia), we chose a ﬁxed value so that all the molds we print could ﬁt within a single scaffolding

framework F . We chose 34 mm as a static diameter value, since the 95th percentile of the widest

adult ﬁnger (the thumb) is 26 mm to 27 mm [62]. As such, the minimum thickness (tmin) of our
mold is computed as tmin = 1/2 ∗ (34 − 27)mm = 3.5 mm. We empirically validated that a mold
thickness of tmin ≥ 3.5 mm provides the durability needed for our casting process.

iii) Split Mold - With the inner and outer surface of the mold in place, we continue the fabri-

cation process by simultaneously splitting C and S along the xy-plane into Cabove, Sabove, Cbelow,

and Sbelow. Splitting the mold into two semi-cylindrical components will facilitate the extraction

of the ﬁnal ﬁngerprint castings T (from the mold). Cabove, Sabove, Cbelow, and Sbelow are further

post processed by adding new faces and vertices such that all four surfaces lie ﬂat on the xy-plane.

Figure 2.5 (d) illustrates the sliced, trimmed, and post processed components Cbelow, Cabove, Sbelow,

and Sabove.

iv) Stitching and Printing - Finally, the individual surfaces Cbelow and Sbelow and Cabove and

Sabove are stitched together into two three-dimensional, semi-cylindrical mold halves by adding

triangular faces around the periphery of the respective surfaces. Upon completion of this stitching,

a high ﬁdelity ﬁngerprint mold M has been electronically fabricated (Fig. 2.5 (e)).

To minimize the variability of ﬁngerprint targets during consecutive castings, two “lock” com-

ponents are attached to the bottom of C (Fig. 2.5 (f)). These lock pieces, having length equal to

34 mm (Cdia) will prevent C from rotating inside of the scaffolding framework F .

At this point, M is physically realized by using a high resolution, state-of-the-art 3D printer

that has the ability to print in slices as small as 16 microns [162]. A printer with such ﬁne resolution

34

is necessary to capture the minute details of the mapped ﬁngerprint onto M 4. As in [5], the mold is

printed in 30 micron layers as this captures the necessary detail of the mapped ﬁngerprints, while

simultaneously decreasing the print time of M from 8 hours to 4 hours [5]. At the conclusion of

printing, the mold is soaked in 2M NaOH5 for about 4 hours to dissolve away the support material

from the printed mold in a manner that does not damage the ﬁngerprint ridges. After chemical

cleaning a high ﬁdelity ﬁngerprint mold is ready for casting ﬁngerprint targets (Fig. 2.6).

The resultant mold will only produce a solid casting, since casting material will ﬁll the entire

mold cavity. To make the cast wearable (e.g. mounting to a robotic gripper) or manual evaluation

(e.g. human placement of the target) a “scaffolding framework” F is fabricated, which, when used

in conjunction with M, creates a wearable 3D target T (Fig. 2.7). The process for generating F is

further expounded upon below.

Figure 2.6 (a) High ﬁdelity 3D printed ﬁngerprint mold M. (b) View of ﬁngerprint engraving on
M at 20X magniﬁcation. The magniﬁed image in (b) shows that all the friction ridge patterns are
clearly present in the mold M. These friction ridge patterns are inverted, since negative molds are
necessary to produce positive ﬁngerprint targets (Fig 7 (c)).

4We also experimented with low resolution printers, however, the resolution was insufﬁcient to cleanly separate

from the mold.

the ridges and valleys of a ﬁngerprint pattern.

5NaOH (Sodium Hydroxide) is a basic (alkaline) solution that cleans the residual printing support material away

35

Figure 2.7 3D wearable Universal Fingerprint Target (a) front view, (b) rear view, and (c) view of
the Universal Fingerprint Target ridges at 20X magniﬁcation.

2.2.2 Scaffolding Fabrication

To create a wearable ﬁngerprint casting, a hollow, appropriately shaped void must be cured into

the casted material as it resides in M. This void enables wearability as it creates the space where

an end user’s ﬁnger (or robotic attachment) would reside during evaluation.

We build upon the above idea by developing (based upon the dimensions M) a scaffolding
framework F used to insert a ﬁngerprint surface S(cid:48) (with diameter slightly smaller than Sdia) into
M during successive ﬁngerprint target casts (Fig. 2.8 (a)). In doing so, we ensure that when casting
material is injected into the mold, the space between S and S(cid:48) will be ﬁlled to form a wearable
ﬁngerprint target T .

The scaffolding F consists of several components: a base platform that holds the mold M

in place, two sides extending beyond the top of M, and a top piece from which the ﬁngerprint
surface S(cid:48) is suspended. Aside from S(cid:48), all of these pieces are generated by creating a simple
cuboid shape and applying afﬁne transformations until the component is of the correct size and in

the correct position. The thickness of scaffolding walls is chosen to be 9 mm, which provides the

structural robustness and durability needed for repeated castings of ﬁngerprint targets. In addition,

a concentric rectangular prism is cut from the inside of the base component. The length and width

of this rectangular prism share the same dimension (Cdia) as the diameter of M. This ensures that

36

Figure 2.8 Fabricating scaffolding F using the dimensions of the mold, M. (a) scaffolding frame-
work F is electronically modeled; (b) the electronic scaffolding system is physically generated in
acrylonitrile butadiene styrene (ABS) using a high resolution 3D printer. Using F in conjunction
with M, 3D wearable ﬁngerprint targets T are repeatably produced.

M will attach securely into the base unit, thus controlling the thickness of the casted targets. Since

the diameter of M is ﬁxed (based upon the 95th percentile of the human ﬁnger width at 34 mm),

any mold can be attached interchangeably into a single scaffolding system.

Given that S(cid:48) is a ﬁngerprint surface with a diameter smaller than S, we can derive S(cid:48) from the
same scanned ﬁngerprint surface that we originally used to generate S. That is, given a smooth
scanned 3D ﬁngerprint surface Ssmooth, we can generate S(cid:48) by shrinking Ssmooth along the direc-
is a vertex of Ssmooth and
tion of its normals by 1.5 mm. More formally, if v1 =

(cid:20)

(cid:21)

vx vy vz

(cid:20)

(cid:21)

n1 =

nx ny nz

is the corresponding normal vector to v1, then generating the new vertex

37

(cid:20)

x v(cid:48)
v(cid:48)

y v(cid:48)

z

(cid:21)

v(cid:48) =

for S(cid:48) is computed as:



 −

vx

vy

vz

v(cid:48) =

 × 1.5

(2.2.1)



nx

ny

nz

After all vertices of Ssmooth have been iteratively shrunken along the direction of their correspond-
ing normals, the top of S(cid:48) is stitched shut using a triangle fan6.

As with M, the electronic model of F is 3D printed using the same high resolution printer and

parameters (Fig. 2.8 (b)). F is also cleaned with 2M NaOH solution to remove residual printing

support material. Although F does not have the minute detail that M does, high resolution printing

is still needed for printing F so that registration between F and M is consistent and reproducible.

This ensures the high ﬁdelity of the casted targets is preserved.

Upon completed fabrication of both M and F , we have tools for repeatably casting high ﬁdelity,

3D wearable ﬁngerprint targets T .

2.3 Casting

With tools developed for molding and casting in place, we next discuss the characteristics nec-

essary (to emulate human skin) in the casting material for the 3D Universal Fingerprint Target.

Additionally, we prescribe a process for concocting a material consisting of these characteristics

and subsequently casting the material into a ﬁngerprint mold and scaffolding system.

2.3.1 Material Requirements

Our material selection needs to carefully consider the optical, electrical, and mechanical properties

inherent to the human ﬁnger.

6A triangle fan is a circular mesh surface, formed by placing a center vertex and ﬁlling in the circle with triangles

that all share the center vertex.

38

• Optical Property: Optical readers rely on proper reﬂectance and refraction of light rays
on the human ﬁnger surface to detect a ﬁngerprint. Therefore, the optical properties of

the targets must be similar to that of human skin to be accurately sensed by optical readers.

Materials that are black will improperly absorb all light rays and materials of high reﬂectivity

will improperly scatter all light rays, in both cases preventing targets of these materials from

being imaged by many optical readers.

• Electrical Property: In addition to the color attribute, the targets must also be inherently
conductive to act as a conductive plate and create capacitive differences between ridges and

valleys on the cells within the semiconductor chips on capacitive sensors.

• Mechanical Property: Finally, the mechanical properties of the target material must lie
within the range inherent to the human epidermis to ensure high quality ﬁngerprint target

image acquisition. Materials that deviate from the elasticity of the human epidermis could

negatively impact the target in several ways. If the elasticity is too large, the minute details of

the minutia will be lost as the target is compressed against the sensor and the ridges collapse

under the force being exerted (Fig. 2.9 (a)). If, on the other hand, the elasticity is too small,

or the hardness is too great, the ﬁngerprint target will not ﬂatten around the sensor platen,

resulting in only partial print images of the ﬁngerprint surface (Fig. 2.9 (b)).

2.3.2 Material Fabrication and Casting Procedure

To achieve the electrical, mechanical, and optical criteria necessary for the universal ﬁngerprint

target, electrically conductive silicone (SS-27S) [154] is sheer mixed [58] with silicone thinner

[158] (at 4 % by mass), and a ﬂesh-toned pigment [157] (at 3 % by mass)7. This casting material

mixture is transferred to the mold from a disposable pipet. Prior to the transfer, both the mold and

7A simpler casting material - useful for interoperability assessment of contact and contactless optical readers - can
be fabricated by mixing (with the FlakTek [58]) pure PDMS and PMS 488C pigment. These targets are not conductive,
and are therefore unusable for capacitive reader evaluation, but they are optically and mechanically similar to the
human ﬁnger and are cheaper to manufacture (Fig. 2.3 (d)) (Tables 2.1 and 2.3).

39

(a) 900 % elongation at break

(b) Shore A 50

Figure 2.9 Fingerprint impressions captured from targets lacking proper mechanical characteris-
tics. Notice (a) the presence of aberrations resulting from excessive elasticity in the target and (b)
partial impression due to excessive hardness of the target.

scaffolding system are spray coated with silicone release agent [156]. After the material transfer,

vacuum degassing at 98 kPa (0.97 atm) removes all air bubbles from the material. Finally, the

mold and scaffolding system are assembled and left to cure (Fig. 2.4 (d)). After 72 hours, a high

ﬁdelity, 3D wearable universal ﬁngerprint target, T , can be carefully extracted from the ﬁngerprint

mold and scaffolding system (Fig. 2.4 (e)).

2.3.3 Material Characterization

To verify the optical similarity of our fabricated material to human skin, we obtain a spectro-

gram [137] of the Universal Fingerprint Target material and compare it to a range of human skin

spectrograms obtained by NIST [30] from 51 human subjects (Fig. 2.10). From this spectrogram, it

can be seen that the spectral reﬂectance of the Universal Fingerprint Target material lies within the

range of human skin for almost the entire visible spectrum (400 nm - 700 nm). At approximately

625 nm to 700 nm the Universal Fingerprint Target material does deviate from the range of human

skin (.05 - .1 reﬂectance factor). Based on the NIST report, spectral reﬂectance varies signiﬁcantly

even across multiple readings of the same subject. Furthermore, only 51 subjects were evaluated

to establish the range shown in Figure 2.10. Therefore, it is entirely possible that the Universal

40

Fingerprint Target material does lie within the spectral reﬂectance of human skin from 625 nm to

700 nm as well, given a larger number of subjects.

Figure 2.10 Comparison of the Universal Fingerprint Target material spectrogram to a range of
spectrograms obtained by NIST from 51 human subjects. We plotted the range using estimated
data points from the ﬁgure in [30].

In addition to verifying optical similarity of the material to human skin, we also verify that

the material is electrically conductive by obtaining a resistivity reading (using [61]) from 4 square

samples of the material. The average resistivity of the 4 samples is reported in Table 2.3.

Finally, the mechanical properties of the material are computed (using the data-sheets in [154]

and [158]) and reported in Table 2.3. From the mechanical values reported in Tables 2.1 and 2.3,

it can be seen that the chosen material is indeed within the range of the mechanical properties of

human skin.

2.4 Target Fidelity and Reproducibility

To establish the universal ﬁngerprint targets as standard evaluation artifacts, we must show that the

proposed fabrication process (i) is of high ﬁdelity and (ii) is reproducible. Both of these criterion

are veriﬁed in the following subsections.

41

Figure 2.11 Images of the universal ﬁngerprint target (mapped with circular sine gratings) captured
using a Keyence optical microscope [98]. Point-to-point ridge distances are measured. (a) Image
at 50X magniﬁcation and annotated with 20 point-to-point distances. (b) Image at 100X magniﬁ-
cation and annotated with 10 point-to-point distances. (c) 3-D image generated by the microscope
which qualitatively illustrates the uniformity in ridge height of the circular gratings on the univer-
sal ﬁngerprint target. The granular texture in (a), (b), and (c) is evidence of the aluminum coated
silver particles mixed into the universal ﬁngerprint target which allows the target to be imaged by
capacitive ﬁngerprint readers.

2.4.1 Fidelity

A 3D universal ﬁngerprint target is of high ﬁdelity if its 3D ridges retain the topology inherent to

the original 2D image it was fabricated from. We posit that ﬁdelity of universal ﬁngerprint targets

can be objectively determined and compensated by quantifying the errors (as a deviation of the

3D target topology from the topology of the 2D mapping pattern) at each step in the fabrication

process (Fig. 2.4) and accounting for these errors during fabrication.

(i) Error in Electronic Modeling of Fingerprint Mold - Arora et al. [5] showed that the pro-

jection algorithm used to map the 2D ﬁngerprint image to a 3D ﬁnger surface results in a 5.8 %

decrease in point-to-point distances inherent to the original 2D ﬁngerprint image. Because the

electronic fabrication of the ﬁngerprint mold (Fig. 2.4 (a)) uses the same 2D to 3D projection

algorithm as [5], the same error will be encountered in our universal ﬁngerprint target fabrication

process.

(ii) Error in 3D printing - Arora et al. [5] also observed an 11.42 % decrease in point-to-point

distances (inherent to the original 2D ﬁngerprint image) when fabricating the physical 3D target

on a high resolution 3D printer. Since printing the ﬁngerprint mold in (Fig. 2.4 (b)) was performed

42

using the same printer as [5], the universal ﬁngerprint target fabrication process will encounter the

same error.

While the errors introduced in both electronic projection and 3D printing may seem signiﬁcant,

they can be rectiﬁed (as shown in [6]) by setting the scale during 2D/3D projection from 19.685

pixels/mm to 16.79 pixels/mm. In doing so, the errors introduced during mold modeling (Fig. 2.4

(a)) and 3D mold printing (Fig. 2.4 (b)) are compensated.

(iii) Error in Casting - The ﬁdelity in the universal ﬁngerprint target post casting (Figs. 2.4 (d),

(e)) is validated in the following manner. First, three universal ﬁngerprint target castings are fab-

ricated using three different molds; each mapped with different 2D calibration patterns (vertical,

horizontal, and circular sine gratings with a frequency of 10 pixels). At a projection scale of 16.79

pixels/mm (at 500 ppi) and the reduction in point-to-point distances during electronic modeling

and 3D printing, 10 pixel ridge distances on the calibration pattern should correspond to an actual

ridge distance of 0.508 mm on the casted calibration target. Using an optical microscope, 5 images

of each universal ﬁngerprint target are captured at both 50X magniﬁcation and 100X magniﬁca-

tion (Fig. 2.11) [98]. A software tool available with the optical microscope is used to mark 20

point-to-point ridge distances at 50X magniﬁcation and 10 point-to-point ridge distances at 100X

magniﬁcation in all the acquired optical microscope images. The microscope software was cali-

brated using a micrometer resolution calibration target. Table 2.4 shows the average point-to-point

ridge distances at both magniﬁcations for all 3 casted targets. In comparison to the ground truth

distance of 0.508 mm, the optical microscope reveals the empirical mean point-to-point ridge dis-

tances to be 0.499 mm, attributing to a 1.8 % reduction in point-to-point distances on the universal

ﬁngerprint target during casting. This reduction of 1.8 % in point-to-point ridge distances is not

unexpected, since the conductive silicone used to fabricate the universal ﬁngerprint targets is esti-

mated to shrink by 2 % during vulcanization. Again, this error can be compensated by adjusting

the projection scale during 2D/3D mapping.

In addition to measuring point-to-point distances, we also measure the height of the ridges

on the casted targets using a high resolution proﬁlometer [2]. The ridge height of the ﬁngerprint

43

Table 2.4 Average point-to-point ridge distances observed on universal ﬁngerprint targets, mea-
sured using the Keyence Optical Microscope at 50X and 100X magniﬁcation. The expected point-
to-point ridge distance is 0.508 mm. (standard deviation is recorded in parenthesis).

Calibration Pattern
Vertical Gratings
Horizontal Gratings
Circular Gratings

50X Magniﬁcation
0.509 mm (.031)
0.501 mm (.026)
0.513 mm (.029)

100X Magniﬁcation

0.496 mm (.023)
0.490 mm (.028)
0.486 mm (.035)

targets is set to 0.33 mm during electronic projection. Due to mold shrinkage during 3D printing,

we expect the ridge height of the casted targets to be 0.29 mm. The measurements obtained by

the proﬁlometer show all ridge heights to be 0.16 mm. This further reduction in ridge height is

not unexpected since a thin coating of release agent is ﬁrst applied to the mold prior to casting.

Furthermore, the reduction in ridge height is beneﬁcial as it brings the ridge height of the targets

even closer in value to the human ﬁnger ridges at 0.06 mm. Note, the ridge height had to be set

to 0.33 mm during electronic projection due to current limitations in state-of-the-art 3D printing

resolution. Future study could explore novel techniques for fabricating the mold which enable even

higher resolution than 3D printing.

(iv) End-to-end Error- In this ﬁnal error analysis, the full, end-to-end fabrication process

is scrutinized. More speciﬁcally, an experiment is conducted which demonstrates that features

present on a 2D ﬁngerprint image are preserved after converting the 2D ﬁngerprint image into a

wearable, 3D, universal ﬁngerprint target.

To conduct this experiment, six different universal ﬁngerprint target molds are fabricated using

six ﬁngerprint images from the NIST SD4 database [129]. Subsequently, six universal ﬁngerprint

targets are cast from the ﬁngerprint molds. Finally, comparison scores are generated between the

NIST SD4 rolled ﬁngerprint images and 2D ﬁngerprint images acquired from the corresponding

six universal ﬁngerprint targets. Fingerprint images of the universal ﬁngerprint targets are obtained

using an Appendix F certiﬁed, 500 ppi, contact-optical reader, a PIV certiﬁed, 500 ppi, contactless-

optical reader, and a PIV certiﬁed, 500 ppi, capacitive reader. Figure 2.12 illustrates corresponding

minutia points between a NIST SD4 rolled ﬁngerprint image and a ﬁngerprint image acquired from

44

Figure 2.12 Comparing the source ﬁngerprint image to the image of the corresponding universal
ﬁngerprint target. (a) NIST SD4 S0083 rolled ﬁngerprint image is compared to (b) a universal
ﬁngerprint target image; (b) is fabricated using (a) and imaged using an Appendix F certiﬁed,
optical, 500 ppi ﬁngerprint reader. A similarity score of 608 is computed between (a) and (b)
using Veriﬁnger 6.3 SDK (threshold is 33 at FAR=0.01 %). The minutia points in correspondence
between (a) and (b) are shown.

its corresponding universal ﬁngerprint target. Table 2.5 reports similarity scores for each of the six

universal ﬁngerprint targets in comparison to the NIST SD4 rolled print used to fabricate them.

The key ﬁndings of this experiment are as follows:

• The corresponding minutia points between images captured using the universal ﬁngerprint
targets and the images used to generate each target (Fig. 2.12) show that salient 2D features

inherent to the NIST rolled ﬁngerprint images are retained following their fabrication into a

universal ﬁngerprint target.

• The universal ﬁngerprint targets (Table 2.5) almost always outperform previous 3D optical
targets [5] (Table 2.6) by achieving higher similarity scores between the ﬁnished 3D target

images and the ground truth image used to fabricate the respective target. Furthermore,

the universal ﬁngerprint targets perform comparably to goldﬁngers [6] on capacitive readers

(Tables 2.5 and 2.6).

• Unlike past research in 3D ﬁngerprint targets, the universal ﬁngerprint target achieves com-
parison scores on contactless-optical readers well above the acceptance threshold8 of 33. We

8We use Veriﬁnger 6.3 which has a threshold of 33 at a FAR=0.01%.

45

Table 2.5 Universal Fingerprint Target Similarity Scores1 (SD4 ﬁnger-
print image vs. corresponding target image). Proposed Targets.

SD4 Fingerprint

Contact

Contactless

Optical (500 ppi)

Optical (500 ppi)

Capacitive
(500 ppi)

S0005
S0010
S0031
S0044
S0068
S0083

584
539
600
498
327
608

152
137
105
150
146
176

161
305
221
323
368
323

1 Veriﬁnger 6.3 SDK was used for generating similarity scores. The
score threshold at 0.01 % FAR is 33. Veriﬁnger was chosen so that com-
parisons could be made between the universal ﬁngerprint targets and
previous studies [5] [4] [6]

do note that the universal ﬁngerprint targets achieve lower comparison scores against the SD4

images when using the contactless-optical reader as opposed to the contact-optical reader for

image acquisition. One plausible explanation is that the universal ﬁngerprint targets have a

ridge height greater than the ridge height of the adult human ﬁnger. This discrepancy may

cause errors as the contactless-reader unrolls a 3D ﬁngerprint into a 2D ﬁngerprint image.

In summary, the 2D ground truth ﬁngerprint features are found to be preserved during fabrica-

tion into a 3D universal ﬁngerprint target and subsequent image acquisition (with high accuracy)

by contact-based optical readers, contactless-optical readers, and capacitive readers, as evidenced

by the high minutiae-based match scores.

2.4.2 Reproducibility

In the previous section, the fabrication process for creating universal ﬁngerprint targets was quan-

titatively shown to be of high ﬁdelity. One remaining criterion that must be objectively veriﬁed

to solidify the use of universal ﬁngerprint targets as standardized evaluation artifacts is the re-

46

Table 2.6 3D Printed Target1 Similarity Scores (SD4 ﬁngerprint
image vs. corresponding target image). Targets from [4–6].

SD4 Fingerprint

Contact-Optical Reader

Capacitive Reader

(500 ppi) [4, 5]

(500 ppi) [6]

S0005
S0010
S0031
S0044
S0068
S0083

719
129
N/A
371
N/A
441

471
333
N/A
N/A
N/A
183

1 These targets were fabricated using processes reported in [4–6].
They are not interoperable across optical and capacitive readers as
are the Universal Fingerprint Targets.

producibility of high ﬁdelity universal ﬁngerprint target fabrication. To that end, we individually

examine the reproducibility of each step in the universal ﬁngerprint target fabrication process.

The electronic model of the universal ﬁngerprint target mold and scaffolding system can be

easily reproduced by simply executing a program. Additionally, the mold and scaffolding system

can be physically reproduced via 3D printing with accuracy as high as 20 microns [162]. Therefore,

the only step in the universal ﬁngerprint target fabrication process that must still be veriﬁed as

reproducible is the casting step.

To demonstrate reproducibility in casting, 12 universal ﬁngerprint targets are fabricated from

6 ﬁngerprint molds. The 12 universal ﬁngerprint targets correspond to 6 different targets each

fabricated 2 times (with a time lapse of several weeks between target replication). Each mold is

mapped with one of 6 NIST SD4 rolled ﬁngerprint images (S0005, S0010, S0031, S0044, S0068,

and S0083). Let the two sets of universal ﬁngerprint targets be formally deﬁned as T1 and T2,

where T1 is the ﬁrst set of castings and T2 is the set of castings produced several weeks later.

Next, the average and standard deviation of genuine scores between 10 impressions from each

target in the two target sets T1 and T2 collected on 3 types of ﬁngerprint readers (COR A, CLOR,

and CPR A (Table 2.8)) and the corresponding ﬁngerprint image in NIST SD4 are computed using

47

Table 2.7 Universal Fingerprint Target Genuine Similarity Scores1 (SD4 Fin-
gerprint Image vs. Corresponding Target Image) Mean and (Standard Devi-
ation) of Scores for 10 Impressions are Reported

Target Set Reader S0005 S0010 S0031 S0044 S0068 S0083
254.5
(20.1)
248.7
(8.4)
172.8
(17.8)
172.0
(11.5)
190.9
(17.1)
194.0
(6.7)

COR A 212.2
(20.0)
COR A 247.4
(10.3)
203.9
(15.8)
205.1
(14.1)
CPR A 163.2
(21.8)
CPR A 188.1
(16.7)

177.0
(23.1)
230.6
(16.0)
154.3
(11.8)
143.4
(15.7)
141.3
(22.7)
173.3
(10.4)

141.3
(9.0)
166.5
(10.0)
169.9
(17.3)
170.5
(22.2)
121.3
(14.2)
156.5
(7.0)

204.4
(19.3)
226.8
(17.3)
140.6
(7.8)
150.7
(10.3)
177.1
(14.5)
194.4
(21.6)

200.3
(15.8)
207.0
(9.2)
127.5
(15.2)
134.3
(23.0)
128.6
(25.2)
183.8
(16.5)

CLOR

CLOR

T1

T2

T1

T2

T1

T2

1 Innovatrics matcher was used to generate similarity scores. The threshold
of the matcher at FAR = 0.01 % was computed to be 49 on the FVC 2002
and 2004 databases [108, 109].

the Innovatrics ﬁngerprint SDK9 [80]. The averages and standard deviations of genuine similarity

scores between target impressions from each target in T1 and its corresponding ﬁngerprint image in

NIST SD4 are formally deﬁned as GS1. Conversely, GS2 is deﬁned as the averages and standard

deviations of genuine similarity scores between target impressions for each target in T2 and its

corresponding ﬁngerprint image in NIST SD4.

By analyzing the means of the similarity scores in GS1 and GS2, reproducibility in casting

universal ﬁngerprint targets is veriﬁed. In particular, by showing that the means of the similarity

scores in GS1 and GS2 are all well above the genuine acceptance threshold, we demonstrate that

targets (from multiple castings) in T1 and T2 are all of high ﬁdelity, since impressions from both

sets of targets (on multiple types of ﬁngerprint readers) achieve high similarity scores against the

9We use the Innovatrics ﬁngerprint SDK since we recently acquired this matcher, and it is shown to have high
accuracy. Mention of any products or manufacturers does not imply endorsement by the authors or their institutions
of these products or their manufacturers.

48

ground truth images (SD4) from which they were fabricated. The means and standard deviations

of the genuine similarity scores in GS1 and GS2 are reported in Table 2.7.

We note that the means of all similarity scores in GS1 and GS2 are within 0.72 % when using

the contactless ﬁngerprint reader for image acquisition (Table 2.7). This indicates high similarity

between 3D ﬁngerprint topologies on targets in T1 and T2. Additionally we note that the means

of similarity scores in GS1 and GS2 differ slightly when using contact based ﬁngerprint readers

for image acquisition. This is not surprising since the targets in T1 were fabricated with smaller

amounts of silicone thinner than the targets in T2. As such, the softer targets in T2 morphed around

the ﬁngerprint reader platen more than the targets in T1 and produced images with larger friction

ridge area and number of minutia (recall Fig. 2.9 (b)). Subsequently, the larger ﬁngerprint images

acquired from targets in T2 achieved higher match scores against SD4 images than ﬁngerprint im-

ages acquired from targets in T1. This ﬁnding underscores one of the key advantages of contactless

ﬁngerprint readers. In particular, it shows that contactless readers are robust to small mechanical

variations in human ﬁnger epidermis.

Table 2.8 Speciﬁcations of the Fingerprint Readers Used in Our
Experiments

Reader

NDA Alias1
COR A

COR B

Contact-Optical
Contact-Optical

CLOR

Contactless-Optical

CP R A

CP R B

Capacitive
Capacitive

Reader Type

Resolution Certiﬁcations

500 ppi
500 ppi
500 ppi
500 ppi
500 ppi

Appendix F
Appendix F

PIV
PIV
PIV

1 Because of a Nondisclosure agreement (NDA) with our vendors,
we do not release the names of the ﬁngerprint readers.

49

Reader Type

S0005

S0010

S0083

Circular
Gratings

Horizontal
Gratings

Vertical
Gratings

Contact
Optical

Contactless

Optical

Capacitive

Figure 2.13 Example ﬁngerprint impressions from 6 universal ﬁngerprint targets (one per column)
on 3 types of ﬁngerprint readers.

2.5 Experiments

With the ﬁdelity and reproducibility of the universal ﬁngerprint target fabrication process estab-

lished, multiple experiments are performed on all three major types of ﬁngerprint readers using uni-

versal ﬁngerprint targets as operational evaluation targets. First, three ﬁngerprint readers (COR A,

CLOR, and CPR A (Table 2.8)) are individually assessed using three different universal ﬁnger-

print targets mapped with controlled calibration patterns (horizontal gratings, vertical gratings,

and circular gratings). Next, the same three ﬁngerprint readers are individually evaluated using

impressions acquired from ﬁngerprint targets in T2. Finally, a ﬁngerprint reader interoperability

study is performed by comparing images acquired from one of three reader types (contact-optical,

contactless-optical, and capacitive) against images acquired from another of the three reader types.

50

Table 2.9 Mean (µ) and std. deviation (σ) of center-to-center ridge spacings (in
pixels) on images acquired from 3 universal ﬁngerprint targets. The expected
ridge spacing is 9.8 pixels.

Sine Gratings

Pattern

Circular

Contact Optical

(COR A)
µ = 9.50
σ = 0.56

Contactless Optical

(CLOR)
µ = 8.99
σ = 0.06

Horizontal

Vertical

µ = 9.21
σ = 0.65

µ = 8.90
σ = 0.88

µ = 8.94
σ = 0.16

µ = 7.63
σ = 0.51

Capacitive (CPR A)

µ = 9.75
σ = 0.12

µ = 9.45
σ = 0.10

µ = 9.17
σ = 0.09

2.5.1 Evaluating Readers with Calibration Patterns

To evaluate the directional imaging capability of ﬁngerprint readers, we design a similar experi-

ment to that which is proposed in [5]. In particular, we collect 10 impressions on 3 different types

of ﬁngerprint readers using 3 different universal ﬁngerprint targets mapped with controlled cal-

ibration patterns (example impressions shown in Fig. 2.13). Then, using the method in [74] the

average ridge-to-ridge spacing (in pixels) is computed for the captured impressions. Unlike the tar-

gets proposed in [4–6] which could only perform directional assessment of one type of ﬁngerprint

reader, our proposed universal ﬁngerprint targets are capable of performing directional assessment

on contact-optical, contactless-optical, and capacitive ﬁngerprint readers alike. Therefore, in Table

2.9, we report the average ridge-to-ridge spacing of the 3 different universal ﬁngerprint targets

across all three of the major ﬁngerprint reader types. For comparison to previous work [4–6], the

ridge spacing values acquired from sine grating mapped, 3D printed targets (using several fabrica-

tion processes) are reported in Table 2.10.

All three of the calibration patterns that were mapped to universal ﬁngerprint targets have a

10 pixel peak-to-peak frequency. Given our earlier ﬁndings of an approximately 2 % decrease

in point-to-point distances on the universal ﬁngerprint targets during fabrication (due to silicone

shrinkage), ridge-to-ridge distances on the 3 calibration mapped universal ﬁngerprint targets are

51

Table 2.10 Mean (µ) and std. deviation (σ) of center-to-center ridge spacings (in
pixels) on images acquired from 3D printed targets1. The expected ridge spacing
is 8.28 pixels.

Sine Gratings

Pattern

Circular

Horizontal

Vertical

Contact Optical [5] Contactless Optical [4] Capacitive [6]2

µ = 8.92
σ = 0.04

µ = 8.31
σ = 0.10

µ = 8.87
σ = 0.08

µ = 8.12
σ = 0.16

N/A

N/A

N/A

N/A

N/A

1 These targets were fabricated using processes reported in [4–6]. They are not
interoperable across optical and capacitive readers as are the Universal Finger-
print Targets.
2 No ridge spacing results were reported for sine grating mapped gold ﬁngers
in [6].

expected to be 9.8 pixels. Given this ground truth value and the results of Table 2.9, we can evaluate

the three types of ﬁngerprint readers used in this experiment.

The summary of our ﬁndings are as follows:

• Similar to the ﬁndings of [5], impressions of targets mapped with circular gratings have
larger ridge-to-ridge spacing than impressions of targets mapped with horizontal or vertical

gratings. As noted in [5], this is likely due to the radial ﬂattening of the target with circular

gratings as it is applied with pressure to the ﬁngerprint reader platen. This radial ﬂatten-

ing results in larger ridge-to-ridge spacing than the ﬂattening of the horizontal and vertical

calibration targets.

• Unlike the ﬁndings of [5], all of the captured impressions of universal ﬁngerprint targets
have smaller ridge-to-ridge spacing than the expected ridge-to-ridge spacing. In [5] a larger

than expected ridge-to-ridge spacing was explained as a result of ridge-to-ridge distance

expansion during the ﬂattening of the target against the reader platen. We hypothesize that

universal ﬁngerprint targets have smaller ridge-to-ridge expansion during contact with the

reader platen than [5] since universal ﬁngerprint targets are less elastic than the targets in [5].

52

Universal ﬁngerprint targets are closer in elasticity to the human skin than [5] and so the

results shown in Table 2.9 are more indicative of the ridge-to-ridge spacing the readers used

in this study are able to capture from real human ﬁngers.

• Consistent with the ﬁndings of [4], the ridge-to-ridge distances are smaller on the contactless
ﬁngerprint reader than on the contact ﬁngerprint readers. In particular, the captured ridge-

to-ridge spacing of the vertical gratings was lower than expected. We hypothesize that the

ridge-to-ridge spacing on the contactless reader is smaller due to the fact that no distortion

occurs during image acquisition (as no pressure is applied onto a reader platen). Further

analysis needs to be undertaken to understand why the vertical gratings deviated most from

the expected ridge spacing.

Table 2.11 Mean (µ) and std. deviation (σ) of center-to-center ridge spacings (in
pixels) on images acquired from 6 universal ﬁngerprint targets. Expected ridge
spacing (in pixels) for each target is reported in parenthesis

SD4 Fingerprint

S0005 (9.25 )

S0010 (9.98)

S0031 (10.37)

S0044 (9.07)

S0068 (9.48)

S0083 (10.23)

Contact Optical

(COR A)
µ = 8.77
σ = 1.17

Contactless Optical

(CLOR)
µ = 8.77
σ = 0.31

µ = 9.87
σ = 1.46

µ = 10.02
σ = 1.40

µ = 8.49
σ = 1.24

µ = 9.60
σ = 1.29

µ = 9.70
σ = 1.23

µ = 9.52
σ = 0.29

µ = 9.04
σ = 0.37

µ = 8.25
σ = 0.18

µ = 9.18
σ = 0.29

µ = 8.16
σ = 0.15

Capacitive (CPR A)

µ = 9.01
σ = 0.18

µ = 10.42
σ = 0.41

µ = 10.45
σ = 0.28

µ = 9.04
σ = 0.23

µ = 9.86
σ = 0.19

µ = 10.23
σ = 0.14

53

2.5.2 Evaluating Readers with Fingerprint Patterns

Similar to the previous experiment, we conduct an analysis of the ridge-to-ridge distances captured

by three of the major ﬁngerprint reader types. However, in this experiment, rather than mapping

controlled calibration patterns to universal ﬁngerprint targets, we use the targets from T2 which

are each mapped with real ﬁngerprint images from SD4. In doing so, we evaluate the readers with

targets very similar to the real ﬁngers the readers will see in an operational setting.

Again, 10 impressions are captured on all 3 ﬁngerprint readers, this time with each of the 6 uni-

versal ﬁngerprint targets in T2 (example impressions shown in Fig. 2.13). Then, using the method

in [74], the average ridge spacing of the captured impressions is computed (Table 2.11). Addition-

ally, the average ridge spacing is computed (using the method in [74]) on the original ﬁngerprint

images from SD4 and established as the ground truth ridge spacing values. By comparing these

ground truth values with the results of Table 2.11, we perform an assessment of the three ﬁnger-

print readers. Finally, for comparison to previous work [4–6], the ridge spacing values acquired

from ﬁngerprint mapped, 3D printed targets (using several fabrication processes) are reported in

Table 2.12.

In summary, the ﬁndings of this experiment are as follows:

• Consistent with the ﬁndings of our previous experiment with calibration pattern mapped
universal ﬁngerprint targets, the images captured by the contactless-optical ﬁngerprint reader

have smaller ridge-to-ridge distances than the impressions captured by contact based readers.

This is likely due to the absence of ﬁngerprint distortions in contactless ﬁngerprint readers.

Additionally, errors in the contactless reader may be introduced when the three-dimensional

ﬁnger surface captured by the reader is projected into two dimensions (due to the ridge height

of universal ﬁngerprint targets being greater than the ridge height of human ﬁngers).

• In almost all of the target impressions, capacitive ﬁngerprint readers captured the ridge-to-
ridge distances more closely to ground truth than contact-optical readers did. Further studies

54

Table 2.12 Mean (µ) and std. deviation (σ) of center-to-center ridge spacings (in
pixels) on images acquired from 3D printed ﬁngerprint targets1. Expected ridge
spacing (in pixels) for each target is reported in parenthesis

SD4 Fingerprint Contact Optical [5] Contactless Optical [4]2 Capacitive [6]

S0005

S0010

S0083

µ = 8.49
σ = 0.10

(7.82)

µ = 9.22
σ = 0.16

(8.43)

µ = 9.10
σ = 0.19

(8.62)

N/A

N/A

N/A

µ = 9.57
σ = 0.14

(9.45)

µ = 10.34
σ = 0.21
(10.20)

µ = 10.60
σ = 0.14
(10.44)

1 These targets were fabricated using different processes reported in [4–6]. They
are not interoperable across optical and capacitive readers as are the Universal Fin-
gerprint Targets.
2 No ridge spacing results were reported for ﬁngerprint mapped targets imaged by
a contactless reader in [4].

and analysis need to be performed to determine if this ﬁnding is consistent, and also, the

explanation behind this.

2.5.3 Reader Interoperability Evaluations

Whereas our previous two experiments with universal ﬁngerprint targets evaluated the three major

types of ﬁngerprint readers individually, in this ﬁnal experiment, we perform ﬁngerprint reader

interoperability evaluations using the universal ﬁngerprint targets.

To set up this experiment, 10 impressions from each target in T2 are captured on 5 different

ﬁngerprint readers (Table 2.8). Then, for all pairs of ﬁngerprint readers in our set of 5 readers,

images from one reader are used as enrollment images and images from the other reader are used

as probe images to generate genuine and imposter scores using the Innovatrics matcher [80]. In

Table 2.13, we report the means of the genuine and imposter scores. Additionally we report the

True Accept Rate (TAR) and the False Accept Rate (FAR) of the scores using a threshold of 49

55

Table 2.13 Genuine and Imposter Score1 Statistics and Matching Performance Measures when
Comparing Fingerprint Images Acquired from Different Types of Fingerprint Readers. Mean of
Genuine Scores (µG), Mean of Imposter Scores (µI), True Accept Rate (TAR) and False Accept
Rate (FAR2) are Reported.

Probe Image

Fingerprint Readers

COR A

COR B

CLOR

CPR A

CPR B

µG = 440.7,

µG = 399.6,

µG = 182.2,

µG = 276.0,

µG = 202.4,

µI = 0.5

µI = 0.3

µI = 1.2

T AR = 100%
µG = 399.3,

T AR : 100%
µG = 438.1,

T AR : 100%
µG = 171.3,

µI = 0.2

µI = 0.5

µI = 1.9

T AR : 100%
µG = 278.5,

µI = 1.6

Enrollment

Reader

COR A

COR B

CLOR

CPR A

CPR B

µI = 0.3

T AR : 100%
µG = 183.8,

µI = 1.4

T AR : 100%
µG = 271.1,

µI = 0.8

T AR : 100%
µG = 196.4,

µI = 2.3

T AR : 100%
µG = 174.3,

T AR : 99.8%
µG = 334.1,

T AR : 100%
µG = 154.1,

µI = 0.5

µI = 9.0

µI = 2.1

T AR : 100%
µG = 274.7,

T AR : 100%
µG = 147.0,

T AR : 99.8%
µG = 353.0,

µI = 0.8

µI = 1.6

T AR : 100%
µG = 195.4,

T AR : 99.7%
µG = 105.4,

µI = 2.2

µI = 3.2

µI = 7.6

T AR : 100%
µG = 268.2,
µI = 10.1

T AR : 100%

µI = 4.8

T AR : 100%
µG = 200.0,

µI = 4.3

T AR : 100%
µG = 113.2,

µI = 4.6

T AR : 94.8%
µG = 269.4,
µI = 12.1

T AR : 100%
µG = 277.5,
µI = 14.4

T AR : 100%

T AR : 100%

T AR : 100%

T AR : 91.8%

1 Innovatrics matcher was used to generate similarity scores. The threshold of the matcher at
FAR = 0.01 % was computed to be 49 on the FVC 2002 and 2004 databases [108, 109].
2 The False Accept Rate in all cases was 0.0%

(this threshold was precomputed on the FVC 2002 and 2004 databases [108, 109], because we do

not have a sufﬁcient number of images from the targets to set the threshold).

Although the performance results of Table 2.13 seem to indicate that all of the readers used are

highly interoperable, these results are likely too optimistic as only 6 different targets were used. For

this reason, we also report the genuine and imposter score means to show how the scores deteriorate

when different readers are used for enrollment and veriﬁcation. Similar to the ﬁndings of past

ﬁngerprint reader interoperability studies [120,144,145], we note that genuine scores decrease and

imposter scores increase when different ﬁngerprint readers are used to acquire enrollment images

and probe images, especially when the two readers use different sensing technology to acquire

56

images. While past studies reported these ﬁndings using real ﬁngers for data collection, we report

the same ﬁndings, for the ﬁrst time ever, using realistic, 3D, wearable, ﬁngerprint targets. By

demonstrating the same results as past studies with the universal ﬁngerprint targets, we validate

the utility in using universal ﬁngerprint targets for advancing ﬁngerprint reader interoperability

studies. In particular, the universal ﬁngerprint targets could be mounted to a robot and imaged on

different readers at known pressure and orientation. This standardized data could then be used to

learn calibration mappings between different ﬁngerprint readers which could be used to improve

ﬁngerprint reader interoperability.

2.6 Summary

We have designed a molding and casting system capable of fabricating wearable, 3D ﬁngerprint tar-

gets from a plethora of casting materials. By selecting a casting material with similar mechanical,

optical, and electrical properties to the human skin, we cast universal ﬁngerprint targets, which can

be imaged on the three major ﬁngerprint reader types in use (contact-optical, contactless-optical,

and capacitive). Previous studies were unable to produce a single 3D ﬁngerprint target which could

be imaged on multiple types of ﬁngerprint readers. We demonstrate that the process for fabricat-

ing universal ﬁngerprint targets is of high ﬁdelity, and that it is reproducible. Finally, we use the

universal ﬁngerprint targets as evaluation targets on multiple types of PIV/Appendix F certiﬁed

ﬁngerprint readers. Our results verify the utility in using the universal ﬁngerprint targets for both

individual ﬁngerprint reader assessments and also ﬁngerprint reader interoperability studies. We

believe that the universal 3D ﬁngerprint targets introduced here will advance state of the art in

ﬁngerprint reader evaluation and interoperability studies.

In the future, the universal ﬁngerprint targets will be mounted to a robotic hand and imaged

on various ﬁngerprint readers at known pressure and orientation. With this data, objective evalu-

ations can be performed on ﬁngerprint readers. Additionally, the data collected could be utilized

to learn ﬁngerprint distortion models, ﬁngerprint reader interoperability calibration models, and

57

latent ﬁngerprint distortion models. Finally, the universal ﬁngerprint targets will be used to assess

the spooﬁng vulnerability of various ﬁngerprint recognition systems (such as smartphones).

2.7 Acknowledgment

This research was supported by grant no. 70NANB15H280 from the NIST Measurement Science

program. The views and conclusions contained herein are those of the authors and should not be

interpreted as necessarily representing the ofﬁcial policies, either expressed or implied, of the U.S.

Government.

The authors would like to thank Brian Wright, Michigan State University, for his help in 3D

printing of molds. We would also like to thank Edward Drown, Michigan State University, for his

help mixing castings materials for the universal ﬁngerprint targets. Finally, we would like to thank

Nick Paulter for his valuable insights throughout the research.

58

Chapter 3

RaspiReader: Open Source Fingerprint

Reader

In the previous chapter, we demonstrated a manufacturing process for creating high-ﬁdelity, real-

istic, universal, 3D ﬁngerprint targets. While we proposed the use of these targets for evaluating

ﬁngerprint readers, it has been shown that hackers can use ﬁngerprint targets, like the Universal

Target, to “spoof” ﬁngerprint recognition systems. That is, hackers can use fake ﬁngerprints (also

known as presentation attacks, or spoofs) to impersonate a victim or to obfuscate their own iden-

tity. To thwart such attacks, in this chapter, we develop a ﬁngerprint reader, called RaspiReader,

with the built-in capability of automatically detecting and ﬂagging ﬁngerprint spoof attacks prior to

performing recognition [47]. RaspiReader is an open-source, easy to assemble, high resolution, op-

tical ﬁngerprint reader, built entirely from ubiquitous components. More importantly, RaspiReader

is specially customized with two cameras for ﬁngerprint image acquisition. One camera provides

high contrast, frustrated total internal reﬂection (FTIR) ﬁngerprint images, and the other outputs

direct images of the ﬁnger in contact with the platen. Using both of these image streams, we extract

complementary information which, when fused together and used for spoof detection, results in

marked performance improvement over previous methods relying only on grayscale FTIR images

provided by COTS optical readers. Fingerprint matching experiments between images acquired

59

Figure 3.1 Prototype of RaspiReader: two ﬁngerprint images (b, (i)) and (b, (ii)) of the input ﬁnger
(a) are captured. The raw direct image (b, (i)) and the raw, high contrast FTIR image (b, (ii)) both
contain useful information for spoof detection. Following the use of (b, (ii)) for spoof detection,
image calibration and processing are performed on the raw FTIR image to output a high quality,
500 ppi ﬁngerprint for matching (b, (iii)). The dimensions of the RaspiReader shown in (a) are 100
mm x 100 mm x 105 mm (about the size of a 4 inch cube).

from the FTIR output of RaspiReader and images acquired from a COTS reader verify the inter-

operability of the RaspiReader with existing COTS optical readers. By using our open source STL

ﬁles and software, RaspiReader can be built in under one hour for only US $175.

3.1

Introduction

In an effort to mitigate the costs associated with ﬁngerprint spoof attacks, a number of spoof

detection techniques involving both hardware and software have been proposed in the literature.

Special hardware embedded in ﬁngerprint readers1 enables capture of features such as heartbeat,

1Several ﬁngerprint vendors have developed hardware spoof detection solutions by employing multispectral imag-
ing, infrared imaging (useful for sub-dermal ﬁnger analysis), and pulse capture to distinguish live ﬁngers from spoof
ﬁngers [67, 130].

60

Figure 3.2 Fingerprint images acquired using the RaspiReader. Images in (a) were collected from
a live ﬁnger. Images in (b) were collected from a spoof ﬁnger. Using features extracted from
both raw image outputs ((i), direct) and ((ii), FTIR) of the RaspiReader, our spoof detectors are
better able to discriminate between live ﬁngers and spoof ﬁngers. The raw FTIR image output of
the RaspiReader (ii) can be post processed (after spoof detection) to output images suitable for
ﬁngerprint matching. Images in (c) were acquired from the same live ﬁnger (a) and spoof ﬁnger
(b) on a commercial off-the-shelf (COTS) 500 ppi optical reader. The close similarity between
the two images in (c) qualitatively illustrates why current spoof detectors are limited by the low
information content, processed ﬁngerprint images (c, (iii)) output by COTS readers.

thermal output, blood ﬂow, odor, and sub-dermal ﬁnger characteristics useful for distinguishing a

live ﬁnger from a spoof [8,100,113,130,146,151–153,176]. Spoof detection methods in software

are based on extracting textural [65, 66, 69, 70, 127], anatomical [64], and physiological [1, 114]

features from processed2 ﬁngerprint images which are used in conjunction with a classiﬁer such

as Support Vector Machines (SVM). Alternatively, a Convolutional Neural Network (CNN) can be

trained to distinguish a live ﬁnger from a spoof [26, 27, 117, 131].

While existing hardware and software spoof detection schemes provide a reasonable starting

point for solving the spoof detection problem, current solutions have a plethora of shortcomings.

As noted in [151,152,176] most hardware based approaches can be easily bypassed by developing

very thin spoofs (Fig. 3.3 (a)), since heartbeat, thermal output, and blood ﬂow can still be read from

2Raw ﬁngerprint images are “processed” (such as RGB to grayscale conversion, contrast enhancement, and scaling)
by COTS readers to boost matching performance. However, useful spoof detection information (such as color and/or
minute textural abberations) is lost during this processing.

61

Figure 3.3 Example spoof ﬁngers and live ﬁngers in our database. (a) Spoof ﬁngers and (b) live
ﬁngers used to acquire both spoof ﬁngerprint impressions and live ﬁngerprint impressions for
conducting the experiments reported in this thesis. The spoofs in (a) and the live ﬁngers in (b)
are not in 1-to-1 correspondence.

the live human skin behind the thin spoof. Additionally, some of the characteristics (such as odor

and heartbeat) acquired by the hardware vary tremendously amongst different human subjects,

making it very difﬁcult to build an adequate model representative of all live subjects [152, 176].

Current spoof detection software solutions have their own limitations. Although the LivDet

2015 competition reported state-of-the-art spoof detection software to have an average accuracy

of 95.51% [121], the spoof detection performance at desired operating points such as False Detect

Rate (FDR) of 0.1% was not reported, and very limited evaluation was performed to determine

the effects of testing spoof detectors with spoofs fabricated from materials not seen during train-

ing (cross-material evaluation). In the limited cross material evaluation that was performed, the

rate of spoofs correctly classiﬁed as spoofs was shown to drop from 96.57% to 94.20% [121].

While this slight drop in accuracy seems promising, without knowing the performance at ﬁeld

conditions, namely False Detect Rate (FDR)3 of 0.1% on a larger collection of unknown materials,

the reported levels of total accuracy should be accepted with caution. Chugh et al. [26] pushed

3The required operating point for the ODIN program supporting this research is FDR = 0.2%

62

state-of-the-art ﬁngerprint spoof detection performance on the LivDet 2015 dataset from 95.51%

average accuracy to 98.61% average accuracy using a CNN trained on patches around minutiae

points, but they also demonstrated that performance at strict operating points dropped signiﬁcantly

in some experiments. For example, Chugh et al. reported an average accuracy on the LivDet 2011

dataset of 97.41%, however, at a FDR of 1.0%, the TDR was only 90.32%, indicating that current

state-of-the-art spoof detection systems leave room for improvement at desired operating points.

Finally, several other studies have reported up to a three-fold increase in error when testing spoof

detectors on unknown material types [112, 167, 185].

Because of the less than desired performance of spoof detection software to adapt to spoofs

fabricated from unseen materials, studies in [143], [142], and [38] developed open-set recognition

classiﬁers to better detect spoofs fabricated with novel material types. However, while these clas-

siﬁers are able to generalize to spoofs made with new materials better than closed-set recognition

algorithms, their overall accuracy (approx. 85% - 90%) still does not meet the desired perfor-

mance for ﬁeld deployments. Other attempts to bridge the gap between seen and unseen material

spoof detection performance include synthetic spoof generation, and adversarial representation

learning [28, 71].

Given the limitations of state-of-the-art ﬁngerprint spoof detection (both in hardware and soft-

ware), it is evident that much work remains to be done in developing robust and generalizable

spoof detection solutions. We posit that one of the biggest limitations facing the most successful

spoof detection solutions to date (such as use of textural features [185] and CNNs [26,117,131]), is

the processed COTS ﬁngerprint reader images used to train spoof detectors. In particular, because

COTS ﬁngerprint readers output ﬁngerprint images which have undergone a number of image pro-

cessing operations (in an effort to achieve high matching performance), they are not optimal for

ﬁngerprint spoof detection, since valuable information such as color and textural aberrations is lost

during the image processing operations. By removing color and minute textural details from the

raw ﬁngerprint images, spoof ﬁngerprint impressions and live ﬁngerprint impressions (acquired on

63

COTS optical readers) appear very similar (Fig. 3.2 (c)), even when the physical live/spoof ﬁngers

used to collect the respective ﬁngerprint impressions appear very different (Fig. 3.3).

This limitation inherent to many existing spoof detection solutions motivated us to develop a

custom, optical ﬁngerprint reader, called RaspiReader, with the capability to output 2 raw images

(from 2 different cameras) for spoof detection. By mounting two cameras at appropriate angles

to a glass prism (Fig. 4.1), one camera is able to capture high contrast FTIR ﬁngerprint images

(useful for both ﬁngerprint spoof detection and ﬁngerprint matching) (Fig. 3.2 (ii)), while the other

camera captures direct images of the ﬁnger skin in contact with the platen (useful for ﬁngerprint

spoof detection) (Fig. 3.2 (i)). Both images of the RaspiReader visually differentiate between

live ﬁngers and spoof ﬁngers much more than the processed ﬁngerprint images output by COTS

ﬁngerprint readers (Fig. 3.2 (c)).

RaspiReader’s two camera approach is similar to that which was prescribed by Rowe et al.

in [130, 146] where both an FTIR image and a direct view image were acquired using different

wavelength LEDs, however, the commercial products developed around the ideas in [130,146] act

as a proprietary black box outputting only a single processed composite image of a collection of

raw image frames captured under various wavelengths. As such, ﬁngerprint researchers cannot

implement new spoof detection schemes on the individual raw frames captured by the reader.

Furthermore, unlike the patented ideas in [130], RaspiReader is built with ubiquitous components

and open source software packages, enabling ﬁngerprint researchers to very easily prototype their

own RaspiReader, further customize it with new spoof detection hardware, and gain direct access

to the raw images captured by the reader. In short, the low cost ($175 USD) and easy to implement

(1 hour build time) RaspiReader is a truly unique concept which we posit will push the boundaries

of state-of-the-art ﬁngerprint spoof detection, by facilitating spoof detection schemes which use

both hardware and software.

Experiments demonstrate that by utilizing the two cameras of RaspiReader, we are able to

signiﬁcantly boost the performance of state-of-the-art spoof detectors previously trained on COTS

grayscale images (both on known-material and cross-material testing scenarios).

In particular,

64

Figure 3.4 Schematic illustrating RaspiReader functionality. Incoming white light from three LEDs
enters the prism. Camera 2 receives light rays reﬂected from the ﬁngerprint ridges only (light rays
are not reﬂected back from the ﬁngerprint valleys due to total internal reﬂection (TIR)). This image
from Camera 2, with high contrast between ridges and valleys can be used for both spoof detection
and ﬁngerprint matching. Camera 1 receives light rays reﬂected from both the ridges and valleys.
This image from Camera 1 provides complementary information for spoof detection.

because both image outputs of the RaspiReader are raw and contain useful color information, we

can extract discriminative and complementary information from each of the image outputs. By

fusing this complementary information (at a feature level or score level) the performance of spoof

detectors is signiﬁcantly higher than when features are extracted from COTS grayscale images.

Finally, by calibrating and processing the FTIR image output of the RaspiReader (post spoof

detection), we demonstrate that RaspiReader is not only interoperable with existing COTS optical

readers but is also capable of achieving state-of-the-art ﬁngerprint matching accuracy. Note that

interoperability with existing COTS readers is absolutely vital in any new hardware based spoof

detection solution as it makes the spoof resistant device compatible (in terms of matching) with

65

legacy ﬁngerprint databases4. Furthermore, by making the RaspiReader compatible with existing

COTS readers, we further extend the utility of RaspiReader beyond spoof detection. In particular,

RaspiReader is not only useful for providing direct access to multiple raw images for spoof detec-

tion; it also provides researchers in ﬁngerprint matching the easy ability to ﬁne tune (resolution and

processing) the images being output by the ﬁngerprint reader. In any imaging system, the recogni-

tion performance depends on the quality of the image output by the sensor. This is particularly true

of ﬁngerprint recognition systems. As shown in the NIST FpVTE 2012 [178] results, the single

most important factor responsible for degrading ﬁngerprint recognition performance is the ﬁnger-

print image quality. However, most ﬁngerprint researchers have no control over the quality of the

ﬁngerprint images being used to develop ﬁngerprint recognition algorithms since they must rely on

blackbox COTS ﬁngerprint readers. RaspiReader changes this by providing ﬁngerprint matching

algorithm designers an easy method for prototyping their own ﬁngerprint reader and optimizing

ﬁngerprint image quality and ﬁngerprint matching algorithms jointly in an effort to further improve

ﬁngerprint recognition performance.

In summary, our work on RaspiReader removes the mystery of designing and understanding

the internals of a ﬁngerprint reader. Using the open-source fabrication process of this ﬁngerprint

reader, any ﬁngerprint algorithm designer can quickly and affordably construct his or her own

reader with the capabilities (spoof detection and matching image quality) necessary to meet their

application requirements.

More concisely, the contributions of this chapter are:

• An open source, easy to assemble, cost effective ﬁngerprint reader, called RaspiReader, ca-
pable of producing ﬁngerprint images useful for spoof detection and that are of high quality

and resolution (1,500 ppi - 3,300 ppi native resolution) for ﬁngerprint matching. The custom

RaspiReader can be easily modiﬁed to facilitate spoof detection and ﬁngerprint matching

studies.

4Interoperability with existing COTS readers is a strict requirement of the IARPA ODIN program supporting this

research [76].

66

Table 3.1 Primary Components Used to Construct RaspiReader. Total Cost is $175.20

Component Image

Name and Description

Quantity Cost (USD)1

Raspberry Pi 3B: A single board
computer (SBC) with 1.2 GHz 64-bit

quad-core CPU, 1 GB RAM,

MicroSDHC storage, and Broadcom

VideoCore IV Graphic card

Raspberry Pi Camera Module V1:

A 5.0 megapixel, 30 frames per
second, ﬁxed focal length camera
Multi-Camera Adapter: Splits
Raspberry Pi camera slot into two
slots, enabling connection of two

cameras

LEDs: white light, 5 mm, 1 watt

Resistors: 1 kΩ

Right Angle Prism:2 25 mm leg, 35.4

mm hypotenuse

1

2

1

3

3

1

$38.27

$13.49

$49.99

$0.10

$5.16

$54.50

1 All items except the glass prism were purchased for the listed prices on Ama-
zon.com
2 The glass prism was purchased from ThorLabs [169].

• A customized ﬁngerprint reader with two cameras for image acquisition rather than a single
camera. Use of two cameras enables robust ﬁngerprint spoof detection, since we can extract

features from two complementary, information rich images instead of processed grayscale

images output by traditional COTS optical ﬁngerprint readers.

• A signiﬁcant boost in spoof detection performance (both known-material and seven cross-
material testing scenarios) using current state-of-the-art software based spoof detection

methods in conjunction with RaspiReader images as opposed to COTS optical grayscale

images. Spoofs of seven materials were used in both known-material and cross-material

testing scenarios.

67

• Demonstrated matching interoperability of RaspiReader with a COTS optical ﬁngerprint
reader. Since RaspiReader is shown to be interoperable with COTS readers, it could imme-

diately be deployed in the real world since interoperability makes the device compatible with

legacy ﬁngerprint databases.

3.2 RaspiReader Construction and Calibration

In this section, the construction of the RaspiReader using ubiquitous, off-the-shelf components

(Table 1) is explained. In particular, the main steps involved in constructing RaspiReader consist

of (i) properly mounting cameras (angle and position) with respect to a glass prism, (ii) fabricating

a plastic case to house the hardware components, (iii) assembling the cameras and hardware within

the plastic case, and (iv) writing software to capture ﬁngerprint images with the assembled hard-

ware. Each of these steps is described in more detail in the following subsections. Finally, we pro-

vide the steps for calibrating and processing the raw FTIR ﬁngerprint images of the RaspiReader

for ﬁngerprint matching.

3.2.1 Camera Placement

The most important step in constructing RaspiReader is the placement (angle and position) of the

two cameras capturing ﬁngerprint images. In particular, to collect an FTIR image of a ﬁngerprint,

a camera needs to be mounted at an angle greater than the critical angle, and to collect a direct view

image, a camera needs to be mounted an an angle less than the critical angle (both with respect to

the platen). Here, the critical angle is deﬁned as the angle at which total internal reﬂection occurs

when light passes from a medium with an index of refraction n1 to another medium with index of

refraction n2 (Eq. 3.2.1):

θc = arcsin(

n2
n1

)

68

(3.2.1)

In the case of ﬁngerprint sensing, the ﬁrst medium is glass which has an index of refraction

n1 = 1.5, and the second medium is air which has an index of refraction of n2 = 1.0 leading to
a critical angle (θc) of 41.8◦. Therefore, as shown in (Fig. 4.1), we mount the direct view camera
(camera1) at an angle of θ1 = 10◦ and we mount the FTIR camera (camera2) at an angle of
θ2 = 45◦.

With respect to the position of each camera lens to the glass prism, there is a tradeoff between

resolution and ﬁngerprint area to consider. As the camera is moved closer to the prism, the ﬁn-

gerprint image resolution (pixels per inch) is increased. However, if the cameras are too close to

the platen, only part of the ﬁngerprint image is within the ﬁeld of view (FOV). In constructing

RaspiReader, we wanted to maximize the ﬁngerprint image resolution, while still capturing the

entire ﬁngerprint image within the FOV. We experimentally determined that at a distance of 23 mm

from the prism, the cameras would capture the entire ﬁngerprint area. At closer distances, part of

the ﬁngerprint image would start to be outside the FOV. As a ﬁnal step in camera placement, the

focal length of the Raspicams (cameras used in RaspiReader) must be increased so that the camera

will focus on the nearby glass prism (the default focus-length of the Raspicams is 1 meter; much

greater than the 23 mm distant prism). By default, the Raspicams have a ﬁxed-focal length of 3.6
mm. However, by rotating the Raspicam lens 652.5◦ counterclockwise (for the FTIR imaging cam-
era) and 405◦ counterclockwise (for the direct imaging camera), the focal length can be slightly
increased to bring the nearby ﬁngerprint images into focus.

3.2.2 Case Fabrication

After determining the angle and position of both cameras, an outer casing (Fig. 3.5) accommodat-

ing these positions is electronically modeled using Meshlab [29] and subsequently 3D printed on

a high resolution 3D printer (Stratasys Objet350 Connex)5. To make the fabrication process easily

reproducible, the camera mounts and light source mounts are modeled in place on the front part of

the ﬁngerprint reader case (Fig. 3.5). As such, one only needs to 3D print the open-source STL

5We are currently investigating alternative case manufacturing methods such as CNC milling.

69

ﬁles and clip the LEDs and Raspicams to their respective mounts (Fig. 3.5) in order to quickly

build their own RaspiReader replica.

Figure 3.5 Electronic CAD model of the RaspiReader case. The camera and LED mounts are
positioned at the necessary angles and distance to the glass prism, making the reproduction of
RaspiReader as simple as 3D printing the open-sourced STL ﬁles.

3.2.3 Image Acquisition Hardware and Software

The backbone of the RaspiReader is the popular Raspberry Pi 3B single board computer, which

enables easy interfacing with GPIO pins (for controlling LEDs) and image acquisition (with its

standard camera and camera connection port). Because the Raspberry Pi only has a single camera

connection port, a camera port multiplexer is used to enable the use of multiple cameras on a single

Pi [3]. Using the Raspberry Pi GPIO pins, the code available in [3], and the camera multiplexer,

one can easily extend the Raspberry Pi to use multiple cameras.

After assembling the camera port multiplexer to the Pi (with two Raspicams), wiring 3 LEDs

to the Raspberry Pi GPIO pins, and attaching the Raspicams and LEDs to the 3D printed casing

mounts (Fig. 3.5), open source python libraries [3] can be used to illuminate the glass prism

and subsequently acquire two images (Fig. 3.2 (a)) from the ﬁngerprint reader (one raw FTIR

ﬁngerprint image and another raw direct ﬁngerprint image).

70

Figure 3.6 Processing a RaspiReader raw FTIR ﬁngerprint image into a 500 ppi ﬁngerprint image
compatible for matching with existing COTS ﬁngerprint readers. (a) The RGB FTIR image is ﬁrst
converted to grayscale. (b) Histogram equalization is performed to enhance the contrast between
the ﬁngerprint ridges and valleys. (c) The ﬁngerprint is negated so that the ridges appear dark,
and the valleys appear white. (d), (f) Calibration (estimated using the checkerboard calibration
pattern in (e)) is applied to frontalize the ﬁngerprint image to the image plane and down sample
(by averaging neighborhood pixels) to 500 ppi in both the x and y axis.

3.2.4 Fingerprint Image Processing

In order for the RaspiReader to be used for spoof detection, it must also demonstrate the ability to

output high quality ﬁngerprint images suitable for ﬁngerprint matching. As previously mentioned,

the RaspiReader performs spoof detection on non-processed, raw ﬁngerprint images. While these

raw images are shown to provide discriminatory information for spoof detection, they need to be

made compatible with processed images output by other COTS ﬁngerprint readers. Therefore,

after spoof detection, the RaspiReader performs image processing operations on the raw high con-

trast, FTIR image frames in order to output high ﬁdelity images compatible with COTS optical

ﬁngerprint readers.

Let a raw (unprocessed) FTIR ﬁngerprint image from the RaspiReader be denoted as F T IRraw.

This raw image F T IRraw is ﬁrst converted from the RGB color space to grayscale (F T IRgray)

(Fig. 3.6 (a)). Then, in order to further contrast the ridges from the valleys of the ﬁngerprint,

histogram equalization is performed on F T IRgray (Fig. 3.6 (b)). Finally, F T IRgray is negated so

71

Figure 3.7 Acquiring Image Transformation Parameters. A 2D printed checkerboard pattern (a) is
imaged by the RaspiReader (b). Corresponding points between the frontalized checkerboard pat-
tern (a) and the distorted checkerboard pattern (b) are deﬁned so that perspective transformation
parameters can be estimated to map (b) into (c). These transformation parameters are subsequently
used to frontalize ﬁngerprint images acquired by RaspiReader for the purpose of ﬁngerprint match-
ing. The checkerboard imaged in (b) is also used to acquire the native resolution of RaspiReader
in order to scale matching images to 500 ppi in both the x and y axis as shown in (c).

that the ridges of the ﬁngerprint image are dark, and the background of the image is white (as are

ﬁngerprint images acquired from COTS readers) (Fig. 3.6 (c)).

Following the aforementioned image processing techniques, the RaspiReader FTIR ﬁngerprint

images are further processed by performing a perspective transformation (to frontalize the ﬁnger-

print to the image plane) and scaling to 500 ppi (Figs. 3.6 (d), (f)). Note, we also experimented

with non-linear distortion corrections (camera barrel distortion), but found no improvement (over a

simple linear distortion correction) in RaspiReader matching performance and little improvement

in error between landmarks on the ground truth calibration pattern and the non-linear distortion

corrected images. This makes sense, since the distortion of the raw FTIR images can be seen to be

predominantly tangential (Fig. 3.7 (b)).

A perspective transformation is performed using Equation 3.2.2,

72

 =

1
λ



x(cid:48)
y(cid:48)

1



a b

c

d e f

g h 1



(3.2.2)





x

y

1

where x and y are the source coordinates, x(cid:48) and y(cid:48) are the transformed coordinates,
(a, b, c, d, e, f, g, h) is the set of transformation parameters, and λ = gx + hy + 1 is a scale param-

eter. In this work, we image a 2D printed checkerboard pattern6 to deﬁne source and destination

coordinate pairs such that the transformation parameters could be estimated (Fig. 3.7). Once

the perspective transformation has been completed, the RaspiReader image is downsampled (by

averaging neighborhood pixels) to 500 ppi (Fig. 3.6 (f)). Note that the native resolution of the

RaspiReader images was acquired using a 2D printed checkerboard calibration pattern (Fig. 3.7

(b)) and ranges from approx. 1594 ppi to 2480 ppi in the x-axis (Fig. 3.8 (a)) and 2463 ppi to 3320

ppi in the y-axis (Fig. 3.8 (b)). While the high resolution images captured by the RaspiReader 5

Megapixel cameras far exceed the resolution of COTS ﬁngerprint readers (providing added minute

textural details for distinguishing live ﬁngers from spoof ﬁngers), we observed that the focus of

the native images captured by RaspiReader does deteriorate on the left and right edges (Fig. 3.7

(b)). We are currently investigating methods for properly focusing the lens on the entire FOV, so

that minute textural details are not lost on the edges of the RaspiReader images.

Upon completion of this entire ﬁngerprint reader assembly and image processing procedure, the

RaspiReader is fully functional and ready for use in both spoof detection and subsequent ﬁngerprint

matching.

6A checkerboard can be imaged by RaspiReader by printing a checkerboard pattern on glossy paper and applying

several drops of water to the platen prior to placing the printed checkerboard.

73

(a)

(b)

Figure 3.8 Native resolution (ppi) in (a) x-axis and (b) y-axis over the raw FTIR RaspiReader
image. As is normal, native resolution changes across the image because the right side of the
image is closer to the camera than the left side.

3.3 Live and Spoof Fingerprint Database Construction

To test the utility of the RaspiReader for spoof detection and its interoperability for ﬁngerprint

matching, a database of live and spoof ﬁngerprint impressions was collected for performing exper-

iments. This database is constructed as follows.

Using 7 different materials (Fig. 3.3 (a)), 66 spoofs were fabricated7. Then, for each of these

spoofs, 10 impressions were captured at varying orientations and pressure on both the RaspiReader

(Rpi) and a COTS 500 ppi, PIV-certiﬁed, optical FTIR ﬁngerprint reader (COT SA). Note, we

also experimented with an Appendix-F certiﬁed slap scanner but found little difference in spoof

detection performance between the PIV-certiﬁed device and the Appendix-F certiﬁed device. The

summary of this data collection is enumerated in Table 3.1.

7Our spoofs were shipped to us by Precise Biometrics [139], a company specializing in evaluating spoof detection
capability and that also has close ties to the LivDet dataset authors. As such, our spoofs are of high quality and are
similar to the spoofs used in the LivDet competition.

74

Table 3.2 Summary of Spoof Fingerprints Collected

Material1

Ecoﬂex

Wood Glue

Monster Liquid

Latex

Liquid Latex
Body Paint

Gelatin

Silver Coated

Ecoﬂex

Crayola Model

Magic
Total

Spoof Count2 RPi Direct
Images

RPi FTIR
Images

COT SA FTIR

Images

10
10

10

10

10

10

6

66

100
100

100

100

100

100

60

660

100
100

100

100

100

100

60

660

100
100

100

100

100

100

60

660

1 The spoof materials used to fabricate these spoofs were in accordance with
the approved materials by the IARPA ODIN project [76].
2 The spoofs are all of unique ﬁngerprint patterns.

To collect a sufﬁcient variety of live ﬁnger data, we enlisted 15 human subjects with different

skin colors (Fig. 3.3 (b)). Each of these subjects gave 5 ﬁnger impressions (at different orientations

and pressures) from all 10 of their ﬁngers on both the RaspiReader and COT SA. A summary of

this data collection is enumerated in Table 3.2.

Table 3.3 Summary of Live Finger Data Collected

Number of
Subjects

Number of

RPi Direct

Fingers

Images

RPi FTIR
Images

COT SA FTIR

Images

15

150

750

750

750

In addition to the images of live ﬁnger impressions and spoof ﬁnger impressions we collected

for conducting spoof detection experiments, we also veriﬁed that for spoofs with optical properties

too far from that of live ﬁnger skin (Fig. 3.9), images would not be captured by the RaspiReader.

75

These “failure to capture” spoofs are therefore ﬁltered out as attacks before any software based

spoof detection methods need to be performed.

Figure 3.9 Failure to Capture. Several spoofs are unable to be imaged by the RaspiReader due to
their dissimilarity in color. In particular, because spoofs in (a) and (b) are black, all light rays will
be absorbed preventing light rays from reﬂecting back to the FTIR imaging sensor. In (c), the dark
blue color again prevents enough light from reﬂecting back to the camera. (a) and (b) are both
ecoﬂex spoofs coated with two different conductive coatings. (c) is a blue crayola model magic
spoof attack.

3.4 Spoof Detection Experiments and Results

Given the database of live and spoof ﬁngerprint images collected on both COT SA, and the pro-

totype RaspiReader, a number of spoof detection experiments are conducted to demonstrate the

superiority of the raw images from the RaspiReader for training spoof detectors in comparison to

the grayscale images output by COTS optical readers. In particular, we (i) take several successful

spoof detection techniques from the literature, (ii) train and test the spoof detectors on COT SA

images, (iii) train and test the spoof detectors on RaspiReader images, and (iv) compare the results

to show the signiﬁcant boost in performance when RaspiReader images are used to train spoof

detectors rather than COT SA images. In addition, experiments are conducted to demonstrate that

ﬁngerprint images from the RaspiReader are compatible for matching with ﬁngerprint images ac-

quired from COT SA.

76

3.4.1 Spoof Detection Methods

To thoroughly demonstrate the value RaspiReader images provide in training spoof detectors, we

select two different spoof detection methods, namely, (i) textural features (LBP [132]) in conjunc-

tion with a linear Support Vector Machine (SVM) and (ii) a Convolutional Neural Network (CNN).

Textural features were chosen because of their popularity and their demonstrated superior spoof

detection performance in comparison to other “hand-crafted features” such as anatomical or phys-

iological features in the literature [185]. CNNs were chosen as a second spoof detection method

in our experiments given that they are currently state-of-the-art on the publicly available LivDet

datasets. [26, 117, 131, 132]. The details of the experiments performed with both of these spoof

detection methods are provided in the following subsections.

3.4.1.1 LBP Features From COT SA Images

We begin our experiments using grayscale processed ﬁngerprint images acquired from COT SA

(Fig. 3.2 (c)). From these images, we extract the very prevalent grayscale and rotation invariant

local binary patterns (LBP) [132]. LBP features are extracted by constructing a histogram of bit

string values determined by thresholding pixels in the local neighborhoods around each pixel in the

image. Since image texture can be observed at different spatial resolutions, parameters R and P

are speciﬁed in LBP construction to indicate the length (in pixels) of the neighborhood radius used

for selecting pixels and also the number of neighbors to consider in a local neighborhood. Previous

studies have shown that more than 90% of fundamental textures in an image can belong to a small

subset of binary patterns called “uniform” textures (local binary patterns containing two or fewer

0/1 bit transitions) [132]. Therefore, in line with previous studies using local binary patterns for

ﬁngerprint spoof detection, we also employ the use of uniform local binary patterns.

More formally, let LBP (P, R) be the uniform local binary pattern histogram constructed by

binning the local binary patterns for each pixel in an image according to the well known LBP

operation [132] with parameters P and R. In our experiments, we extract LBP (8, 1), LBP (16, 2),

and LBP (24, 3) in order to capture textures at different spatial resolutions. These histograms (each

77

having P + 2 bins) are individually normalized and concatenated into a single feature vector X of

dimension 54.

For classiﬁcation of these features, we employ a binary linear SVM. As known in the art, an

initially “hard margin” SVM can be “softened” by a parameter C to enable better generalization of

the classiﬁer to the testing dataset. In our case, we use ﬁve-fold cross validation to select the value

(cid:20)

(cid:21)

of C (from the list of

10−5 10−4

... 104 105

) such that the best performance is achieved in

different folds. In our experiments, the best classiﬁcation results were achieved with C = 102.

3.4.1.2 CLBP Features From RaspiReader Images

In this experiment, we make use of the information rich images from the RaspiReader (Figs. 3.2 (a,

b)) for spoof detection. As with Experiment 1, we again pursue the use of LBP textural features.

However, since the raw images from the RaspiReader contain color information, rather than using

the traditional grayscale LBP features, we employ the use of color local binary patterns (CLBP).

Previous works have shown the efﬁcacy of CLBP for both face recognition and face spoof detection

[15,25]. However, because ﬁngerprint images from COTS ﬁngerprint readers are grayscale, CLBP

features have, to our knowledge, not been investigated for use in ﬁngerprint spoof detection until

now.

Unlike traditional grayscale LBP patterns, color local binary patterns (CLBP) encode discrimi-

native spatiochromatic textures from across multiple spectral channels [25]. In other words, CLBP

extracts textures across all the different image bands in a given input image. More formally, given

an input image I with K spectral channels, let the set of all spectral channels for I be deﬁned as
S = {S1, ..., SK}. Then, the CLBP feature vector X of dimension 486 can be extracted from I
using Algorithm 1. Note that in Algorithm 1, LBP (Si, Sj, P, R) returns a normalized histogram

of local binary patterns using Si as the image channel that the center (thresholding) pixels are se-

lected from, and Sj as the image channel from which the neighborhood pixels are selected from
in the same computation of LBP as performed in Experiment 1. Also note that in Algorithm 1, (cid:107)
indicates vector concatenation. Finally, in our experiments, we preprocess the RaspiReader input

78

image I prior to CLBP extraction by (i) downsampling (FTIR images from 1450 x 1944 to 108 x

145 and direct view images from 1290 x 1944 to 96 x 145), and (ii) converting to the HSV color

space8.

Algorithm 1 Extraction of Color Local Binary Patterns

X ← [ ]
for i ← 1, K do

for j ← 1, K do

X ← X(cid:107)LBP (Si, Sj, 8, 1)(cid:107)

LBP (Si, Sj, 16, 2)(cid:107)LBP (Si, Sj, 24, 3)

end for

end for

return X

As in Experiment 1, a binary linear SVM with a parameter of C = 102 is trained with these

features and subsequently used for classiﬁcation. We again choose the parameter C using 5-fold

(cid:20)

(cid:21)

cross validation and a selection list of

10−5 10−4

... 104 105

. Since RaspiReader outputs

two color images (one raw FTIR image and one direct view image), we perform multiple experi-

ments using the proposed CLBP features in conjunction with the SVM. In particular, we (i) extract

CLBP features from the RaspiReader raw FTIR images to train/test a SVM, (ii) extract CLBP fea-

tures from RaspiReader direct view images to train/test a SVM, and (iii) fuse CLBP features from

both image outputs to train and test a SVM. We also attempted fusing CLBP features from the

RaspiReader raw images with grayscale LBP features from RaspiReader processed FTIR images,

but found no signiﬁcant performance gains under this last fusion scheme.

8Other color spaces were experimented with, but HSV consistently provided the highest performance. This is likely
because HSV separates the luminance and chrominance components in an image, allowing extraction of features on
more complementary image channels.

79

3.4.1.3 MobileNet

In addition to performing experiments involving “handcrafted” textural features, we also perform

experiments where the features are directly learned and classiﬁed by a Convolutional Neural Net-

work (CNN). In choosing a CNN architecture, we carefully considered both the size and compu-

tational overhead, since in future works, we will optimize the architecture to directly run on the

RaspiReader’s Raspberry Pi Processor. The need for a “low over-head” architecture prompted us

to select MobileNet [75]. MobileNet has been shown to perform very closely (within 1 % ac-

curacy) to popular CNN models (VGG and Inception v3) on the ImageNet and Stanford Dogs

datasets while being 32 times smaller than VGG, 7 times smaller than Inception v3, 27 times less

computationally expensive than VGG, and 8 times less computationally expensive than Inception

v3.

In our experiments, we employ the Tensorﬂow Slim implementation of MobileNet9. MobileNet

is comprised of 28 convolutional layers, and in our case, a ﬁnal 2 class softmax layer for classiﬁca-

tion of live or spoof. In all of our experiments involving MobileNet, the RMSProp optimizer was

used for training the network along with a batch size of 32, and an adaptive (exponential decay)

learning rate. To increase the generalization ability of the networks, we employ various data aug-

mentation methods such as brightness adjustment, random cropping, and horizontal and vertical

reﬂections.

Using the aforementioned MobileNet architecture and hyper-parameters, we train/test the net-

work with (i) COT SA grayscale ﬁngerprint images, (ii) RaspiReader raw FTIR images, (iii)

RaspiReader direct view images, and (iv) RaspiReader processed FTIR images. Additionally, we

perform experiments in which we fuse the score outputs of MobileNet models trained on the differ-

ent image outputs from RaspiReader to take advantage of the complementary information within

the different RaspiReader image outputs. When training and testing MobileNet with COT SA im-

ages or RaspiReader processed FTIR images, the three input channels of the network are each fed

with the same down sampled (357 x 392 to 224 x 224) grayscale COT SA image or (290 x 267

9https://github.com/tensorﬂow/models/tree/master/research/slim

80

to 224 x 224) RaspiReader processed FTIR image. When training the network with RaspiReader

raw images, we again down sample the images (1450 x 1944 to 224 x 224 for raw FTIR and 1290

x 1944 to 224 x 224 for direct image), however, in this case, each of the three color channels

are fed as input to the three input channels of the network. More speciﬁcally, we ﬁrst convert the

RaspiReader image to HSV (given our earlier ﬁndings of superior performance in this color space),

and then feed each channel H, S, and V into the network’s input channels.

3.4.2 Spoof Detection Results

Using the spoof detection schemas previously described, we train and test classiﬁers under two

main scenarios. In the ﬁrst scenario, we train the classiﬁer on a subset of spoof images from every

type in the dataset (Table 3.1). During testing, spoof images from the same spoof types seen during

training will be passed to the spoof detector for classiﬁcation. We hereafter refer to this training

and testing scenario as a “known-material” scenario. In the second scenario, we train the classiﬁer

with images from all of the spoof types in the dataset except one (i.e. the spoof impressions from

one type of spoof are withheld). Then, during testing the impressions of the withheld spoof type are

used for testing. In the literature, this type of spoof detection evaluation is referred to as a “cross-

material” scenario. In the following experimental results, we demonstrate that the RaspiReader

images signiﬁcantly boost the spoof detection performance in both the known-material evaluations

and the cross-material evaluations.

3.4.2.1 Known-Material Scenarios

The ﬁrst known-material results are reported in accordance with spoof detection methods 1 and

2. That is, we extract textural features from both COT SA images and RaspiReader images re-

spectively, train and test linear SVMs, and ﬁnally, compare the results (Table 3.3). In all of our

known-material scenario experiments, we report the average spoof detection performance and stan-

dard deviation over 5-folds. That is, for spoof data, we select 80% of the spoof impressions from

each spoof material for training (each fold) and use the remaining 20% for testing. For live ﬁnger

81

data, we select the ﬁnger impressions of 12 subjects each fold (600 total images) for training, and

use the live ﬁnger impressions of the remaining 3 subjects for testing.

Table 3.4 Textural Features and Known Testing Materials

Method

COT SA + LBP

Rpi raw FTIR + CLBP

Rpi Direct + CLBP
Rpi Fusion + CLBP2

TDR @ FDR = 1.0%

µ ± σ1

75.9% ± 30.8
91.5% ± 11.0
98.10% ± 1.9
97.7% ± 3.0

Detection Time (msecs)

236
243
243
486

1 These results are reported over 5-folds.
2 Rpi Fusion + CLBP is a feature level fusion (concatenation) of CLBP fea-
tures extracted from both Rpi raw FTIR images and Rpi Direct Images, re-
spectively.

From the results of Table 3.3, one can observe that both image outputs of the RaspiReader

contain far more discriminative information for spoof detection than the processed grayscale im-

ages output by COT SA. In particular, spoof detection performance is signiﬁcantly higher when

extracting textural (CLBP) features from the RaspiReader images, than when extracting textural

features (LBP) from COT SA images. While in these ﬁrst results, the fusion of features from both

RaspiReader image outputs actually hurts the classiﬁcation performance slightly (compared to

extracting features only from the direct view images), in subsequent experiments, we will demon-

strate that different feature extraction and classiﬁcation techniques can better utilize the multiple

outputs of RaspiReader in a complementary manner to instead boost the classiﬁcation performance.

The second known-material results are reported in accordance with spoof detection scheme 3.

More speciﬁcally, the results are reported (over 5-folds) when MobileNet is trained and tested with

(i) COT SA images, (ii) RaspiReader processed FTIR images, (iii) RaspiReader raw FTIR images,

and (iv) RaspiReader direct images. In addition, we report the results when fusing the score outputs

of multiple MobileNet models trained on the different image outputs of RaspiReader (Table 3.4).

82

Table 3.5 MobileNet and Known Testing Materials

Method

COT SA + MobileNet

Rpi processed FTIR + MobileNet

Rpi raw FTIR + MobileNet

Rpi Direct + MobileNet
Rpi Fusion 2 + MobileNet2
Rpi Fusion 3 + MobileNet3

TDR @ FDR = 1.0%

µ ± σ1

91.9% ± 8.0
94.5% ± 3.7
95.1% ± 5.6
95.3% ± 3.5
98.4% ± 2.3
98.9% ± 1.5

Detection Time

(msecs)

22
22
22
22
45
67

1 These results are reported over 5-folds.
2 Rpi Fusion 2 + MobileNet is a score level fusion (averaging) of a MobileNet
model trained on Rpi raw FTIR images and a MobileNet model trained on Rpi
Direct Images.
3 Rpi Fusion 3 + MobileNet is a score level fusion (averaging) of separate Mo-
bileNet models trained on Rpi raw FTIR images, Rpi Direct Images, and on
Rpi processed FTIR images.

The results of Table 3.4 show that both the raw image outputs of RaspiReader and the processed

image output of RaspiReader contain more discriminative information for spoof detection than the

processed images output by COT SA. The MobileNet models trained on RaspiReader images al-

ways outperform the MobileNet model trained on COT SA grayscale images both in average spoof

detection performance and stability (signiﬁcantly lower s.d.). What is further interesting about the

results of Table 3.4 is that the features extracted by MobileNet from each RaspiReader output are

quite complementary, demonstrated by the fact that spoof detection performance is improved when

fusing the scores of MobileNet models trained on each RaspiReader image output. So, while CLBP

features outperform MobileNet on the RaspiReader direct images, the fused MobileNet classiﬁers

outperform the fused CLBP classiﬁer.

3.4.2.2 Cross-Material Scenarios

The cross-material results use the same spoof detection schemas as enumerated in the known-

material results with a primary difference being the training and testing data splits provided to

83

Table 3.6 Textural Features and Cross-Material Testing1

Testing Material

COT SA + LBP Rpi Fusion + CLBP2

Crayola Model Magic

Ecoﬂex

Silver Coated Ecoﬂex

Gelatin

Liquid Latex Body Paint
Monster Liquid Latex

Wood Glue

91.7%

66.0%

88.0%

62.0%

84.0%

68.0%

100.0%

98.3%

77.0%

100.0%

87.0%

100.0%

98.0%

81.0%

1 TDR @ FDR = 1.0% is reported
2 Rpi Fusion + CLBP is a feature level fusion (concatentation) of
CLBP features extracted from both Rpi raw FTIR images and Rpi
Direct Images, respectively.

the various classiﬁers. In all the cross-material scenarios, spoof impressions of six materials are

partitioned to the classiﬁer for training, and the spoof impressions of one “unseen” material are

kept aside for testing. In this manner the generalization capability of the spoof detector to novel

spoof types is thoroughly assessed. For live ﬁnger data, we randomly select the ﬁnger impressions

of two subjects (100 total images) for testing, and use the live ﬁnger impressions of the remaining

thirteen subjects for training. Since there are seven different spoof materials in our training set (Ta-

ble 3.1), we conduct seven different cross-material experiments for each spoof detection schema

(where one of the seven spoof types is left aside for testing). The cross material results when using

textural features in conjunction with SVMs is reported in Table 3.5. The cross-material results

when using MobileNet extracted features is reported in Table 3.6. Finally, we report the spoof

detection accuracy of a state-of-the-art, commercial ﬁngerprint reader (with embedded spoof de-

tection capabilities) (Lumidigm V-Series [73]) to further provide a fair comparison of our proposed

RaspiReader to current state-of-the-art hardware based spoof detection systems (Table ??). Note,

we only report the best textural fusion and CNN fusion methods in the cross-material results. The

84

Table 3.7 MobileNet and Cross-Material Testing1

Testing Material COT SA + MobileNet

Rpi Fusion 2 +
MobileNet2

Rpi Fusion 3 +
MobileNet3

Crayola Model

Magic
Ecoﬂex

Silver Coated

Ecoﬂex
Gelatin

Liquid Latex
Body Paint

Monster Liquid

Latex

50.0%

100.0%

77.0%

88.0%

97.0%

86.0%

100.0%

100.0%

8.0%

56.0%

100.0%

100.0%

100.0%

100.0%

100.0%

100.0%

100.0%

100.0%

Wood Glue

96.0%

96.0%

94.0%
1 TDR @ FDR = 1.0% is reported.
2 Rpi Fusion 2 + MobileNet is a score level fusion (max) of separate MobileNet
models trained on Rpi raw FTIR images and on Rpi Direct Images, respec-
tively.
3 Rpi Fusion 3 + MobileNet is a score level fusion (max) of separate MobileNet
models trained on Rpi raw FTIR images, Rpi Direct Images, and on Rpi pro-
cessed FTIR images.

Table 3.8 Lumidigm Spoof Detection Accuracy

Crayola Model

Magic
100%

Ecoﬂex

100%

Silver Coated

Ecoﬂex
100%

Gelatin

Liquid Latex
Body Paint

90%

100%

Monster Liquid

Latex
50%

Wood
Glue
45%

1 Lumidigm classiﬁcation accuracy is reported to enable further comparison of RaspiReader to
2 For each material, the same spoofs used
state-of-the-art spoof resistant ﬁngerprint readers.
to construct Table 2 were imaged on Lumidigm (20 impressions per material).

other non-fusion based methods were experimented with, but did not provide as high of perfor-

mance in the cross-material scenarios.

The key ﬁndings of the cross-material experiments as revealed in Tables 3.5, 3.6, and 3.7 are

as follows. First, in both textural based spoof detection methods and CNN based spoof detec-

85

tion methods, the raw images output by RaspiReader almost always provide more discriminative

information than COTS grayscale ﬁngerprint images. This enables much higher spoof detection

performance on spoofs fabricated from materials not seen by the classiﬁer during training, a major

ﬂaw in many existing spoof detection methods relying on only COTS grayscale images. We also

note that the RaspiReader cross-material performance is signiﬁcantly higher than Lumidigm (Ta-

bles 3.6 and 3.7) on several spoof materials, further demonstrating RaspiReader’s effectiveness in

comparison to current state-of-the-art, commercial, hardware based spoof detection techniques.

The one case of poor cross-material performance (when using RaspiReader images) came when

the testing material withheld was ecoﬂex (Table 3.6). This can be explained by ecoﬂex being a

very transparent spoof, enabling much of the live ﬁnger color behind the spoof to seep through.

As such, when the MobileNet models were trained on the other non-transparent spoofs and tested

on the transparent ecoﬂex, the performance dropped considerably. However, we also noticed that

the best cross-material performance (when using COT SA images) came when the testing material

withheld was ecoﬂex. The most plausible explanation for this is that the MobileNet model trained

on the COT SA images must focus on textural features rather than color. As such, the transparent

property of ecoﬂex did not affect the classiﬁer trained on the grayscale images. This prompted us

to train a third model on the RaspiReader processed FTIR images (i.e. the raw FTIR images were

converted to grayscale and contrast enhanced). We then fused the score of this third model with

the two MobileNet models trained on the RaspiReader raw FTIR and direct images respectively.

The ﬁnal product was a three CNN model system which performed much better on the ecoﬂex

testing scenario (48% improvement). Note, the standalone performance of the MobileNet model

trained on RaspiReader processed FTIR images was 86.0%, lending evidence to our hypothesis

that RaspiReader performed worse on ecoﬂex due to the transparent nature of the material. While

the ecoﬂex testing scenario is still low under fusion (56.0%), in a real world setting, this limitation

is easily solved by including one transparent spoof in the training set (evidenced by the fact that

in the known-material experiments, ecoﬂex could be differentiated from live ﬁngers with high

accuracy).

86

We also note that in Table 3.5, COT SA outperformed RaspiReader when the testing material

withheld was wood glue. Again, in this speciﬁc testing scenario, the grayscale features were

better able to generalize to the unknown spoof material. This is evidenced by the fact that when

using grayscale LBP features extracted from the RaspiReader matching images, the TDR was

computed to be 93.0%, a signiﬁcant improvement from the CLBP performance of 81.0%. The

remaining difference between the RaspiReader grayscale LBP performance on wood glue and

COT SA performance on wood glue can likely be attributed to subtle differences in the image

processing techniques used by the two readers in converting a raw FTIR image to a processed

FTIR image.

3.5

Interoperability of RaspiReader

In addition to demonstrating the usefulness of the RaspiReader images for ﬁngerprint spoof detec-

tion, we also demonstrate that by processing the RaspiReader FTIR images, we can output images

which are compatible for matching with images from COTS ﬁngerprint readers. Previously, we

discussed how to process and transform a RaspiReader raw FTIR image into an image suitable for

matching. In this experiment, we evaluate the matching performance (of 11,175 imposter pairs and

6,750 genuine pairs) when using (i) the RaspiReader processed images as both the enrollment and

probe images, (ii) the COT SA images as both the enrollment and probe images, (iii) the COT SA

images as the enrollment images and the RaspiReader processed images as the probe images, and

(iv) the RaspiReader images as the enrollment images and the COT SA images as the probe images.

The results for these matching experiments are listed in Table 3.8.

From these results, we make two observations. First, the best performance is achieved for na-

tive comparisons, where the enrolled and search (probe) images are produced by the same capture

device. RaspiReader’s native performance is slightly better than that of COT SA. This indicates

that the RaspiReader is capable of outputting images which are compatible with state of the art

ﬁngerprint matchers. Second, we note that the performance does drop slightly when conducting

87

Table 3.9 Fingerprint Matching Results

Enrollment Reader Probe Reader TAR @ FAR = 0.1%†

COT SA

RaspiReader

COT SA

RaspiReader

COT SA

RaspiReader
RaspiReader

COT SA

98.62%
99.21%
95.56%
95.10%

† We use the Innovatrics ﬁngerprint SDK which is shown to
have high accuracy in the NIST FpVTE evaluation [178].

the interoperability experiment (COT SA is used for enrollment images and RaspiReader is used

for probe images). However, the matching performance is still quite high considering the stringent

operating point (FAR = 0.1%). Furthermore, studies have shown that when different ﬁngerprint

readers are used for enrollment and subsequent veriﬁcation or identiﬁcation, the matching per-

formance indeed drops [43, 94, 144]. Finally, we are currently investigating other approaches for

processing and downsampling RaspiReader images to reduce some of the drop in cross-reader

performance.

3.6 Computational Resources

All image preprocessing, LBP and CLBP feature extractions, and SVM classiﬁcations were per-

formed with a single CPU core on a Macbook Pro running a 2.9 GHz Intel Core i5 processor.

MobileNet training and classiﬁcation was performed on a single Nvidia GTX Titan GPU. The to-

tal time from image capture to spoof detection with our best MobileNet model (RpiFusion3) is

approximately 3.067 seconds (1.5 seconds for image capture, 1.5 seconds to transmit data to GPU,

and 67 milliseconds for classiﬁcation).

88

3.7 Summary

We have open sourced10 the design and assembly of a custom ﬁngerprint reader, called

RaspiReader, with Raspberry Pi and other ubiquitous components. This ﬁngerprint reader is both

low cost (US $175) and easy to assemble11, enabling other researchers to easily and seamlessly

develop their own novel ﬁngerprint spoof detection solutions which use both hardware and soft-

ware. By customizing RaspiReader with two cameras for ﬁngerprint image acquisition rather

than the customary one, we were able to extract discriminative information from both raw images

which, when fused together, enabled us to achieve higher spoof detection performance (in both

known-material and cross-material testing scenarios) compared to when features were extracted

from COTS grayscale images. Finally, by processing the raw FTIR images of the RaspiReader,

we were able to output ﬁngerprint images compatible for matching with COTS optical ﬁngerprint

readers demonstrating the interoperability of RaspiReader.

Future directions could be to integrate specialized hardware into RaspiReader such as IR cam-

eras for vein detection, or microscopes for capturing extremely high resolution images of the ﬁn-

gerprint. Because the RaspiReader uses ubiquitous components running open source software,

RaspiReader enables integration of these additional hardware components. In addition to the in-

tegration of specialized hardware, one could also pursue use of the raw, information rich images

from the RaspiReader to pursue one-class classiﬁcation schemes for ﬁngerprint spoof detection

such as that proposed in [45].

3.8 Acknowledgment

This research was supported by grant no. 70NANB17H027 from the NIST Measurement Science

program and by the Ofﬁce of the Director of National Intelligence (ODNI), Intelligence Advanced

Research Projects Activity (IARPA), via IARPA R&D Contract No. 2017 - 17020200004. The

10https://github.com/engelsjo/RaspiReader
11https://bit.do/RaspiReader

89

views and conclusions contained herein are those of the authors and should not be interpreted as

necessarily representing the ofﬁcial policies, either expressed or implied, of ODNI, IARPA, or

the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for

governmental purposes notwithstanding any copyright annotation therein.

We would like to thank Chris Perry for all of his help and support as a system integrator

throughout this project.

90

Chapter 4

Learning a Fixed Length Fingerprint

Representation

In the previous chapters, our focus was primarily on improving the performance and security of

the ﬁngerprint reader sub-module of ﬁngerprint recognition systems via Universal Target evalu-

ations and the spoof resistant RaspiReader, respectively. In this chapter, we turn our focus to-

wards improving the performance (accuracy and speed) and security of the ﬁngerprint feature

extraction and matching sub-modules of ﬁngerprint recognition systems. In particular, we present

DeepPrint, a deep network, which learns to extract ﬁxed-length ﬁngerprint representations of only

200 bytes [48]. The DeepPrint representation enables accuracy levels comparable to state-of-the-

art AFIS based upon minutiae representations and can be matched at orders of magnitude faster

speeds. Furthermore, the DeepPrint representation can be more easily secured and matched in the

encrypted domain via a fully homomorphic encryption scheme.

To arrive at a discriminative ﬁxed-length representation, DeepPrint incorporates ﬁngerprint

domain knowledge, including alignment and minutiae detection, into a deep network architecture.

We benchmark DeepPrint against two top performing COTS SDKs (Veriﬁnger and Innovatrics)

from the NIST and FVC evaluations. Coupled with a re-ranking scheme, the DeepPrint rank-1

search accuracy on the NIST SD4 dataset against a gallery of 1.1 million ﬁngerprints is comparable

91

to the top COTS matcher, but it is signiﬁcantly faster (DeepPrint: 98.80% in 0.3 seconds vs.

COTS A: 98.85% in 27 seconds). To the best of our knowledge, the DeepPrint representation is the

most compact and discriminative ﬁxed-length ﬁngerprint representation reported in the academic

literature.

4.1

Introduction

To overcome the limitations of the variable length minutiae representation (Table 4.1), we present

a reformulation of the ﬁngerprint recognition problem. In particular, rather than extracting varying

length minutiae-sets for matching (i.e. handcrafted features), we design a deep network embedded

with ﬁngerprint domain knowledge, called DeepPrint, to learn a ﬁxed-length representation of

200 bytes which discriminates between ﬁngerprint images from different ﬁngers (Fig. 4.1).

Table 4.1 Comparison of variable length minutiae representation with ﬁxed-length DeepPrint rep-
resentation

(12, 196)
(12, 225)
N.A.2

(Min, Max) # of Minutiae1 (Min, Max) Template Size (kB)

Matcher
COTS A
COTS B
Proposed
1 Statistics from NIST SD4 and FVC 2004 DB1.
2 Template is not explicitly comprised of minutiae.
† Template size is ﬁxed at 200 bytes, irrespective of the number of minutiae
(192 bytes for the features and 8 bytes for 2 decompression scalars).

(1.5, 23.7)
(0.6, 5.3)

0.2†

To arrive at a compact and discriminative representation of only 200 bytes, the DeepPrint ar-

chitecture is embedded with ﬁngerprint domain knowledge via an automatic alignment module

and a multi-task learning objective which requires minutiae-detection (in the form of a minutiae-

map) as a side task to representation learning. More speciﬁcally, DeepPrint automatically aligns

an input ﬁngerprint and subsequently extracts both a texture representation and a minutiae-based

representation (both with 96 features). The 192-dimensional concatenation of these two represen-

tations, followed by compression from ﬂoating point features to integer value features comprises

92

Figure 4.1 Flow diagram of DeepPrint: (i) a query ﬁngerprint is aligned via a Localization Net-
work which has been trained end-to-end with the Base-Network and Feature Extraction Networks
(no reference points are needed for alignment); (ii) the aligned ﬁngerprint proceeds to the Base-
Network which is followed by two branches; (iii) the ﬁrst branch extracts a 96-dimensional texture-
based representation; (iv) the second branch extracts a 96-dimensional minutiae-based represen-
tation, guided by a side-task of minutiae detection (via a minutiae map which does not have to
be extracted during testing); (v) the texture-based representation and minutiae-based represen-
tation are concatenated into a 192-dimensional representation of 768 bytes (192 features and 4
bytes per ﬂoat). The 768 byte template is compressed into a 200 byte ﬁxed-length representation
by truncating ﬂoating point value features into integer value features, and saving the scaling and
shifting values (8 bytes) used to truncate from ﬂoating point values to integers. The 200 byte Deep-
Print representations can be used both for authentication and large-scale ﬁngerprint search. The
minutiae-map can be used to further improve system accuracy and interpretability by re-ranking
candidates retrieved by the ﬁxed-length representation.

a 200 byte ﬁxed-length representation (192 bytes for the feature vector and 4 bytes for storing the

2 compression parameters). As a ﬁnal step, we utilize Product Quantization [93] to further com-

press the DeepPrint representations stored in the gallery, signiﬁcantly reducing the computational

requirements and time for large-scale ﬁngerprint search.

Detecting minutiae (in the form of a minutiae-map) as a side-task to representation learning

has several key beneﬁts:

• We guide our representation to incorporate domain inspired features pertaining to minutiae
by sharing parameters between the minutiae-map output task and the representation learning

task in the multi-task learning framework.

93

• Since minutiae representations are the most popular for ﬁngerprint recognition, we posit that
our method for guiding the DeepPrint feature extraction via its minutiae-map side-task falls

in line with the goal of “Explainable AI” [34].

• Given a probe ﬁngerprint, we ﬁrst use its DeepPrint representation to ﬁnd the top k candi-
dates and then re-rank the top k candidates using the minutiae-map provided by DeepPrint 1.

This optional re-ranking add-on further improves both accuracy and interpretability.

The primary beneﬁt of the 200 byte representation extracted by DeepPrint comes into play

when performing mega-scale search against millions or even billions of identities (e.g., India’s

Aadhaar [171] and the FBI’s Next Generation Identiﬁcation (NGI) databases [56]). To highlight

the signiﬁcance of this beneﬁt, we benchmark the search performance of DeepPrint against the lat-

est version SDKs (as of July, 2019) of two top performers in the NIST FpVTE 2012 (Innovatrics2

v7.2.1.40 and Veriﬁnger3 v10.04) on the NIST SD4 [129] and NIST SD14 [128] databases aug-

mented with a gallery of nearly 1.1 million rolled ﬁngerprints. Our empirical results demonstrate

that DeepPrint is competitive with these two state-of-the-art COTS matchers in accuracy while

requiring only a fraction of the search time. Furthermore, a given DeepPrint ﬁxed-length represen-

tation can also be matched in the encrypted domain via homomorphic encryption with minor loss

to recognition accuracy as shown in [14] for face recognition.

More concisely, the primary contributions of this chapter are:
• A customized deep network (Fig. 4.1), called DeepPrint, which utilizes ﬁngerprint domain
knowledge (alignment and minutiae detection) to learn and extract a discriminative ﬁxed-

length ﬁngerprint representation.

• Demonstrating in a manner similar to [177] that Product Quantization can be used to com-
press DeepPrint ﬁngerprint representations, enabling even faster mega-scale search (51 ms
1The 128 × 128 × 6 DeepPrint minutiae-map can be easily converted into a minutiae-set with n minutia:

{(x1, y1, θ1), ..., (xn, yn, θn)} and passed to any minutia-matcher (e.g., COTS A, COTS B, or [20]).

2https://www.innovatrics.com/
3https://www.neurotechnology.com/
4We note that Veriﬁnger v10.0 performs signiﬁcantly better than earlier versions of the SDK often used in the

literature.

94

search time against a gallery of 1.1 million ﬁngerprints vs. 27,000 ms for a COTS with

comparable accuracy).

• Demonstrating with a two-stage search scheme similar to [177] that candidates retrieved by
DeepPrint representations can be re-ranked using a minutiae-matcher in conjunction with

the DeepPrint minutiae-map. This further improves system interpretability and accuracy

and demonstrates that the DeepPrint features are complementary to the traditional minutiae

representation.

• Benchmarking DeepPrint against two state-of-the-art COTS matchers (Innovatrics and Ver-
iﬁnger) on NIST SD4 and NIST SD14 against a gallery of 1.1 million ﬁngerprints. Empirical

results demonstrate that DeepPrint is comparable to COTS matchers in accuracy at a signif-

icantly faster search speed.

• Benchmarking the authentication performance of DeepPrint on the NIST SD4 and NIST
SD14 rolled-ﬁngerprints databases and the FVC 2004 DB1 A slap ﬁngerprint database [109].

Again, DeepPrint shows comparable performance against the two COTS matchers, demon-

strating the generalization ability of DeepPrint to both rolled and slap ﬁngerprint databases.

• Demonstrating that homomorphic encryption can be used to match DeepPrint templates in
the encrypted domain, in real time (1.26 ms), with minimal loss to matching accuracy as

shown for ﬁxed-length face representations [14].

• An interpretability visualization which demonstrates our ability to guide DeepPrint to look

at minutiae-related features.

4.2 Prior Work

Several early works [22,89,90] presented ﬁxed-length ﬁngerprint representations using traditional

image processing techniques. In [89, 90], Jain et al. extracted a global ﬁxed-length representation

95

Table 4.2 Published Studies on Fixed-Length Fingerprint Representations

HR @ PR = 1.0%1

(NIST SD4)2

HR @ PR = 1.0%
(NIST SD14)3

Template Size

(bytes)

Algorithm

Fingercode [89, 90]

MCC [22]

Inception v3 [18]

PDC [159]
MDC [160]

N.A.
93.2%
98.65%
93.3%
99.2%
99.83%
99.75%

N.A.
91.0%
98.93%
N.A.
99.6%
99.89%
99.93%

Gallery
Size4
N.A.
2,700
250,000
2,000
2,700
2,700
1.1M

640
1,913
8,192
N.A.
1,200
1,024
200†

Finger Patches [102]
DeepPrint (proposed)
1 In some baselines we estimated the data points from a Figure (speciﬁc data points were not
reported in the paper).
2 Only 2,000 ﬁngerprints are included in the gallery to enable comparison with previous works.
(HR = Hit Rate, PR = Penetration Rate)
3 Only last 2,700 pairs (2,700 probes; 2,700 gallery) are used to enable comparison with previ-
ous works.
4 Largest gallery size used in the paper.
† The DeepPrint representation can be further compressed to only 64 bytes using product quan-
tization with minor loss in accuracy.

of 640 bytes, called Fingercode, using a set of Gabor Filters. Cappelli et al. introduced a ﬁxed-

length minutiae descriptor, called Minutiae Cylinder Code (MCC), using 3D cylindrical structures

computed with minutiae points [22]. While both of these representations demonstrated success at

the time they were proposed, their accuracy is now signiﬁcantly inferior to state-of-the-art COTS

matchers

Following the seminal contributions of [89,90] and [22], the past 10 years of research on ﬁxed-

length ﬁngerprint representations [16, 54, 96, 105, 122, 123, 164, 165, 184] has not produced a rep-

resentation competitive in terms of ﬁngerprint recognition accuracy with the traditional minutiae-

based representation. However, recent studies [18, 102, 159, 160] have utilized deep networks to

extract highly discriminative ﬁxed-length ﬁngerprint representations. More speciﬁcally, (i) Cao

and Jain [18] used global alignment and Inception v3 to learn ﬁxed-length ﬁngerprint representa-

tions. (ii) Song and Feng [159] used deep networks to extract representations at various resolutions

which were then aggregated into a global ﬁxed-length representation. (iii) Song et al. [160] further

learned ﬁxed-length minutiae descriptors which were aggregated into a global ﬁxed-length repre-

96

sentation via an aggregation network. Finally, (v) Li et al. [102] extracted local descriptors from

predeﬁned “ﬁngerprint classes” which were then aggregated into a global ﬁxed-length representa-

tion through global average pooling.

While these efforts show tremendous promise, each method has some limitations. In particu-

lar, (i) the algorithms proposed in [18] and [159] both required computationally demanding global

alignment as a preprocessing step, and the accuracy is inferior to state-of-the-art COTS matchers.

(ii) The representations extracted in [160] require the arduous process of minutiae-detection, patch

extraction, patch-level inference, and an aggregation network to build a single global feature repre-

sentation. (iii) While the algorithm in [102] obtains high performance on rolled ﬁngerprints (with

small gallery size), the accuracy was not reported for slap ﬁngerprints. Since [102] aggregates lo-

cal descriptors by averaging them together, it is unlikely that the approach would work well when

areas of the ﬁngerprint are occluded or missing (often times the case in slap ﬁngerprint databases

like FVC 2004 DB1 A), and (v) all of the algorithms, suffer from lack of interpretability compared

to traditional minutiae representations.

In addition, existing studies targeting deep, ﬁxed-length ﬁngerprint representations all lack

an extensive, large-scale evaluation of the deep features. Indeed, one of the primary motivations

for ﬁxed-length ﬁngerprint representations is to perform orders of magnitude faster large scale

search. However, with the exception of Cao and Jain [18], who evaluate against a database of

250K ﬁngerprints, the next largest gallery size used in any of the aforementioned studies is only

2,700.

As an addendum, deep networks have also been used to improve speciﬁc sub-modules of

ﬁngerprint recognition systems such as segmentation [32, 50, 124, 194], orientation ﬁeld estima-

tion [17, 140, 150], minutiae extraction [33, 125, 168], and minutiae descriptor extraction [19].

However, these works all still operate within the conventional paradigm of extracting an unordered,

variable length set of minutiae for ﬁngerprint matching.

97

Figure 4.2 Fingerprint impressions from one subject in the DeepPrint training dataset [188]. Im-
pressions were captured longitudinally, resulting in the variability across impressions (contrast and
intensity from environmental conditions; distortion and alignment from user placement). Impor-
tantly, training with longitudinal data enables learning compact representations which are invariant
to the typical noise observed across ﬁngerprint impressions over time, a necessity in any ﬁngerprint
recognition system.

4.3 DeepPrint

In the following section, we (i) provide a high-level overview and intuition of DeepPrint, (ii)

present how we incorporate automatic alignment into DeepPrint, and (iii) demonstrate how the

accuracy and interpretability of DeepPrint is improved through the injection of ﬁngerprint domain

knowledge.

4.3.1 Overview

A high level overview of DeepPrint is provided in Figure 4.1 with pseudocode in Algorithm 2.

DeepPrint is trained with a longitudinal database (Fig. 4.2) comprised of 455K rolled ﬁngerprint

images stemming from 38,291 unique ﬁngers [188]. Longitudinal ﬁngerprint databases consist of

ﬁngerprints from distinct subjects captured over time (Fig. 4.2) [188]. It is necessary to train Deep-

Print with a large, longitudinal database so that it can learn compact, ﬁxed-length representations

which are invariant to the differences introduced during ﬁngerprint image acquisition at different

times and in different environments (humidity, temperature, user interaction with the reader, and
ﬁnger injuries). The primary task during training is to predict the ﬁnger identity label c ∈ [0, 38291]
(encoded as a one-hot vector) of each of the 455K training ﬁngerprint images (≈ 12 ﬁngerprint

98

Algorithm 2 Extract DeepPrint Representation
1: L(If ): Shallow localization network, outputs x, y, θ
2: A: Afﬁne matrix composed with parameters x, y, θ
3: G(If , A): Bilinear grid sampler, outputs aligned ﬁngerprint
4: S(It): Inception v4 stem
5: E(x): Shared minutiae parameters
6: M (x): Minutia representation branch
7: D(x): Minutiae map estimation
8: T (x): Texture representation branch
9:
10: Input: Unaligned 448 × 448 ﬁngerprint image If
11: A ← (x, y, θ) ← L(If )
12: It ← G(If , A)
13: Fmap ← S(It)
14: Mmap ← E(Fmap)
15: R1 ← M (Mmap)
16: H ← D(Mmap)
17: R2 ← T (Fmap)
18: R ← R1 ⊕ R2
19: Output: Fingerprint representation R ∈ R192 and minutiae-map H. (H can be optionally
utilized for (i) visualization and (ii) fusion of DeepPrint scores obtained via R with minutiae-
matching scores.)

impressions / ﬁnger). The last fully connected layer is taken as the representation for ﬁngerprint

comparison during authentication and search.

The input to DeepPrint is a 448 × 448 5 grayscale ﬁngerprint image, If, which is ﬁrst passed
through the alignment module (Fig. 4.1). The alignment module consists of a localization network,

L, and a grid sampler, G [83]. After applying the localization network and grid sampler to If, an

aligned ﬁngerprint It is passed to the base-network, S.

The base-network is the stem of the Inception v4 architecture (Inception v4 minus Inception

modules). Following the base-network are two different branches (Fig. 4.1) comprised primarily

of the three Inception modules (A, B, and C) described in [166]. The ﬁrst branch, T (x), completes
5Fingerprint images in our training dataset vary in size from ≈ 512 × 512 to ≈ 800 × 800. As a pre-
processing step, we do a center cropping (using Gaussian ﬁltering, dilation and erosion, and thresholding) to all images
to ≈ 448 × 448. This size is sufﬁcient to cover most of the rolled ﬁngerprint area without extraneous background
pixels.

99

the Inception v4 architecture 6 as T (S(It)) and performs the primary learning task of predicting a

ﬁnger identity label directly from the cropped, aligned ﬁngerprint It. It is included in order to learn

the textural cues in the ﬁngerprint image. The second branch (Figs. 4.1 and 4.5), M (E(S(It))),

again predicts the ﬁnger identity label from the aligned ﬁngerprint It, but it also has a related side

task (Fig. 4.5) of detecting the minutiae locations and orientations in It via D(E(S(It))). In this

manner, we guide this branch of the network to extract representations inﬂuenced by ﬁngerprint

minutiae (since parameters between the minutiae detection task and representation learning task

are shared in E(x)). The textural cues act as complementary discriminative information to the

minutiae-guided representation. The two 96-dimensional representations (each dimension is a

ﬂoat, consuming 4 bytes of space) are concatenated into a 192-dimensional representation (768

total bytes). Finally, the ﬂoats are truncated from 32 bits to 8 bit integer values, compressing the

template size to 200 bytes (192 bytes for features and 8 bytes for 2 decompression parameters).

Note that the minutiae set is not explicitly used in the ﬁnal representation. Rather, we use the

minutiae-map to guide our network training. However, for improved accuracy and interpretability,

we can optionally store the minutiae set for use in a re-ranking scheme during large-scale search

operations.

In the following subsections, we provide details of the major sub-components of the proposed

network architecture.

4.3.2 Alignment

In nearly all ﬁngerprint recognition systems, the ﬁrst step is to perform alignment based on some

reference points (such as the core point). However, this alignment is computationally expensive.

This motivated us to adopt attention mechanisms such as the spatial transformers in [83].

The advantages of using the spatial transformer module in place of reference point based align-

ment algorithms are two-fold: (i) it requires only one forward pass through a shallow localization

6We selected Inception v4 after evaluating numerous other architectures such as: ResNet, Inception v3, Inception

ResNet, and MobileNet.

100

Figure 4.3 Unaligned ﬁngerprint images from NIST SD4 (top row) and corresponding DeepPrint
aligned ﬁngerprint images (bottom row).

network (Table 4.3), followed by bilinear grid sampling. This reduces the computational com-
plexity of alignment (we resize the 448 × 448 ﬁngerprints to 128 × 1287 to further speed up
the localization estimation); (ii) The parameters of the localization network are tuned to minimize

the loss (Eq. 4.3.9) of the base-network and representation extraction networks. In other words,

rather than supervising the transformation via reference points (such as the core point), we let the

base-network and representation extraction networks tell the localization network what a “good”

transformation is, so that it can learn a more discriminative representation for the input ﬁngerprint.

Given an unaligned ﬁngerprint image If, a shallow localization network ﬁrst hypothesizes the

translation and rotation parameters (x,y, and θ) of an afﬁne transformation matrix Aθ (Fig. 4.1).

A user speciﬁed scaling parameter λ is used to complete Aθ (Fig. 4.1). This scaling parameter

stipulates the area of the input ﬁngerprint image which will be cropped. We train two DeepPrint

448) meaning a
models, one for rolled ﬁngerprints (λ = 1) and one for slap ﬁngerprints (λ = 285
285 × 285 ﬁngerprint area window will be cropped from the 448 × 448 input ﬁngerprint image.
Given Aθ, a grid sampler G samples the input image If pixels (xf
i ) for every target grid location

i , yf

(xt

i) to output the aligned ﬁngerprint image It in accordance with Equation 4.3.1.

i, yt
7We also tried 64 × 64, however, we could not obtain consistent alignment at this resolution.

101

Table 4.3 Localization Network Architecture

Type

Convolution
Max Pooling
Convolution
Max Pooling
Convolution
Max Pooling
Convolution
Max Pooling

Output Size
128 × 128 × 24
64 × 64 × 24
64 × 64 × 32
32 × 32 × 32
32 × 32 × 48
16 × 16 × 48
16 × 16 × 64
8 × 8 × 64

Filter Size, Stride

5 × 5, 1
2 × 2, 2
3 × 3, 1
2 × 2, 2
3 × 3, 1
2 × 2, 2
3 × 3, 1
2 × 2, 2

Fully Connected
Fully Connected
† These three outputs correspond to x,y,θ shown in
Fig. 4.1.

64
3†

 = Aθ





xf
i

yf
i

1



xt
i

yt
i

1

(4.3.1)

Once It has been computed, it is passed on to the base-network for classiﬁcation. Finally, the

parameters for the localization network are updated based upon the loss in Equation 4.3.9.

The architecture used for our localization network is shown in Table 4.3 and images from

before and after the alignment module are shown in Figure 4.3. In order to get the localization

network to properly converge, (i) the learning rate was scaled by 0.035 and (ii) the upper bound
of the estimated afﬁne matrix translation and rotation parameters was set to 224 pixels and ±60
degrees, respectively. These constraints are based on our domain knowledge on the maximum

extent a user would rotate or translate their ﬁngers during placement on the reader platen.

102

Figure 4.4 Minutiae Map Extraction. The minutiae locations and orientations of an input ﬁnger-
print (a) are encoded as a 6-channel minutiae map (b). The “hot spots” in each channel indicate the
spatial location of the minutiae points. The kth channel of the hot spots indicate the contributions
of each minutiae to the kπ/3 orientation.

4.3.3 Minutiae Map Domain Knowledge

To prevent overﬁtting the network to the training data and to extract interpretable deep features,

we incorporate ﬁngerprint domain knowledge into DeepPrint. The speciﬁc domain knowledge we

incorporate into our network architecture is hereafter referred to as the minutiae map [20]. Note

that the minutiae map is not explicitly used in the ﬁxed-length ﬁngerprint representation, but the

information contained in the map is indirectly embedded in the network during training.

A minutiae map is essentially a 6-channel heatmap quantizing the locations (x, y) and orien-
tations θ ∈ [0, 2π] of the minutiae within a ﬁngerprint image. More formally, let h and w be the
height and width of an input ﬁngerprint image and T = {m1, m2, ..., mn} be its minutiae tem-
plate with n minutiae points, where mt = (xt, yt, θt) and t = 1, ..., n. Then, the minutiae map
H ∈ Rh×w×6 at (i, j, k) can be computed by summing the location and orientation contributions
of each of the minutiae in T to obtain the heat map (Fig. 4.4 (b)).

103

n(cid:88)

t=1

H(i, j, k) =

Cs((xt, yt), (i, j)) · Co(θt, 2kπ/6)

(4.3.2)

where Cs(.) and Co(.) calculate the spatial and orientation contribution of minutiae mt to the

minutiae map at (i, j, k) based upon the euclidean distance of (xt, yt) to (i, j) and the orientation

difference between θt and 2kπ/6 as follows:

Cs((xt, yt), (i, j)) = exp(−||(xt, yt) − (i, j)||2

2

2σ2
s

)

(4.3.3)

Co(θt, 2kπ/6) = exp(−dφ(θt, 2kπ/6)

2σ2
s

)

(4.3.4)

where σ2

s is the parameter which controls the width of the gaussian, and dφ(θ1, θ2) is the orientation

difference between angles θ1 and θ2:

|θ1 − θ2|

2π − |θ1 − θ2|

dφ(θ1, θ2) =

−π ≤ θ1 − θ2 ≤ π

otherwise.

(4.3.5)

An example ﬁngerprint image and its corresponding minutiae map are shown in Figure 4.4. A

minutiae-map can be converted back to a minutiae set by ﬁnding the local maximums in a channel

(location), and individual channel contributions (orientation), followed by non-maximal suppres-

sion to remove spurious minutiae8.

4.3.4 Multi-Task Architecture

The minutiae-map domain knowledge is injected into DeepPrint via multitask learning. Multitask

learning improves generalizability of a model since domain knowledge within the training signals

of related tasks acts as an inductive bias [23,186]. The multi-task branch of the DeepPrint architec-

ture is shown in Figures 4.1 and 4.5. The primary task of the branch is to extract a representation

8 Code for converting a minutiae set to a minutiae map and vice versa is open-sourced: https://bit.ly/2KpbPxV

104

Figure 4.5 The custom multi-task minutiae branch of DeepPrint. The dimensions inside each box
represent the input dimensions, kernel size, and stride length, respectively.

and subsequently classify a given ﬁngerprint image into its “ﬁnger identity”. The secondary task

is to estimate the minutiae-map. Since parameters are shared between the representation learning

task and the minutiae-map extraction task, we guide the minutiae-branch of our network to extract

ﬁngerprint representations that are inﬂuenced by minutiae locations and orientations. At the same

time, a separate branch in DeepPrint aims to extract a complementary texture-based representation

by directly predicting the identity of an input ﬁngerprint without any domain knowledge (Fig. 4.1).
DeepPrint extracts minutiae maps of size 128 × 128 × 6 9 to encode the minutiae locations and
orientations of an input ﬁngerprint image of size 448 × 448 × 1. The ground truth minutiae maps
for training DeepPrint are estimated using the open source minutiae extractor proposed in [20].

Note, we combine the texture branch with the minutiae branch in the DeepPrint architecture

(rather than two separate networks) for the following reasons: (i) the minutiae branch and the

texture branch share a number of parameters (the Inception v4 stem), reducing the model com-

plexity that two separate models would necessitate, and (ii) the spatial transformer (alignment

module) is optimized based on both branches (i.e.

learned alignment beneﬁts both the texture-

based and minutiae-based representations) avoiding two separate spatial transformer modules and

alignments.

9We extract maps of 128 × 128 × 6 to save GPU memory during training (enabling a larger batch size), and to

reduce disk space requirements for storage of the maps.

105

More formally, we incorporate domain knowledge into the DeepPrint representation by com-

puting the network’s loss in the following manner. First, given R1 and R2 as computed in Algo-
rithm 2, fully connected layers are applied for identity classiﬁcation logits, outputting y1 ∈ Rc
and y2 ∈ Rc, where c is the number of identities in the training set. Next, y1 and y2 are both
passed to a softmax layer to compute the probabilities ˆy1 and ˆy2 of R1 and R2 belonging to each

identity. Finally, ˆy1 and ˆy2, the ground truth label y, and the network’s parameters w, can be used

to compute the combined cross-entropy loss of the two branches and an image It:

L1(It, y) = −log(ˆyj=y

1

|It, w) − log(ˆyj=y

2

|It, w)

(4.3.6)

where j ∈ {1,··· , c}. To further reduce the intra-class variance of the learned features, we also
employ the widely used center-loss ﬁrst proposed in [179] for face recognition. In particular, we

compute two center-loss terms, one for each branch in our multi-task architecture as:

L2(It) = ||R1 − ctrn
1||2

2 + ||R2 − ctrn
2||2

2

(4.3.7)

where ctrn

i , are the branch, i, and subject, n, speciﬁc centers for a ﬁngerprint image It.

For computing the loss of the minutiae map estimation side task, we employ the Mean Squared

Error Loss between the estimated minutiae map H and the ground truth minutiae map 10 H as

follows:

L3(It, H) =

(cid:88)

(Hi,j,k − Hi,j,k)2

i,j,k

(4.3.8)

Finally, using the addition of all these loss terms, and a dataset comprised of N training images,

our model parameters w are trained in accordance with:

N(cid:88)

i=1

arg min

w

λ1L1(I i

t , yi) + λ2L2(I i

t ) + λ3L3(I i

t , H i)

(4.3.9)

10The ground truth minutiae maps are estimated using the open-source minutiae extractor in [20].

106

where {λ1 = 1, λ2 = 0.00125, λ3 = 0.095} are empirically set to obtain convergence. Note, during
the training, we augment our dataset with random rotations, translations, brightness, and cropping.

We use the RMSProp optimizer with a batch size of 30. Weights are initialized with the variance

scaling initializer. Regularization included dropout (before the embedding fully connected layer)

with a keep probability of 0.8 and weight decay of 0.00004. We trained for 140K steps, which

lasted 25 hours.

After the multitask architecture has converged, a ﬁxed length feature representation can be

acquired by extracting the fully connected layer before the softmax layers in both of the network’s
branches. Let R1 ∈ R96 be the unit-length minutiae representation and R2 ∈ R96 be the unit-length
texture representation. Then, a ﬁnal feature representation is obtained by concatenation of R1 and
R2 into R ∈ R192, followed by normalization of R to unit length.

4.3.5 Template Compression

The ﬁnal step in the DeepPrint representation extraction is template compression. In particular,

the 192-dimensional DeepPrint representation consumes a total of 768 bytes. We can compress

this size to 200 bytes by truncating the 32 bit ﬂoating point feature values to 8-bit integer values in

the range of [0,255] using min-max normalization. In particular, given a DeepPrint representation
R ∈ R192, we transfer the domain of R to R ∈ N192 and output R(cid:48), where we restrict the set of
the natural numbers N to the range of [0,255]. More formally:

(cid:22) 255(R − min(R))

max(R) − min(R)

(cid:23)

R(cid:48) =

(4.3.10)

where min(R) and max(R) output the minimum and maximum feature values of the vector R,

respectively. In order to decompress the features back to ﬂoat values for matching, we need to

save the minimum and maximum values for each representation. Thus, our ﬁnal representation is

200 bytes, 192 bytes for the features, 4 bytes for the minimum value and 4 bytes for the maximum

value. To decompress the representations (when loading them into RAM), we simply reverse the

107

min-max normalization using the saved minimum and maximum values. Table ?? shows that

compression only minimally impacts the matching accuracy.

Table 4.4 Effect of Compression on Accuracy

DeepPrint

DeepPrint

Uncompressed Features

Compressed Features

97.90%
97.50%

Dataset
NIST SD4†

97.95%
FVC 2004 DB1 A†
97.53%
† TAR @ FAR = 0.01% is reported.
† TAR @ FAR = 0.1% is reported.

4.4 DeepPrint Matching

Two, unit length, DeepPrint representations Rp and Rg can be easily matched using the cosine

similarity between the two representations. In particular:

s(Rp, Rg) = R

p · Rg
(cid:124)

(4.4.1)

Thus, DeepPrint authentication (1:1 matching) requires only 192 multiplications and 191 ad-

ditions. We also experimented with euclidian distance as a scoring function, but consistently ob-

tained higher performance with cosine similarity. Note that if compression is added, there would

be an additional d subtractions and d multiplications to reverse the min-max normalization of the

enrolled representation. Therefore, the authentication time effectively doubles. However, depend-

ing on the application or implementation, compression does not necessarily effect the search speed

since the gallery of representations could be already decompressed and in RAM before performing

a search.

108

4.4.1 Fusion of DeepPrint Score with Minutiae Score

Given the speed of matching two DeepPrint representations, the minutiae-based match scores of

any existing AFIS can also be fused together with the DeepPrint scores with minimal loss to the

overall AFIS authentication speed (i.e. DeepPrint can be easily used as an add-on to existing

minutiae-based AFIS to improve recognition accuracy). In our experimental analysis, we demon-

strate this by fusing DeepPrint scores together with the scores of minutiae-based matchers COTS

A, COTS B, and [20] and subsequently improving authentication accuracy. This indicates that the

information contained in the compact DeepPrint representation is complementary to that of minu-

tiae representations. Note, since DeepPrint already extracts minutiae as a side task, fusion with

a minutiae-based matcher requires little extra computational overhead (simply feed the minutiae

extracted by DeepPrint directly to the minutiae matcher, eliminating the need to extract minutiae a

second time).

4.5 DeepPrint Search

Fingerprint search entails ﬁnding the top k candidates, in a database (gallery or background) of

N ﬁngerprints, for an input probe ﬁngerprint. The simplest algorithm for obtaining the top k

candidates is to (i) compute a similarity measure between the probe template and every enrolled

template in the database, (ii) sort the enrolled templates by their similarity to the probe 11, and

(iii) select the top k most similar enrollees. More formally, ﬁnding the top k candidates Ck(.) in a

gallery G for a probe ﬁngerprint Rp is formulated as:

Ck(Rp) = Rankk({s(Rp, Rg)|Rg ∈ G})

(4.5.1)

where Rankk(.) returns the k most similar candidates from an input set of candidates and s is a

similarity function such as deﬁned in Equation 4.4.1.

11In our search experiments, we reduce the typical sorting time from N log(N ) to N log(k) (where k << N) by
maintaining a priority queue of size k since we only care about the scores of the top k candidates. This trick reduces
sorting time from 23 seconds to 8 seconds when the gallery size N = 100, 000, 000 and the candidate list size k = 100.

109

Since minutiae-based matching is computationally expensive, comparing the probe to every

template enrolled in the database in a timely manner is not feasible with minutiae matchers. This

has led to a number of schemes to either signiﬁcantly reduce the search space, or utilize high-level

features to quickly index top candidates [12, 21, 95, 106, 163]. However, such methods have not

achieved high-levels of accuracy on public benchmark datasets such as NIST SD4 or NIST SD14.

In contrast to minutiae-matchers, the ﬁxed-length, 200 byte DeepPrint representations can be

matched extremely quickly using Equation 4.4.1. Therefore, large scale search with DeepPrint

can be performed by exhaustive comparison of the probe template to every gallery template in

accordance with Equation 4.5.1. The complexity of exhaustive search is linear with respect to both

the gallery size N and the dimensionality d of the DeepPrint representation (d = 192 in this case).

4.5.1 Faster Search

Although exhaustive search can be effectively utilized with DeepPrint representations in conjunc-

tion with Equation 4.5.1, it may be desirable to even further decrease the search time. For example,

when searching against 100 million ﬁngerprints, the DeepPrint search time is still (11 seconds on

an i9 processor with 64 GB of RAM) 12. A natural way to reduce the search time further with

minimal loss to accuracy is to utilize an effective approximate nearest neighbor (ANN) algorithm.

Product Quantization is one such ANN algorithm which has been successfully utilized in large-

scale face search [177]. Product quantization is still an exhaustive search algorithm, however,

representations are ﬁrst compressed via keys to a lookup table, which signiﬁcantly reduces the

comparison time between two representations. In other words, product quantization reformulates

the comparison function in Equation 4.4.1 to a series of lookup operations in a table stored in RAM.

More formally, a given DeepPrint representation Rg of dimensionality d, is ﬁrst decomposed into

m sub-vectors as:

12Search time for 100 million gallery was simulated by generating 100 million random representations, where each

feature was a 32-bit ﬂoat value drawn from a uniform distribution from 0 to 1.

110

Rg = (R1, R2, ..., Rm)

(4.5.2)

j in a codebook Ci =
Next, each mth sub-vector Ri ∈ Rd/m is mapped to a codeword ci
j ∈ Rd/m} where z is the size of the codebook. The index j of each codeword ci
j=1,2,...,z|ci
{ci
can be represented as a binary code of log2(z) bits. Therefore, after mapping each sub-vector

j

to its codeword, the original d-dimensional representation Rg (d = 192 for DeepPrint) can be
compressed to only m ∗ log2(z) bits!

The codewords ci

j ∈ Rd/m for each codebook Ci are computed ofﬂine (before search time) using
k-means clustering for each sub-vector. Thus each codebook Ci contains z centroids computed
from the corresponding sub-vectors Ri. Given all m codebooks {C1,C2,C3, ...,Cm}, the product
quantizer of Rg is computed as:

q(Rg) = (q1(R1), ..., qm(Rm))

(4.5.3)

where qi(Ri) is the index of the nearest centroid in the codebook Ci, i = 1, ..., m.

Finally, given a DeepPrint probe representation Rp, and the now quantized gallery template

Rg, a match score can be obtained in accordance with Equation 4.5.4:

s(Rp, Rg) = ||Rp − q(Rg)||2

2 =

m(cid:88)

||Ri

p − qi(Ri)||2

2

(4.5.4)

i=1

Thus matching a probe template to each quantized template in the gallery requires a one-time
build up of a m × z table which is stored in RAM, followed by m lookups and additions for each
quantized template in the gallery. In our experiments, we set z = 256 and m = 64. A quantized

template in the gallery is compressed to 64 bytes, and search is reduced from 192 additions and
multiplications (N times, where N is the gallery size) to a one-time m × z table build up, followed
by 64 lookups and additions for each gallery template (a signiﬁcant savings on memory and search

time)13.

13We used the Facebook Faiss PQ implementation: https://github.com/facebookresearch/faiss

111

4.5.2 Two-stage DeepPrint Search

In addition to increasing the speed of large-scale ﬁngerprint search using DeepPrint with prod-

uct quantization, we also propose a method whereby a negligible amount of search speed can be

sacriﬁced in order to further improve the search accuracy. In particular, we ﬁrst use the Deep-

Print representations to ﬁnd the top-k14 candidates for a probe Rp in a gallery G. Then, the top-k

candidates are re-ranked using the scores of a minutiae-matcher fused together with the Deep-

Print similarity scores. More formally, given a minutiae-matcher function m(.), the k re-ranked

candidates can be computed by:

Sortk({m(mp, mg) + s(Rp, Rg)|g ∈ G})

(4.5.5)

where mp and mg two varying length minutiae templates, Rp and Rg are the two ﬁxed-length

DeepPrint templates, s(.) is the DeepPrint similarity score (either Equation 4.4.1 or Equa-

tion 4.5.4), and Sortk returns a list of k candidates sorted in descending order by similarity score.

We note that since DeepPrint already outputs a minutiae-map, which can easily be converted to

a minutiae-set, fusing DeepPrint with a minutiae matcher is quite seamless. We simply convert the

DeepPrint minutiae-maps to minutiae-sets, and subsequently input the minutiae-sets to a minutiae-

matcher such as the open-source minutiae matcher in [20].

4.6 Secure DeepPrint Matching

One of the primary beneﬁts of the ﬁxed-length, 192-dimensional DeepPrint representation is that

it can be encrypted and matched in the encrypted domain (with 192 bits of security [14]) with

fully homomorphic encryption (FHE). In particular, FHE enables performing any number of both

addition and multiplication operations in the encrypted domain. Since DeepPrint representations

can be matched using only multiplication and addition operations (Eq. 4.4.1), they can be matched

14The value of k depends on the gallery size N. For the gallery size of N = 1.1 million, we empirically selected

k = 500.

112

in the encrypted domain with minimal loss to system accuracy (only loss in accuracy comes from

converting ﬂoating point features to integer value features, resulting in a loss of precision).

In contrast, minutiae-based representations cannot be matched under FHE, since the matching

function cannot be reduced to simple addition and multiplication operations. Furthermore, exist-

ing encryption schemes for minutiae-based templates such as the fuzzy-vault, result in a loss of

matching accuracy, and are very sensitive to ﬁngerprint pre-alignment [172]. We demonstrate in

our experiments that the DeepPrint authentication performance remains almost unaltered following

FHE matching. We utilize the Fan-Vercauteren FHE Scheme [53] with improvements from [14]

for improved speed and efﬁciency15.

4.7 Datasets

We use four sources of data in our experiments. Our training data is a longitudinal dataset com-

prised of 455K rolled ﬁngerprint images from 38,291 unique ﬁngers taken from [188]. Our testing

data is comprised of both large area rolled ﬁngerprint images taken from NIST SD4 and NIST

SD14 (similar to the training data) and small area slap ﬁngerprint images from FVC 2004 DB1 A.

4.7.1 NIST SD4 & NIST SD14

The NIST SD4 and NIST SD14 databases are both comprised of rolled ﬁngerprint images

(Fig. 4.6). Due to the number of challenging ﬁngerprint images contained in both datasets (even

for commercial matchers), they continue to be popular benchmark datasets for automated ﬁnger-

print recognition algorithms. NIST SD4 is comprised of 2,000 unique ﬁngerprint pairs (total of

4,000 images), evenly distributed across the 5 ﬁngerprint types (arch, left loop, right loop, tented

arch, and whorl). NIST SD14 is a much larger dataset comprised of 27,000 unique ﬁngerprint

pairs. However, in most papers published on ﬁngerprint search, only the last 2,700 pairs from

15We use the following open-source implementation: https://github.com/human-analysis/secure-face-matching

113

Figure 4.6 Examples of poor quality ﬁngerprint images from benchmark datasets. Row 1: Rolled
ﬁngerprint impressions from NIST SD4. Row 2: Slap ﬁngerprint images from FVC 2004 DB1 A.
Rolled ﬁngerprints are often heavily smudged, making them challenging to accurately recognize.
FVC 2004 DB1 A also has several distinct challenges such as small overlapping ﬁngerprint area
between two ﬁngerprint images, heavy non-linear distortions, and extreme ﬁnger conditions (wet
or dry). Minutiae annotated with COTS A.

NIST SD14 are utilized for evaluation. To fairly compare DeepPrint with previous approaches, we

also use the last 2,700 pairs of NIST SD14 for evaluation.

4.7.2 FVC 2004 DB1 A

The FVC 2004 DB1 A dataset is an extremely challenging benchmark dataset (even for commercial

matchers) for several reasons: (i) small overlapping ﬁngerprint area between ﬁngerprint images

from the same subject, (ii) heavy non-linear distortion, and (iii) extremely wet and dry ﬁngers

(Fig. 4.6). Another major motivation for selecting FVC 2004 DB1 A as a benchmark dataset is

that it is comprised of slap ﬁngerprint images. Because of this, we are able to demonstrate that even

though DeepPrint was trained on rolled ﬁngerprint images similar to NIST SD4 and NIST SD14,

our incorporation of domain knowledge into the network architecture enables it to generalize well

to slap ﬁngerprint datasets.

114

4.8 COTS Matchers

In most all of our experiments, we benchmark DeepPrint against COTS A and COTS B (Veriﬁn-

ger 10.0 or Innovatrics v7.2.1.40, the latest version of the SDK as of July, 2019). Due to our

Non-disclosure agreement, we cannot provide a link between aliases COTS A and COTS B and

Veriﬁnger or Innovatrics. Both of these SDKs provide an ISO minutia-only template as well as

a proprietary template comprised of minutiae and other features. To obtain the best performance

from each SDK, we extracted the more discriminative proprietary templates. The proprietary tem-

plates are comprised of minutiae and other features unknown to us. We note that both Veriﬁnger

and Innovatrics are top performers in the NIST and FVC evaluations [107, 178].

4.9 Benchmark Evaluations

We begin our experiments by comparing the DeepPrint search performance to the state-of-the-art

ﬁxed-length representations reported in the literature. Then, we show that the DeepPrint repre-

sentation can also be used for state-of-the-art authentication by benchmarking against two of the

top COTS ﬁngerprint matchers in the market. We further show that this authentication can be

performed in the encrypted domain using fully homomorphic encryption. Finally, we conclude

our experiments by benchmarking the large-scale search accuracy of the DeepPrint representation

against the same two COTS search algorithms.

4.9.1 Search (1:N Comparison)

Our ﬁrst experimental objective is to demonstrate that the ﬁxed-length DeepPrint representation

can compete with the best ﬁxed-length representations reported in the academic literature [18,102]

in terms of its search accuracy on popular benchmark datasets and protocols. In particular, we

compute the Rank-1 search accuracy of the DeepPrint representation on both NIST SD4 and the

last 2,700 pairs of NIST SD14 to follow the protocol of the earlier studies.

115

Table 4.5 Benchmarking DeepPrint Search Accuracy against Fixed-Length Representations in the
Literature and COTS

Template
Description

NIST SD41
Rank-1 (%)

NIST SD142
Rank-1 (%)

Template Size
Range (kB)

Search Time

(ms)3

Fixed-Length

97.80

N.A.

8

1.0

0.2

(1.5,23.7)

(0.6,5.3)

175

16

11

72

20

Algorithm†

Inception v3
+ COTS [18]

Finger

Patches [102]

Fixed-Length

99.27

DeepPrint

Fixed-Length

COTS A

Minutiae-based4

98.70

99.55

COTS B

Minutiae-based4

92.9

99.04

99.22

99.92

92.6

1 Only 2,000 ﬁngerprints are included in the gallery to enable comparison with previous works.
2 Last 2,700 pairs are used to enable comparison with previous works.
3 Search times for all algorithms benchmarked on NIST SD4 with an Intel Core i9-7900X CPU
@ 3.30GHz
4 We use the proprietary COTS templates which are comprised of minutiae together with other
proprietary features.
† These results primarily show that (i) DeepPrints is competitive with the best ﬁxed-length rep-
resentation in the literature [102] (with a smaller template size) and state-of-the-art COTS, but
also (ii) the benchmark dataset performances are saturated due to small gallery sizes. There-
fore, in subsequent experiments we compare with state-of-the-art COTS against a background
of 1.1 million.

The results, reported in Table 4.5, indicate that the DeepPrint representation is competitive with

the most accurate search algorithm previously published in [102] (slightly lower performance on

NIST4 and slightly higher on NIST14). However, we also note that the existing benchmarks (NIST

SD4 and NISTSD14) for ﬁngerprint search have now become saturated, making it difﬁcult to

showcase the differences between published approaches. Therefore, in subsequent experiments, we

better demonstrate the efﬁcacy of the DeepPrint representation by evaluating against a background
of 1.1 million ﬁngerprints (instead of the ≈ 2K in existing benchmarks).

We highlight once again that DeepPrint has the smallest template among state-of-the-art ﬁxed

length representations (200 bytes vs 1,024 bytes for the next smallest).

116

The search performance on FVC 2004 DB1 A is not reported, since the background is not of

sufﬁcient size (only 700 slap prints) to provide any meaningful search results.

4.9.2 Authentication

We benchmark the authentication performance of DeepPrint against two state-of-the-art COTS

minutiae-based matchers, namely COTS A and COTS B. We note that none of the more recent

works on ﬁxed-length ﬁngerprint representation [18, 102, 159, 160] have considered authentica-

tion performance, making it difﬁcult for us to compare with these approaches (to the best of our

knowledge, the code for these methods is not open-sourced).

From the experimental results (Tables 4.6 and 4.7), we note that DeepPrint outperforms COTS

B on all benchmark testing protocols. We further note that DeepPrint outperforms both COTS A

and COTS B on the very challenging FVC 2004 DB1 A (Fig. 4.6). The ability of DeepPrint to

surpass COTS A and COTS B on the FVC slap ﬁngerprint dataset is a very exciting ﬁnd, given the

DeepPrint network was trained on rolled ﬁngerprint images which are comprised of very different

textural characteristics than slap ﬁngerprint impressions (Fig. 4.6). In comparison to rolled ﬁnger-

prints, slap ﬁngerprints often (i) require more severe alignment, (ii) can contain heavier non-linear

distortion, (iii) and are much smaller with respect to impression area. We posit that our injection of

domain knowledge (both alignment and minutiae detection) into the DeepPrint architecture help it

to generalize well from the rolled ﬁngerprints it was trained on to the slap ﬁngerprints comprising

FVC 2004 DB1 A. We demonstrate this further in a later ablation study.

Table 4.6 Authentication Accuracy (FVC 2004 DB1 A)

DeepPrint COTS A COTS B

DeepPrint

DeepPrint

DeepPrint

+

COTS A1

+

COTS B1

+
[20]1

97.5%†

96.75% 96.57% 98.93% 98.46%

97.6%

1 Sum score fusion is used.
† TAR @ FAR of 0.1% is reported since there are only 4,950 imposter
pairs in the FVC protocol.

117

Table 4.7 Authentication Accuracy (Rolled-Fingerprints)

Algorithm

COTS A

COTS B

Cao et al. [20]1

DeepPrint

DeepPrint +
Minutiae [20]

NIST SD4

NIST SD14

TAR @ FAR = 0.01%

99.70

TAR @ FAR = 0.01%

99.89

97.80

96.75

97.90

98.70

97.85

95.96

98.55

99.0

1 Minutiae extracted from DeepPrint minutiae-map (H) and fed di-
rectly into minutiae matcher proposed in [20].

4.9.2.1 Fusion with Minutiae-Matchers

Another interesting result with respect to the DeepPrint authentication performance is that of the

score distributions. In particular, we found that minutiae-based matchers COTS A and COTS B

have very peaked imposter distributions near 0. Indeed, this is very typical of minutiae-matchers.

In contrast, DeepPrint, has a peaked genuine distribution around 1.0, and a much ﬂatter imposter

distribution. In other words, COTS is generally stronger at true rejects, while DeepPrint is stronger

at true accepts. This complementary phenomena motivated us to fuse DeepPrint with minutiae-

based matchers to further improve their authentication performance (Table 4.6). Indeed, our re-

sults (Table 4.6) indicate that the DeepPrint representation does contain features complementary to

minutiae-based matchers, given the improvement in authentication performance under score level

fusion. We note that since a DeepPrints score can be computed with only 192 multiplications and

191 additions, it requires very little overhead for existing COTS matchers to integrate the DeepPrint

representation into their matcher.

118

4.9.2.2 Secure Authentication

In addition to being competitive in authentication accuracy with state-of-the-art minutiae match-

ers, the ﬁxed-length DeepPrint representation also offers the distinct advantage of matching in the

encrypted domain (using FHE). Here we verify that the DeepPrint authentication accuracy remains

intact following encryption. We also benchmark the authentication speed in the encrypted domain.

Our empirical results (Table 4.8) demonstrate that the authentication accuracy remains nearly the

same following FHE, and that authentication between a pair of templates takes only 1.26 millisec-

onds in the encrypted domain.

Table 4.8 Encrypted Authentication using DeepPrint Representation

Algorithm

NIST SD42 NIST SD142 FVC 2004 DB1 A3

DeepPrint

97.9%

98.55%

97.5%

DeepPrint + FHE1
1 Fully homomorphic encryption is utilized (match time: 1.26 ms).
2 TAR @ FAR = 0.01%.

3 TAR @ FAR = 0.1%

97.0%

96.9%

97.3%

4.10 Large Scale Search

Perhaps the most important attribute of the compact DeepPrint representation is its ability to per-

form extremely fast ﬁngerprint search against large galleries. To adequately showcase this feature,

we benchmark the DeepPrint search accuracy against COTS A and COTS B on a gallery of over 1.1

million rolled ﬁngerprint images. The experimental results show that DeepPrint is able to obtain

competitive search accuracy with the top COTS algorithm, at orders of magnitude faster speeds.

Note, we are unable to benchmark other recent ﬁxed-length representations in the literature against

the large scale background, since code for these algorithms has not been open-sourced.

119

4.10.1 DeepPrint Search

First, we show the search performance of DeepPrints using a simple exhaustive search technique

previously described. In particular, we match a probe template to every template in the gallery, and

select the k candidates with the highest similarity scores. We use the NIST SD4 and NIST SD14

databases in conjunction with a gallery of 1.1 million rolled ﬁngerprints. Under this exhaustive

search scheme, the DeepPrint representation enables obtaining Rank-1 identiﬁcation accuracies

of 95.15% and 94.44%, respectively (Table 4.10) and (Fig. 4.7). Notably, the search time is

only 160 milliseconds. At Rank-100, the search accuracies for both datasets cross over 99%. In

our subsequent experiments, we demonstrate how we can re-rank the top k candidates to further

improve the Rank-1 accuracy with minimal cost to the search time.

Figure 4.7 Closed-Set Identiﬁcation Accuracy of DeepPrint (with and without Product Quantiza-
tion (PQ)) on NIST SD4 and NIST SD14 (last 2,700 pairs) supplemented with a gallery of 1.1
Million. Rank-1 Identiﬁcation accuracies are 95.15% and 94.44%, respectively. Search time is
only 160 milliseconds. After adding product quantization, the search time is reduced to 51 mil-
liseconds and the Rank-1 accuracies only drop to 94.8% and 94.2%, respectively.

4.10.2 Minutiae Re-ranking

Using the open-source minutiae matcher proposed in [20], COTS A and COTS B, we re-rank

the top-500 candidates retrieved by the DeepPrint representation to further improve the Rank-1

120

Table 4.9 DeepPrint + Minutiae Re-ranking Search Accuracy (1.1 million background)

Metric

NIST SD4

Rank-1

NIST SD14

Rank-1

Search Accuracy

Search Accuracy

Search Time
(milliseconds)1

DeepPrint + [20]

98.8%

DeepPrint + COTS A2

99.45%

DeepPrint + COTS B2

98.25%

COTS A3

98.85%

98.22%

99.48%

98.41%

99.51%

300

11,000

13,000

27,472

89.2%

COTS B3

428
1 Search times benchmarked on an Intel Core i9-7900X CPU @ 3.30GHz
2 COTS only used for re-ranking the top 500 DeepPrint candidates.
3 COTS used to perform search against the entire 1.1 million gallery.

85.6%

Table 4.10 DeepPrint + PQ: Search Accuracy (1.1 million background)

Algorithm

NIST SD4

Rank 1

NIST SD14

Rank1

Search Accuracy

Search Accuracy

Search Time
(milliseconds)1

DeepPrint

95.15%

94.44%

160

DeepPrint + PQ
1 Search times benchmarked on an Intel Core i9-7900X CPU @ 3.30GHz

94.18%

94.80%

51

identiﬁcation accuracy. Following this re-ranking, we obtain competitive search accuracy as the

top COTS SDK, but at signiﬁcantly faster speeds (Table 4.9).

4.10.3 Product Quantization

We further improve the already fast search speed enabled by the DeepPrint representation by per-

forming product quantization on the templates stored in the gallery. This reduces the DeepPrint

template size to only 64 bytes and reduces the search speed down to 51 milliseconds from 160

milliseconds with only marginal loss to search accuracy (Table 4.10) and (Fig. 4.7).

121

Table 4.11 DeepPrint Representation Comparison

Metric

Minutiae

Representation1

Texture

Representation1

Fused

Representation2

FVC 2004 DB1A

TAR @

FAR = 0.1%

NIST SD4
TAR @

FAR = 0.01%

NIST SD14

TAR @

FAR = 0.01%

97.4%

90.0%

97.5%

97.0%

97.15%

97.9%

97.29%

98.14%

98.55%

1 Each representation (96 bytes) is extracted from one branch in the Deep-
2 Scores from the minutiae representation are fused
Print architecture.
with the texture representation using sum score fusion.

Table 4.12 DeepPrint Ablation Study

Metric

w/o all with alignment

with alignment

+ domain knowledge

FVC 2004 DB1A

TAR @ FAR = 0.1%

NIST SD4

TAR @ FAR = 0.1%

NIST SD14

TAR @ FAR = 0.1%

72.86%

88.0%

96.95%

96.65%

97.5%

97.9%

97.96%

96.52%

98.55%

4.11 Ablation Study

Finally, we perform an ablation study to highlight the importance of (i) the automatic alignment

module in the DeepPrint architecture and (ii) the minutiae-map domain knowledge added during

training of the network. In our ablation study, we report the authentication performance of Deep-

Print with/without the constituent modules.

122

We note that in all scenarios, the addition of domain knowledge improves authentication per-

formance (Tables 4.11 and 5.6). This is especially true for the FVC 2004 DB1 A database which

is comprised of slap ﬁngerprints with different characteristics (size, distortion, conditions) than

the rolled ﬁngerprints used for training DeepPrint. Thus we show how adding minutiae domain

knowledge enables better generalization of DeepPrint to datasets which are very disparate from

its training dataset. We note that alignment does not help in the case of NIST SD4 and NIST

SD14 (since rolled ﬁngerprints are already mostly aligned), however, it signiﬁcantly improves the

performance on FVC 2004 DB1 A where ﬁngerprint images are likely to be severely unaligned.

We also note that the minutiae-based representation and the texture-based representation from

DeepPrint are indeed complementary, evidenced by the improvement in accuracy when fusing the

scores from both representations. (Table 4.11).

4.12

Interpretability

As a ﬁnal experiment, we demonstrate the interpretability of the DeepPrint representation using the

deconvolutional network proposed in [189]. In particular, we show in Fig. 4.8 which pixels in an

input ﬁngerprint image are ﬁxated upon by the DeepPrint network as it extracts a representation.

From this ﬁgure, we make some interesting observations. In particular, we note that while the

texture branch of the DeepPrint network seems to only focus on texture surrounding singularity

points in the ﬁngerprint (core points, deltas), the minutiae branch focuses on a larger portion of

the ﬁngerprint in areas where the density of minutiae points are high. This indicates to us that our

guiding the DeepPrint network with minutiae domain knowledge does indeed draw the attention of

the network to minutiae points. Since both branches focus on complementary areas and features,

the fusion of the representations improves the overall matching performance (Table 4.11).

123

Figure 4.8 Illustration of DeepPrint interpretability. The ﬁrst row shows three example ﬁngerprints
from NIST SD4 which act as inputs to DeepPrint. The second row shows which pixels the texture
branch is focusing on as it extracts its feature representation. Singularity points are overlaid to
show that the texture branch ﬁxates primarily on regions surrounding the singularity points. The
last row shows pixels which the minutiae branch focuses on as it extracts its feature representation.
We overlay minutiae to show how the minutiae branch focuses primarily on regions surrounding
minutiae points. Thus, each branch of DeepPrint extracts complementary features which comprise
more accurate and interpretable ﬁxed-length ﬁngerprint representations than previously reported
in the literature.

4.13 Computational Resources

DeepPrint models and training code are implemented in Tensorﬂow 1.14.0. All models were

trained across 2 NVIDIA GeForce RTX 2080 Ti GPUs. All search and authentication experiments

were performed with an Intel Core i9-7900X CPU @ 3.30GHz and 32 GB of RAM.

124

4.14 Summary

We have presented the design of a custom deep network architecture, called DeepPrint, capable

of extracting highly discriminative ﬁxed-length ﬁngerprint representations (200 bytes) for both au-

thentication (1:1 ﬁngerprint comparison) and search (1:N ﬁngerprint comparison). We showed how

alignment and ﬁngerprint domain knowledge could be added to the DeepPrint network architecture

to signiﬁcantly improve the discriminative power of its representations. Then, we benchmarked

DeepPrint against two state-of-the-art COTS matchers on a gallery of 1.1 million ﬁngerprints, and

showed competitive search accuracy (DeepPrint Rank-1 of 98.8% vs. COTS 98.85% on NIST

SD4) at signiﬁcantly faster speeds (300 ms vs. 27,000 ms against a gallery of 1.1 million). We also

showed how the DeepPrint representation could be used for matching in the encrypted domain via

fully homomorphic encryption. We posit that the compact, ﬁxed-length DeepPrint representation

will signiﬁcantly aid in large-scale ﬁngerprint search. Among the three most popular biometric

traits (face, ﬁngerprint, and iris), ﬁngerprint is the only modality for which no state-of-the-art

ﬁxed-length representation is available. This work aims to ﬁll this void.

125

Chapter 5

Infant Fingerprint Recognition

1 week old

2 weeks old

4 weeks old

6 weeks old

8 weeks old

12 weeks old

Figure 5.1 Face images (top row) and corresponding left thumb ﬁngerprints (bottom row) of six
different infants under 3 months of age. Face images were captured by a Xiaomi MI A1 smartphone
camera and ﬁngerprint images were captured by our 1,900 ppi RaspiReader [49, 82] at the Saran
Ashram Hospital, a charitable organization in Dayalbagh, Agra, India.

In the previous three chapters, we focused on improving each of the sub-modules of ﬁngerprint

recognition systems (ﬁngerprint readers via the Universal Target and RaspiReader, and feature

extractors and matchers via DeepPrint). In this ﬁnal chapter, we look to extend ﬁngerprint recog-

nition to all ages with the goal of alleviating infant suffering and mortality around the world. In

particular in many of the least developed and developing countries, a multitude of infants con-

126

tinue to suffer and die from vaccine-preventable diseases and malnutrition. Lamentably, the lack

of ofﬁcial identiﬁcation documentation makes it exceedingly difﬁcult to track which infants have

been vaccinated and which infants have received nutritional supplements. Answering these ques-

tions could prevent this infant suffering and premature death around the world. To that end, we

propose Infant-Prints, an end-to-end, low-cost, infant ﬁngerprint recognition system [49]. Infant-

Prints is comprised of our (i) custom built, compact, low-cost (85 USD), high-resolution (1,900

ppi), ergonomic ﬁngerprint reader, and (ii) high-resolution infant ﬁngerprint matcher. To evaluate

the efﬁcacy of Infant-Prints, we collected a longitudinal infant ﬁngerprint database captured in 4

different sessions over a 12-month time span (December 2018 to January 2020), from 315 infants

at the Saran Ashram Hospital, a charitable hospital in Dayalbagh, Agra, India. Our experimental

results demonstrate, for the ﬁrst time, that Infant-Prints can deliver accurate and reliable recogni-

tion (over time) of infants enrolled between the ages of 2-3 months, in time for effective delivery

of vaccinations, healthcare, and nutritional supplements (TAR=95.2% @ FAR = 1.0% for infants

aged 8-16 weeks at enrollment and authenticated 3 months later).

5.1

Introduction

There are more than 600 million children living worldwide between the ages of 0-5 (years) [77]

with an additional 353,000 more newborns setting foot on the planet each and every day [78]. A

majority of these births take place in the poorest regions of the world, where it is likely that neither

the infants nor their parents will have access to any ofﬁcial identiﬁcation documents1. Even if the

infant has obtained an ofﬁcial ID, it may be fraudulent or shared with others [180–182]. Without

legitimate and veriﬁable identiﬁcation, infants are often denied access to healthcare, immunization,

1Selecting and assigning a name to the newborns can be a drawn out process in developing countries in which
parents consult immediate family members or even an astrologer for a proper name. While deciding upon a name, the
infant is simply referred to as “baby” or “daughter of”, or “son of”.

127

and nutritional supplements. This is especially problematic for infants2 (newborns to 12 months),

given that they are at their most critical stage of development.

The downstream problems caused by lack of proper infant ID in the planet’s least-developed

countries can be quantitatively seen in the ﬂat lining of global vaccination coverage. In particular,

from 2015 to 2018, the percentage of children who have received their full course of three-dose

diphtheria-tetanus-pertussis (DTP3) routine immunizations remains at about 85% [183]. This falls

short of the GAVI Alliance (formerly Global Alliance for Vaccines and Immunization3) target of

achieving global immunization coverage of 90% by 2020. According to UNICEF, 25 million chil-

dren do not receive proper annual vaccination, leading to 1.5 million child deaths per annum from

vaccine-preventable diseases4. The World Health Organization (WHO) suggests that inadequate

monitoring and supervision and lack of ofﬁcial identiﬁcation documents (making it exceedingly

difﬁcult to accurately track vaccination schedules) are key factors5.

Infant identiﬁcation is also urgently needed to effectively provide nutritional supplements. The

World Food Program (WFP), a leading humanitarian organization ﬁghting hunger worldwide, as-

sists close to 100 million people in some of the poorest regions of the world6. However, often the

food never reaches the intended beneﬁciaries because of fraud in the distribution system [180–182].

For example, the WFP found that in Yemen, a country with 12 million starving residents, food dis-

tribution records are falsiﬁed and relief is being given to people not entitled to it, preventing those

who actually need aid from receiving it [180, 181].

Accurate and reliable infant recognition would also assist in baby swapping prevention7, iden-

tifying missing or abducted children, and access to government beneﬁts, healthcare, and ﬁnancial

services throughout an infant’s lifetime.

2Infants are considered to be in the 0-12 months age range, whereas, toddlers and preschoolers are within 1-3 and

3-5 years of age, respectively [24].

3https://bit.ly/1i7s8s2
4https://www.unicef.org/immunization
5https://bit.ly/1pWn6Gn
6https://evaw-un-inventory.unwomen.org/fr/agencies/wfp
7https://bit.ly/2U5eAHn

128

3 months

3 months, 2 days

6 months

12 months

Figure 5.2 Face images (top row) and corresponding left thumb ﬁngerprints (bottom row) of an in-
fant, Meena Kumari, acquired on (a) December 16, 2018 (Meena was 3 months old), (b) December
18, 2018 (3 months, 2 days old), (c) March 5, 2019 (6 months old), and (d) September 17, 2019
(12 months old) at Saran Ashram Hospital, Dayalbagh, India. Note that as Meena ages, ﬁngerprint
details emerge such as visible pores. This level of detail is enabled by our 1,900 ppi reader.

As we show in the next section, ﬁngerprint recognition [111], is the only way to accurately

and reliably establish an infant’s identity. While ﬁngerprint recognition is now a mature ﬁeld and

billions of teenagers and adults have been using it to authenticate themselves, children, particularly

infants and toddlers, cannot yet utilize ﬁngerprint recognition to get a unique and veriﬁable digital

identity.

5.1.1 Fingerprints for Infant-ID

Conventional identiﬁcation documents (paper records) are impractical for infant recognition in

many of the least developed and developing countries because they are not securely linked to

a speciﬁc infant. Furthermore, they may be fraudulent [182], lost, or stolen. We posit that a

more accurate, robust, and veriﬁable means of infant recognition is through the use of biometric

129

Table 5.1 Related work on child and infant ﬁngerprint recognition.

2006-2009

0 year

0 - 4.5 years

5 years

N/A

1 year

Resolution # Subjects Enrollment Lapse

Year
1899
2005
2007

2009
2013
2016
2019
2019
2020
2020

0 - 13 years N/A*
0 - 12 years N/A*
0 - 18 years

Study
Galton [59]
TNO [41]
BIODEV II [141]
UltraScan [149]
Aadhar [68]
JRC [97]
Jain et al. [84]
Saggese et al. [147]1
Infant-Prints [82]
Preciozzi et al. [138]
This study
* No time span available for these studies.
1 Scores from across all time lapses (weeks or months) are aggregated when computing the ﬁn-
gerprint recognition error rates. This inﬂates the true longitudinal recognition performance.

0 - 12 years
0 - 5 years
0 - 6 months variable
3 mos.
0 - 3 mos.
0 - 18 years
10 years
0 - 3 months 1 year

Inked
500 ppi
500 ppi
500 ppi
500ppi
500 ppi
1,270 ppi
3,400 ppi
1,900 ppi
500 ppi
1,900 ppi

1
161
300
308
1.25B
2611
309
142
194

16,865

315

recognition. Of the prominent biometric traits, we posit that ﬁngerprint is the most promising for

infant recognition. This is because, (i) face recognition is challenging due to the rapid aging of the

infant’s face from infanthood to childhood [173]. (ii) Iris recognition [11] is impractical because

the infant will often be sleeping or crying. (iii) Footprint recognition [99, 104] requires removing

socks and shoes and cleaning the infant’s feet, and ﬁnally, (iv) palmprint recognition [101] requires

opening an infant’s entire hand where the concavity of the palm makes it difﬁcult to image. In

contrast, ﬁngerprint recognition has already been shown to be practical for young children [84].

Furthermore, ﬁngerprints have been shown to be (i) unique [136, 195], (ii) present at birth [7, 31,

133], (iii) stable over time in terms of recognition accuracy [85,188], and (iv) a socially acceptable

biometric trait to capture [84].

Fingerprint recognition of infants comes with its own challenges and requirements, including:

1. A compact, low-cost, ergonomic, high-resolution (to accommodate small inter-ridge spac-

ings), and high throughput ﬁngerprint reader.

2. A robust and accurate ﬁngerprint matcher to accommodate low quality (distorted, dirty, wet,

dry, motion blurred), high-resolution ﬁngerprint images.

130

As such, prevailing COTS ﬁngerprint recognition systems, designed primarily for an adult popu-

lation, are not feasible for infant ﬁngerprint recognition. Our goal then is to develop an end-to-end

ﬁngerprint recognition system, speciﬁcally designed for infants.

5.2 Related Work

Table 5.1 summarizes prior work on infant and child ﬁngerprint recognition. These studies are

summarized as follows:

• Beginning in 2004, the Netherlands Organization for Applied Scientiﬁc Research (TNO)
conducted a study [41] wherein they concluded that it was not possible to obtain clear ﬁn-

gerprints from children under 4 years of age due to low ﬁdelity in the ridge pattern on their

ﬁngers.

• A pilot program called BIODEV II was initiated in 2007 for capture, storage and veriﬁcation
of biometric data for Schengen visa applicants [141]. Experimental results based on the

ﬁngerprints of 300 children acquired in Damascus (Syria) and Ulan Bator (Mongolia), show

that it is challenging to acquire ﬁngerprints of children below 12 years of age.

• UltraScan conducted a study from 2006 to 2009 which modeled the growth of the ﬁngerprints
of children as they grow into their adolescence [149]. However, no experimental results were

provided on child ﬁngerprint capture and recognition.

• The Joint Research Center of the European Commission published a technical report [97]
in 2013 on ﬁngerprinting 2,611 children between 0 to 12 years of age. Fingerprints were

acquired using 500 ppi ﬁngerprint readers while passport processing by the Portuguese gov-

ernment. The report concluded that ﬁngerprint recognition of children younger than 6 years

of age is challenging.

131

• In 2016, Jain et al. acquired ﬁngerprints of 309 children in the age range of 0 to 5 years via
a 1,270 ppi ﬁngerprint reader [84]. They concluded that it is feasible to recognize infants

enrolled at the age of 6 months and authenticated one year later.

• In 2019, Saggese et al. acquired ﬁngerprint images of 500 newborns and infants (less than
6 months of age) at the Tijuana General Hospital in Mexico using a custom built 3,400

ppi contactless ﬁngerprint reader [147]. Although the authentication results reported seem

promising, the study does not separate out the longitudinal recognition performance.

• Perciozzi et al.

reported extremely low authentication performance of infants in a study

published in 2020 [138]. The low performance can be attributed to the fact that the infant’s

ﬁngerprints were captured with a standard 500 ppi ﬁngerprint reader.

• In our preliminary study [82], we collected ﬁngerprints of 194 infants via a custom 1,900
ppi ﬁngerprint reader. We found that infants enrolled at ages 0-3 months can be accurately

and reliably recognized 3 months later with TAR=90% @ FAR=0.1%.

Among the aforementioned studies, there are only three studies [82,138,147] which investigate

the feasibility of recognizing infants under the age of 3 months at enrollment. (i) While the infant

ﬁngerprint recognition results reported in [147] by Saggese et al. seem promising, they aggregate

scores from all time lapses (weeks or months) for computing the ﬁngerprint recognition error rates

which inﬂates the true longitudinal recognition performance. (ii) Preciozzi et al. report poor infant

recognition results (TAR = 15.61% @ FAR=0.1% for 2-3 month old age group). (iii) Our pre-

liminary study on infant ﬁngerprint recognition [82] utilized a custom 1,900 ppi infant ﬁngerprint

reader, however, the matcher was not designed to fully utilize the high-resolution imagery (instead

using existing matchers designed for 500 ppi images). Furthermore, the matcher did not incor-

porate any enhancement or aging of the friction ridge pattern. Finally, our preliminary study was

conducted for 194 infants across a maximum time lapse of 3 months. In contrast, the completed

work in this thesis includes 315 infants with longitudinal data of up to a one year time lapse.

132

The key differences between the present and prior work (speciﬁcally targeting infant recogni-

tion [82, 138, 147]) can be concisely summarized as follows:

• The longitudinal infant authentication and search performance has not been adequately ad-
dressed in prior works. In [147], ﬁngerprint pairs captured across time lapses of different

duration were lumped into the same evaluation. In our preliminary study [82], we only as-

sessed the longitudinal performance for a time lapse of 3-months. In completing this thesis,

we extend this longitudinal evaluation out to a full 12 month time lapse (requiring further

in-situ data collection).

• Prior work proposed high-resolution ﬁngerprint readers, but did not exploit the high-
resolution imagery. Instead, the existing works utilize 500 ppi ﬁngerprint matchers (designed

for the adult population).

In this thesis, we design a high-resolution ﬁngerprint matcher

speciﬁcally for infants to further improve the matching performance. Extensive ablation

studies show the impact of these algorithmic improvements.

• This is the ﬁrst comprehensive study to develop an entire, end-to-end infant ﬁngerprint
recognition system (including ﬁngerprint reader, matcher, and mobile application (Fig. 5.3)),

and then rigorously evaluate the system on a longitudinal, in-situ dataset to successfully

demonstrate that infants can be enrolled at ages of less than 3 months, and then recognized

after a time lapse of 12 months with acceptable accuracy. The study in this thesis is more

complete than any of the existing studies targeting infant ﬁngerprint recognition [82, 138,

147].

Figure 5.3 Overview of the Infant-Prints system.

133

Fingerprint CaptureInfantIDAppTextureFeature ExtractionMinutiaeSearchGalleryLatentAge Transformed & EnhancedThe speciﬁc technical contributions of our approach are as follows:

• Design and prototyping of a compact (1”×2”×3”), low-cost (85 USD), high-resolution
(1,900 ppi), ergonomic ﬁngerprint reader for infants (Fig. 5.4). This reader is much smaller

and better designed for infants than our earlier open sourced ﬁngerprint reader proposed

in [47]. We also prototype a contactless version of our ﬁngerprint reader (Fig. 5.6) in order

to compare contact-based sensing technologies with contactless sensing technologies when

used for infants.

• Collection of a longitudinal infant ﬁngerprint database comprised of 315 infants (0-3
months) over 4 separate sessions separated by 13 months (between December 2018 and

January 2020). The data was collected at the Saran Ashram hospital, Dayalbagh, India.

• A ﬁrst-of-its-kind, high resolution ﬁngerprint matcher for infants which incorporates infant
ﬁngerprint aging and enhancement modules together with high resolution texture and minu-

tiae matchers.

• The
that

experimental

results

evaluated on our

longitudinal

infant dataset

indicate

indeed,

it

is possible to enroll

infants at ages younger

than 3 months

and accurately recognize them months

later based only upon their ﬁngerprints

TAR=95.2%@FAR=1.0%,TAR=92.8%@FAR=0.1% (for infants enrolled at 2-3 months

of age, and authenticated 3 months later), TAR=85%@FAR=1.0% for infants enrolled at

2-3 months of age, and authenticated a full year later.

5.3 High-Resolution Fingerprint Reader

Almost all the ﬁngerprint readers used in government and commercial applications capture images

at a resolution of 500 ppi. This resolution is sufﬁcient to resolve adult ﬁngerprint ridges that have

an inter-ridge spacing of about 8-10 pixels. However, 500 ppi resolution is not adequate for infant

134

(a)

(b)

Figure 5.4 Prototype of the 1,900 ppi compact (1” × 2” × 3”), ergonomic ﬁngerprint reader. An
infant’s ﬁnger is placed on the glass prism with the operator applying slight pressure on the ﬁnger.
The capture time is 500 milliseconds. The prototype can be assembled in less than 2 hours. See
the video at: https://www.youtube.com/watch?v=f8tYbE9Cwd0.

(a) U.are.U 4500

(b) 500 ppi

(c) RaspiReader

(d) 1,900 ppi

Figure 5.5 An infant’s ﬁngerprints are acquired via (a) a 500 ppi commercial reader (Digital Per-
sona U.are.U 4500) and (c) our custom 1,900 RaspiReader. The captured ﬁngerprint images of the
right thumb from the commercial reader and the Infant-Prints reader for a 13 day old infant are
shown in (b) and (d), respectively. Manually annotated minutiae are shown in red circles (location)
with a tail (orientation). Blue arrows denote pores on the ridges.

ﬁngerprint capture since infant ﬁngerprints have an inter-ridge spacing of 4-5 pixels (sometimes

the width of a valley may be less than 1 pixel for an infant ﬁngerprint captured at 500 ppi).

Some cheaper readers (50 USD) reach 1,000 ppi only after upsampling the ﬁngerprint im-

age [155]. However, Jain et al. [84] showed that even at a native resolution8 of 1,270 ppi, ﬁnger-

print recognition of young infants (0-6 months) was much lower than infants 6 months and older.

The lack of an affordable, compact and high resolution ﬁngerprint reader motivated us to construct

8Native resolution is the resolution at which the sensor is capable of capturing (no upsampling or downsampling).

135

3”2”1”8MP CameraGlass PrismRaspberry PiZeroWa ﬁrst-of-a-kind, 1,900 ppi ﬁngerprint reader, called RaspiReader (Fig. 5.4), enabling capture of

high-ﬁdelity infant ﬁngerprints (Fig. 5.5), particularly those in the age range 0-3 months. Unlike

our prior efforts to build a compact and cheap reader for adults [44,47], both the cost and size of the
infant ﬁngerprint reader has been signiﬁcantly reduced (from 180 USD to 85 USD and 4”×4”×4”
to 1”× 2”× 3”). Furthermore, the ﬁngerprint reader is now more ergonomic for infant ﬁngerprints
since it has a glass prism towards the front of the reader (Fig. 5.4) rather than ﬂush with the top

of the reader (as is the case with commercial readers). Since infants frequently clench their ﬁsts

and have very short ﬁngers, placing the prism out front signiﬁcantly eases placement of an infant’s

ﬁnger on the platen (Fig. 5.5 (b)).

The entire design and 3D parts for the reader casing along with step by step assembly instruc-

tions are open sourced.9 Figure 5.5 shows that this custom 1,900 ppi ﬁngerprint reader is able to

capture (500 millisecond capture time) the minute friction ridge pattern of a 13 day old infant (both

minutiae and pores) with higher ﬁdelity than the 500 ppi Digital Persona U.are.U. 4500 reader.

We also prototype a contactless variant of our contact-based infant ﬁngerprint reader. Similar

to [147], we adopt a different size ﬁnger rest for different size thumbs. In this manner, we are

able to compare contact-based high resolution ﬁngerprint readers with the high resolution con-

tactless sensing technology. Figure 5.6 shows an example infant ﬁngerprint captured by both our

contactless and contact-based ﬁngerprint reader.

5.4 Longitudinal Fingerprint Dataset

To effectively demonstrate the utility of an infant ﬁngerprint recognition system for the applications

we have highlighted above, we must be able to show its ability to recognize a child based on

ﬁngerprints acquired months after the initial enrollment. Such an evaluation requires a longitudinal

ﬁngerprint dataset which contains ﬁngerprint images of the same infant over time at successive

intervals. Collecting such a dataset is a signiﬁcant challenge as it requires the cooperation of

9https://github.com/engelsjo/RaspiReader

136

(a)

(b) Contactless image

(c) Contact-based image

Figure 5.6 (a) Prototype of our 1,900 ppi contactless ﬁngerprint reader. During capture, an infant’s
ﬁnger is placed on top of a small, contactless, rectangular opening (annotated in red) on the reader
(the size of this opening can be adjusted with different sized slots). A camera captures the infant’s
ﬁngerprint from behind the rectangular opening. Examples of a processed (segmented, contrast
enhanced), contactless infant thumb-print (2 months old) is shown in (b) and the same infant’s
thumb-print acquired via contact-based ﬁngerprint reader in (c).

an infant’s parents in returning to the clinic multiple times for participation in the study. It also

requires working with uncooperative infants who may become hungry or agitated during the data

collection (our ergonomic ﬁngerprint reader alleviated some of these challenges).

We have collected a dataset comprised of longitudinal ﬁngerprint images of 315 infants (all en-

rolled at 0-3 months of age) at the Saran Ashram hospital in Dayalbagh, India across four sessions

(see Fig. 5.7)10:

1. Session 1: December 12-19, 2018

2. Session 2: March 3-9, 2019

3. Session 3: September 12-21, 2019

4. Session 4: January 17-24, 2020

10Our dataset collection was approved by the Institutional Review Board (IRB) of Michigan State University and
ethics committee of Dayalbagh Educational Institute and Saran Ashram Hospital. The ﬁngerprint dataset cannot be
made publicly available per the IRB regulations and parental consents.

137

Table 5.2 Infant Longitudinal Fingerprint Dataset Statistics

# Sessions
# Infants
Total # images
Age at enrollment
# Subjects with no time lapse*
# Subjects with 3 months lapse*
# Subjects with 6 months lapse*
# Subjects with 9 months lapse*
# Subjects with 12 months lapse*
Male to Female Ratio
* Time lapse between enrollment and authentication image.

4
315
3,071

0 - 3 mos.

127
121
29
101
41

43% to 57%

The infants were patients of the pediatrician, Dr. Anjoo Bhatnagar (Fig. 5.7). Prior to data collec-

tion, the parents were required to sign a consent form (approved by authors’ institutional review

board and the ethics committee of Saran Ashram hospital).

In a single session, we attempted to acquire a total of two impressions per thumb (sometimes

we captured more (e.g. 4 impressions) or less (e.g. 1 impression) depending on the cooperative

nature of the infant). Although a modest incentive was offered to parents for their data collection

efforts, it was often difﬁcult for them to meet our ﬁngerprint capture schedule because of festivals,

vacations, moving to a different city or loss of interest in the project. For this reason, out of the

315 total infants that we encountered, 25 infants were present in all four sessions, 54 infants came

to only three sessions, 109 infants came to only two sessions, and 127 infants came to only one

session. During collection, a dry or wet wipe was used, as needed, to clean the infant’s ﬁnger prior

to ﬁngerprint acquisition. On average, data capture time, for 4 ﬁngerprint images (2 per thumb)

and a face image per infant, was 3 minutes11. This enabled a reasonably high throughput during

the in-situ evaluation, akin to the operational scenario in immunization and nutrition distribution

centers. Longitudinal ﬁngerprint dataset statistics are given in Table 5.2.

11Data capture time includes parents signing the consent forms, record-keeping, and pacifying non-cooperative

infants.

138

Figure 5.7 Infant ﬁngerprint collection at Saran Ashram hospital, Dayalbagh, India. Pediatrician,
Dr. Anjoo Bhatnagar, explaining longitudinal ﬁngerprint study to the mothers while the authors
are acquiring an infant’s ﬁngerprints in her clinic. Parents also sign a consent form approved by
the Institutional Review Board (IRB) of our organizations.

5.5

Infant Fingerprint Matching

State-of-the-art ﬁngerprint feature extractors and matchers are designed to operate on 500 ppi adult

ﬁngerprint images. This limitation forced the authors in [84] to down-sample the ﬁngerprint images

captured at 1,270 ppi to enable compatibility with COTS (Commercial Off The Shelf) matchers.

The authors in [147] also had to down-sample images captured at 3,400 ppi in order to make them

compatible with adult ﬁngerprint matching systems. In our preliminary study [82], we developed

a custom Convolutional Neural Network (CNN) based texture-matcher which directly operates on

1,900 ppi ﬁngerprint images so that we did not have to down-sample images and discard valuable

discriminative cues available in high resolution images. The ﬁnal matching score in [82] was

based on the fusion of (i) our CNN-based custom texture matcher and (ii) two state-of-the-art

COTS matchers.

In completing this thesis, we (i) incorporate an enhancement and ﬁngerprint aging preprocess-

ing module, (ii) improve our high-resolution texture matcher from [82], and (iii) propose a high-

resolution minutiae extractor trained on manually annotated infant ﬁngerprint images. Combining

these algorithmic improvements with two state-of-the-art ﬁngerprint matchers (a latent ﬁngerprint

matcher, and a minutiae matcher) enables us to improve our recognition accuracy over that which

was reported in our preliminary study [82]. In the following subsections, we discuss in more detail

each of these algorithmic improvements.

139

5.5.1 Minutiae Matcher

Our high resolution minutiae matcher is comprised of (i) a high-resolution minutiae extractor, (ii)

a minutiae aging model, and (iii) the Veriﬁnger v11.0 ISO minutiae matcher. In the following

subsections, we describe each of these algorithmic components.

5.5.2 Minutiae Extraction

Recent approaches to minutiae extraction in the literature have found that deep networks are

capable of delivering superior minutiae extraction performance in comparison to traditional ap-

proaches [125, 126, 168, 191]. Furthermore, the authors in [20] showed that deep learning based

minutiae extractors are particularly well suited for low quality ﬁngerprint images. Since infant

ﬁngerprints can also be regarded as a “low-quality” ﬁngerprint (heavy non-linear distortion, mo-

tion blur from uncooperative subjects, small inter-ridge spacing, very moist or dry ﬁngers, dirty

ﬁngers), we choose to adopt the deep learning based minutiae extraction approach from [20] (with

modiﬁcations to the architecture and training procedure) for high-resolution infant minutiae ex-

traction. In our experiments, we demonstrate that the high-resolution minutiae extractor is capable

of boosting the infant ﬁngerprint recognition performance.

The core of the minutiae extraction algorithm proposed in [20] is a fully-convolutional auto-
encoder M (.) which is trained to regress from an input ﬁngerprint image I ∈ Rn×m to a ground
truth minutiae map H ∈ Rn×m×12 via ˆH = M (I), where ˆH is the predicted minutiae map. The
spatial locations of hot spots in the minutiae map indicate the locations of minutiae points, and

the 12 different channels of the minutiae map encode the orientation of the minutiae points. The

parameters of M are trained in accordance with Equation (5.5.1).

Lminutiae = || ˆH − H||2

2

(5.5.1)

This estimated 12-channel minutiae map ˆH can be subsequently converted into a variable
length minutiae set {(x1, y1, θ1), ..., (xN , yN , θN )} with N minutiae points via an algorithm which

140

Figure 5.8 Overview of the minutiae extraction algorithm. An input ﬁngerprint of any size (n× m)
is passed to the minutiae extraction network (Table 5.3). The network outputs a n×m×12 minutiae
map H which encodes the minutiae locations and orientations of the input ﬁngerprint. Finally, the
minutiae map is converted to a minutiae set {(x1, y1, θ1), ..., (xN , yN , θN )} of N minutiae.

(a)

(b)

Figure 5.9 An example infant ﬁngerprint patch (a) and the corresponding minutiae map (b). Note,
we only show 3 channels of the 12 channel minutiae map here for illustrative purposes (red channel
is the ﬁrst channel, green is the ﬁfth channel, and blue is the ninth channel). Given the full 12
channels of the minutiae map in (b), we can compute the minutiae locations (x, y) and orientations
θ of the 1,900 ppi ﬁngerprint patch in (a).

locates local maximums in the channels (locations) and individual channel contributions (orienta-

tions) followed by non-maximal suppression to remove spurious minutiae [20].

To obtain ground truth minutiae maps H for computing Lminutiae, we encode a ground truth
minutiae set T for a given infant ﬁngerprint into H following the approach of [20] for latent ﬁn-

gerprints.

An example infant ﬁngerprint patch, and a few channels of its 12 channel ground truth minutiae

map are shown in Figure 5.9. An overview of our end-to-end minutiae extraction algorithm is

shown in Figure 5.8. In contrast to the 500 ppi latent ﬁngerprint minutiae extractor in [20], we

141

Table 5.3 Minutiae Extraction Network

Type

Filter Size, Stride

Output Size
256 × 256 × 64
Convolution
128 × 128 × 64
Convolution
64 × 64 × 128
Convolution
32 × 32 × 256
Convolution
16 × 16 × 384
Convolution
8 × 8 × 512
Convolution
8 × 8 × 1024
Convolution
4 × 4 × 1024
Convolution
4 × 4 × 1024
Deconvolution
8 × 8 × 512
Deconvolution
16 × 16 × 384
Deconvolution
32 × 32 × 256
Deconvolution
64 × 64 × 128
Deconvolution
Deconvolution 128 × 128 × 64
Deconvolution 256 × 256 × 32
Deconvolution 256 × 256 × 12
† During training, input patches are 256 × 256. During
testing, the input can be of any size (the network is fully
convolutional).

4 × 4, 1
4 × 4, 2
4 × 4, 2
4 × 4, 2
4 × 4, 2
4 × 4, 2
4 × 4, 1
4 × 4, 2
4 × 4, 1
4 × 4, 2
4 × 4, 2
4 × 4, 2
4 × 4, 2
4 × 4, 2
4 × 4, 2
4 × 4, 1

directly train our minutiae extractor on infant ﬁngerprint patches at 1,900 ppi resolution. In this

manner, we do not remove any discriminative cues (via down-sampling) from the input infant

ﬁngerprint images prior to performing minutiae extraction. Operating at a high resolution requires

a deeper network architecture than that which was utilized in [20]. Our network architecture is

shown in detail in Table 5.3. Note that while we train our auto-encoder on infant ﬁngerprint

patches, during test time, we input a full size infant ﬁngerprint (of varying width and height) since

our architecture is fully-convolutional and as such, is amenable to different size inputs.

142

5.5.2.1 Manual Minutiae Markup for Training

As seen in the previous section, obtaining ground truth minutiae maps H for training our minutiae

map extraction network M (.) requires a ground truth minutiae set T for each input infant ﬁnger-

print. To obtain these ground truth minutiae sets for training, we manually annotate the minutiae

locations and orientations of 610 infant ﬁngerprints in our dataset for which we had limited lon-

gitudinal data (i.e. the infant only visited 1 or 2 sessions). These ﬁngerprints are separated from

our evaluation dataset. We manually annotated the infant ﬁngerprints using the GUI tool shown in

Figure 5.10. The tool enables the addition of new minutiae and the removal of spurious minutiae.

To make the markup task easier, we ﬁrst automatically annotate the minutiae points on the 610 in-

fant ﬁngerprints using the Veriﬁnger v11.0 minutiae extraction SDK. Then, we manually reﬁne the

Veriﬁnger annotations with our markup GUI. Each manually annotated ﬁngerprint was reviewed

multiple times by one of 4 experts in the ﬁeld of ﬁngerprint recognition.

Figure 5.10 View of the manual minutiae markup/editing software used to markup minutiae lo-
cations on a subset of infant ﬁngerprint images. These markups were later used as ground truth
to train our high resolution infant minutiae extractor. The ﬁngerprint on the left (blue annota-
tions) is coarsely annotated with Veriﬁnger v11 SDK to help speed up the annotation process. The
ﬁngerprint on the right (red annotations) shows the manually edited minutiae.

143

While the 610 manually annotated infant ﬁngerprints provide an accurate ground truth dataset

for training our minutiae extraction network, it is still small for training a deep network (Table 5.3).

Therefore, rather than training our minutiae extraction network from scratch on the 610 man-

ually annotated infant ﬁngerprints, we ﬁrst pretrain our minutiae extraction network on 9, 508

infant/child ﬁngerprints collected in [85] and coarsely annotated with minutiae using the Veriﬁn-

ger v11.0 minutiae extractor. After pretraining our minutiae extraction network on these 9, 508

coarsely annotated (using Veriﬁnger) ﬁngerprints, we ﬁnally ﬁne-tune all parameters of our net-

work (Table 5.3) using our more accurate 610 manually annotated ground truth infant ﬁngerprint

images (560 used for training, 50 used for validation). We optimize our network parameters using
the Adam optimizer and weight decay set to 4 × 10−5. When training the network on the 9, 508
coarsely annotated training data, we use a learning rate of 0.01. When ﬁne-tuning our network

(all parameters ﬁne-tuned) on our manually annotated ﬁngerprint images, we reduce the learning

rate to 0.0001. We use the minutiae detection accuracy on our 50 manually annotated validation
ﬁngerprints as a stopping criteria for the training. Finally, our network is trained on 256 × 256
patches to increase the number of training samples, and we employ data augmentations such as

random rotations, cropping, translations, and ﬂipping.

The efﬁcacy of our high-resolution minutiae extraction algorithm is shown in Fig. 5.11. In

comparison to Veriﬁnger, our algorithm extracts signiﬁcantly fewer spurious minutiae, while de-

tecting nearly all of the true minutiae locations. We show in subsequent experiments that this

results in a boost in infant ﬁngerprint recognition performance.

5.5.2.2 Minutiae Aging

After extracting a minutiae set from an infant ﬁngerprint with our high-resolution minutiae ex-

tractor, we further process the minutiae set via a minutiae aging model (Fig. 5.12). The authors

in [138] showed that by linearly scaling an infant’s ﬁngerprint image, it could be better matched

to an older ﬁngerprint impression of the same infant. Note, that although the aging model in [138]

144

(a)

(b)

(c)

(d)

(e)

(f)

(g)

(h)

(i)

(j)

Figure 5.11 Top row: Veriﬁnger minutiae detections; Bottom row: Minutiae detections from our
high-resolution minutiae extractor. Manually marked minutiae are annotated in red. Note that
Veriﬁnger detects many of the true minutiae, but also extracts a signiﬁcant number of spurious
minutiae. Our proposed minutiae extractor has slightly lower detection accuracy (of true minutiae)
than Veriﬁnger, however, it extracts signiﬁcantly fewer spurious minutiae. We further compare the
two approaches quantitatively in our experimental results.

was shown to be beneﬁcial for infant recognition, it did not result in desired levels of recognition

accuracy due in part to the fact that the infant ﬁngerprint images were captured at 500 ppi.

Rather than scaling an infant’s ﬁngerprint image as was done in [138], we directly scale the

already extracted minutiae set. More formally, given a scale factor λ and a minutiae set T of N
minutiae, where T = {(x1, y1, θ1), ..., (xN , yN , θN )}, our scaled minutiae set ˆT is given by:

ˆT = {(λx1, λy1, θ1), ..., (λxN , λyN , θN )}

(5.5.2)

To determine the scale factor λ at which an infant’s ﬁngerprint pattern grows as they age, we

select 82 pairs of our 610 manually annotated infant ﬁngerprints for which we have longitudinal

145

(a)

(b)

(c)

(d)

Figure 5.12 Effects of aging. (a) Acquired 3 month old enrollment image (orange) is overlaid on
a 1 year old probe image (blue). (b) An aged 3 month old enrollment image (orange) is overlaid
on a 1 year old probe image (blue). (c) 3 month old enrollment minutiae set (green) is overlaid on
a 1 year old probe minutiae set (red). (d) An aged 3 month old enrollment minutiae set (green) is
overlaid on a 1 year old probe minutiae set (red). Following aging (b, d), the enrollment image and
probe image (and corresponding minutiae sets) overlap better.

impressions. The range of the time lapse ∆T (in weeks) for these 82 pairs of ﬁngerprints is
12 ≤ ∆T ≤ 40 (mean ∆T = 34.3 ± 10.3). We then empirically evaluated different scalar
factors in increments of 0.05 such that the minutiae matching accuracy (as computed by Veriﬁnger

v11 SDK) on these validation images was maximized. We found that applying a scalar factor of

λ = 1.1 to infant images enrolled at less than 3 months provided the best recognition performance.

We also tried an adaptive aging model where the scalar factor was dependent upon the en-

rollment age and the elapsed time, but found no improvement in performance (likely because the

majority age group in our experiments is infants enrolled between 2-3 months and recognized 3

months later, where the simple scalar value of λ = 1.1 sufﬁces). Given similar performance, we

kept the simpler static scalar aging model as opposed to the adaptive aging model.

An example of an infant minutiae set T and its corresponding aged minutiae set ˆT is shown in

Figure 5.12. In our experiments, we quantitatively demonstrate that this scaling of the enrollment

minutiae points provides a boost to our recognition performance.

146

5.5.2.3 Minutiae Match Score

After extracting a minutiae set T (via our high-resolution minutiae extractor) and aging T into

ˆT , we compute a minutiae matching score sm between a probe infant ﬁngerprint and an enrolled

infant ﬁngerprint using the Veriﬁnger v11 ISO minutiae matcher.

5.5.3 Texture Matcher

Figure 5.13 Overview of the Infant-Prints texture matcher. We modify DeepPrint [48] to accept
1,900 ppi high resolution infant ﬁngerprint images. The network is pretrained on adult ﬁngerprint
images and then ﬁne-tuned (red layers) with the infant dataset collected in [84].

Similar to latent ﬁngerprints, infant ﬁngerprints are often of poor quality and as such are difﬁ-

cult to accurately extract minutiae from (even with our high resolution minutiae extractor). There-

fore, in addition to a minutiae match score, we also incorporate a texture matching score st using a

state-of-the-art texture ﬁngerprint matcher [48] 12. Engelsma et al. [48] proposed a CNN architec-

ture, called DeepPrint, embedded with ﬁngerprint domain knowledge for extracting discriminative

ﬁxed-length ﬁngerprint representations. Inspired by the success of DeepPrint to learn additional

textural cues that go beyond just minutiae points, we adopt this matcher for infant ﬁngerprint

recognition. In particular, we modify the DeepPrint network architecture as follows: (i) the input
size of 448 × 448 is increased to 1024 × 1024 (through the addition of convolutional layers) to
12Although DeepPrint also incorporates minutiae domain knowledge into the ﬁxed-length representation, we refer

to it as a texture matcher since minutiae points are not explicitly used for matching.

147

support 1,900 ppi images and (ii) the parameters of the added convolutional layers and the last

fully connected layer are re-trained on the 1,270 ppi (upsampled to 1,900 ppi) longitudinal infant

ﬁngerprints acquired by Jain et al. in [84] combined with 610 of our 1,900 ppi images which we

set aside for training. In total, we re-train the network with 9,683 infant ﬁngerprint images from

1,814 different thumbs. An overview of our modiﬁcations to DeepPrint is shown in Figure 5.13.

During the authentication or search stage, the CNN accepts a 1,900 ppi infant ﬁngerprint as

input and outputs a 192-dimensional ﬁxed-length representation of the ﬁngerprint. This represen-

tation can be compared to previously enrolled representations via the cosine distance between two

given representations at 10 million comparisons/second on an Intel i9 processor with 64 GB of
RAM. More formally, given an enrollment representation e ∈ R192 and a probe representation
p ∈ R192, a texture matching score st is computed as the inner product between e and p:

st = eTp

(5.5.3)

Note, in our preliminary study [82], we also used a deep learning based texture matcher sim-

ilar to DeepPrint, however, we did not incorporate minutiae domain knowledge into the texture

matcher as is done in DeepPrint (shown in Fig. 5.13). Adopting the strategy of DeepPrint in incor-

porating minutiae domain knowledge into the deep network further improves the infant recognition

performance. We show this quantitatively in the experimental results.

5.5.4 Latent Fingerprint Matcher

Finally, in addition to a state-of-the-art minutiae matcher (supplemented by our high resolution

minutiae extractor) and the ﬁne-tuned texture matcher, we include a state-of-the-art latent ﬁnger-

print matcher13 to the ﬁnal infant ﬁngerprint recognition algorithm. Before using the latent ﬁnger-

print matcher to enroll a template, we ﬁrst include two preprocessing steps: (i) enhancement, and

(ii) aging. These preprocessing steps are further described in the following subsections.

13We cannot release the name of the matcher because of a NDA, but it is one of the top performing algorithms in

the NIST ELFT evaluation [79].

148

(a)

(b)

Figure 5.14 Infant ﬁngerprint (a) before enhancement and (b) after enhancement. Looking inside
the small window (red square) we can see that the enhanced infant ﬁngerprint (b) has noticeably
improved sharpness and clarity throughout the friction ridge pattern when compared to (a).

5.5.4.1 Enhancement

Due to the low quality of the infant ﬁngerprints (motion blur, wet, dry), we incorporate an en-

hancement module to improve the sharpness and clarity of the infant friction ridge pattern. In

particular, we incorporate a state-of-the-art image super resolution model, Residual Dense Net-

work (RDN) [190]. To retrain RDN for infant ﬁngerprint enhancement, we ﬁrst add random noise

(random kernel) to the training dataset (9,683 images from [84]), followed by a gaussian blur to

simulate various types of noise in the infant ﬁngerprint images. Then, we retrain the RDN network

(8x version with a modiﬁed stride length) to regress to the clean infant ﬁngerprint images. An

example of an infant ﬁngerprint before and after enhancement is shown in Figure 5.14.

5.5.4.2

Image Aging

In a similar manner to the strategy we used to age our extracted minutiae sets, we age the enhanced

ﬁngerprint images prior to passing them to the latent ﬁngerprint matcher. The COTS latent matcher

SDK does not accept a minutiae set and as such, we must directly age the images prior to passing

them to the matcher. Therefore, if an infant’s ﬁngerprint image is captured at an age of less than

3 months, we resize the image with bicubic interpolation by a scalar factor of λ = 1.1. The scalar

factor is the same as that used to scale our minutiae sets. Finally, after enhancement and image

149

aging, we ﬁnish the latent preprocessing by resizing all images by a scalar of 0.5 in order to bring

the 1,900 ppi ﬁngerprint images to similar size as the adult ﬁngerprint images the latent matcher is

designed to operate on (this same procedure was utilized in [84]).

After preprocessing the infant ﬁngerprint images via enhancement and aging, we can enroll the

infant images via the latent SDK, and subsequently compute a match score sl.

5.5.5 Final Match Score

Our ﬁnal match score sf is a fusion of a minutiae matcher, texture matcher, and latent matcher.

In particular, given our minutiae matching score of sm, our texture match score st as deﬁned

in Equation 5.5.3, and our latent match score sl, our ﬁnal match score sf is computed by ﬁrst

normalizing each score (min-max normalization) to a range of (0, 1) and then performing sum

score fusion via:

sf = λm · sm + λt · st + λl · sl

(5.5.4)

where λm, λt, and λl are set to 0.6, 0.1, 0.3 using our validation set of 610 manually marked

ﬁngerprint images in conjunction with a grid search.

5.6 Experimental Results

In our experimental results, we ﬁrst show the authentication and search performance for all the

infants in our dataset where enrollment occurs during 0-3 months of age, and authentication or

search commences 3 months later. We ﬁrst focus on a 3 month time lapse for the following reasons.

(i) Most of our longitudinal data (121 subjects) has a time lapse of 3 months. (ii) Jain et al. already

show that once infants reach the age of 6 months, they can be enrolled and recognized a year later.

In this work, our primary aim is to bridge the gap between 0-3 months (when ﬁrst time vaccinations

commence) and 6 months. If we can effectively recognize the infants enrolled at 2-3 months and

150

Table 5.4 Infant Authentication Accuracy3 (0 − 3 months at enrollment with 3 month time lapse
between enrollment and authentication)

Algorithm

DeepPrint [48]

Veriﬁnger1

Latent Matcher2

DeepPrint + Veriﬁnger

DeepPrint + Latent Matcher
Veriﬁnger + Latent Matcher

DeepPrint + Veriﬁnger +

Latent Matcher

Enrollment Age:

Enrollment Age:

Enrollment Age:

0-1 months
(17 subjects)
17.6%, 29.4%
41.2%, 58.8%
41.2%, 47.1%
52.9%, 64.7%
41.2%, 58.8%
52.9%, 64.7%

1-2 months
(36 subjects)
27.8%, 58.3
47.2%, 55.6%
50.0%, 61.1%
55.6%, 75.0%
52.8%, 72.2%
58.3%, 75.0%

2-3 months
(83 subjects)
45.8%, 68.7%
79.5%, 86.7%
84.3%, 91.6%
86.7%, 89.2%
85.5%, 91.6%
91.6%, 92.8%

64.7%, 70.6% 63.9%, 83.3% 92.8%, 95.2%

1 Minutiae are extracted with our high-resolution minutiae extractor, then aged and fed
into the Veriﬁnger v11 ISO Matcher.
2 Images are enhanced, aged, and then fed into a state-of-the-art COTS Latent Matcher.
3 Reporting TAR @ FAR=0.1%,1.0%

authenticated or searched at 5-6 months, we can re-enroll the infants and continue to recognize the

infants longitudinally as shown in [84].

We conclude the experiments by showing the authentication and search performance of Infant-

Prints when the time lapse between the enrollment and probe images is extended to a year.

5.6.1 Experimental Protocol

To boost the infant recognition performance, we fuse scores from both of the infant’s thumbs and

also across the multiple impressions captured during the enrollment session and authentication or

search session. For example, if we successfully captured 2 ﬁngerprint images of each thumb in

the enrollment session and authentication session, we would compute a total of 8 scores using

Equation 5.5.4. These 8 scores are then fused using average fusion.

We also utilize the gender of the infant to further improve the recognition performance. In

particular, if two infants have a different gender, we set the matching score to 0.

151

Table 5.5 Infant Search Accuracy3 (0 − 3 months at enrollment with 3 month time lapse between
enrollment and search)

Algorithm

DeepPrint [48]

Veriﬁnger1

Latent Matcher2

DeepPrint + Veriﬁnger

DeepPrint + Latent Matcher
Veriﬁnger + Latent Matcher

DeepPrint + Veriﬁnger +

Latent Matcher

Enrollment Age:

Enrollment Age:

Enrollment Age:

0-1 months
(17 subjects)
52.9%, 58.8%
58.8%, 64.7%
52.9%, 58.8%
58.8%, 64.7%
52.9%, 58.8%
58.8%, 58.8%

1-2 months
(36 subjects)
63.9%, 75.0
69.4%, 77.8%
63.9%, 75.0%
69.4%, 77.8%
63.9%, 75.0%
72.2%, 80.6%

2-3 months
(83 subjects)
90.4%, 92.8%
90.4%, 91.6%
90.4%, 92.8%
90.4%, 91.6%
90.4%, 92.8%
90.4%, 91.6%

58.8%, 58.8% 72.2%, 77.8% 90.4%, 91.6%

1 Minutiae are extracted with our high-resolution minutiae extractor, then aged and fed
into the Veriﬁnger v11 ISO Matcher.
2 Images are enhanced, aged, and then fed into a state-of-the-art COTS Latent Matcher.
3 Reporting Rank 1, Rank 5 search accuracy

All imposter scores are computed by comparing impressions from one subject (both thumbs) in

a particular session to impressions from another subject (both thumbs) in another session (making

sure to only compare impressions if they belong to the same thumb).

5.6.2 Infant Authentication

Table 5.4 shows the authentication performance of the different matchers (as well as the fused

matchers) on infants enrolled between the ages of 0-3 months, and authenticated 3 months later.

From these results, we observe that none of the individual matchers perform particularly well on

any of the age groups when run standalone. However, after fusing the 3 matchers together, we start

to get reliable authentication results when the enrollment age is 2-3 months. While the longitudinal

authentication results are not yet robust for the age groups of 0-1 months and 1-2 months, we

note that vaccinations commence by the age of 3 months. By obtaining promising authentication

152

results at enrollment ages of less than 3 months, we show that ﬁngerprint authentication of infants

is indeed a potential solution for providing infants an identity for life.

5.6.3 Infant Search

Table 5.5 shows the Rank 1 search accuracy of Infant-Prints on infants enrolled between the ages

of 0-3 months, and searched 3 months later. The gallery size for our search experiment includes

every infant which was enrolled in our study (315 infants). We acknowledge that this gallery size

is small, however, we note that (i) obtaining a large gallery of infants would require signiﬁcant

resources, man-hours, and IRB regulations and approvals, and (ii) in several applications, it is very

possible that the gallery size would be of similar size to ours. For example, if the clinic which we

collected our data at were to use Infant-Prints, they would only need to manage a gallery of 315

infants, since that is the total number of infants visiting the clinic in a 1 year time period.

We note from the results of Table 5.5 that Infant-Prints is able to enroll infants at an age of

2-3 months, and search them 3 months later with a Rank 1 search accuracy of 90.4%. While work

remains to be done to further improve the performance to say 99%, we note that this is the ﬁrst

study to show promising longitudinal search performance for infants enrolled at ages as young as

2 months.

It can also be seen from Table 5.5 that each individual matcher is able to obtain the same Rank-

1 search performance (for the 2-3 month enrollment group) as the fused matcher. We acknowledge

that this can likely be explained by the small gallery size, i.e. each individual matcher is sufﬁcient

to accurately retrieve the ﬁngerprints from the smaller gallery. Given a larger gallery, it is likely

that the fused matcher would be necessary to maintain accurate search performance. Obtaining a

large scale infant dataset is an area of future research.

We also highlight that DeepPrint is able to obtain much higher search performance than au-

thentication performance (Table 5.4 vs. Table 5.5). This can be attributed to DeepPrint often times

outputting high imposter scores (creating false accepts and reducing the authentication accuracy,

153

Table 5.6 Ablated Infant Authentication Accuracy4 (0− 3 months at enrollment with 3 month time
lapse between enrollment and authentication)

Algorithm†

w/o High Resolution
Minutiae Extractor

Enrollment Age:

Enrollment Age:

Enrollment Age:

0-1 months
(17 subjects)

1-2 months
(36 subjects)

2-3 months
(83 subjects)

35.3%, 70.6%

63.9%, 83.3%

90.4%, 95.2%

w/o Aging and
Enhancement
w/o Finetuning

DeepPrint
w/o Gender
w/o All1
with All2,3

47.1%, 64.7%

50.0%, 72.2%

86.7%, 92.8%

58.8%, 64.7%

58.33%, 69.4%

90.4%, 95.2%

89.2%, 94.0%
58.8%, 64.7%
35.3%, 47.1%
86.7%, 92.8%
64.7%, 70.6% 63.9%, 83.3% 92.8%, 95.2%

52.8%, 80.6%
44.4%, 66.7%

1 Algorithm used in our preliminary study [82].
2 Minutiae are extracted with our high-resolution minutiae extractor, then aged
and fed into the Veriﬁnger v11 ISO Matcher.
3 Images are enhanced, aged, and then fed into a state-of-the-art COTS Latent
Matcher.
4 Reporting TAR @ FAR=0.1%, FAR=1.0%
† Each row removes only the modules mentioned in that row.

whereas in search high imposters are not as problematic as long as the true mate gives the highest

score).

5.6.4 Ablations

To highlight the hardware and algorithmic contributions of Infant-Prints, we show an algorithmic

ablation study in Tables 5.6, 5.7, 5.8, and 5.9, and a hardware ablation study in Table 5.10.

From Table 5.6, we see the performance of the “fused matcher” (Veriﬁnger + COTS Latent

Matcher + DeepPrint) following every algorithmic improvement (high-resolution minutiae extrac-

tion, aging, enhancement, ﬁnetuning DeepPrint, and gender meta-data). Notably, each algorithmic

improvement contributes to the overall best performance shown in the ﬁnal row. We also note that

154

Table 5.7 Ablated Veriﬁnger Performance

Algorithm

Veriﬁnger

Veriﬁnger + Aging
Veriﬁnger + Aging

+ Enhancement

Veriﬁnger + Aging

+ Enhancement
+ HR Minutiae3

0-1 months
(17 subjects)2

1-2 months
(36 subjects)

2-3 months
(83 subjects)

17.6%1
23.5%
29.4%
(35.3%)4

41.2%
(64.7%)

36.1%
44.4%
52.8%
(63.9%)

47.2%
(63.9%)

74.7%
74.7%
85.5%
(90.4%)

79.5%
(92.8%)

1 TAR @ FAR = 0.1% after a time lapse of 3 months from enroll-
ment age.
2 Indicates enrollment ages (authentication occurs 3 months later).
3 HR Minutiae denotes a minutiae set extracted by our high-
resolution,
infant minutiae extractor, and fed into Veriﬁnger’s
matcher.
4 Performance when fused with other matchers (shown in paren-
thesis) demonstrates that although HR Minutiae does not help the
stand-alone performance of Veriﬁnger, it does help when fusing
with the other matchers. This is explained further in the text.

our algorithm (last row of Table 5.6) is signiﬁcantly improved over our previous algorithm (second

to last row of Table 5.6) used in our preliminary study [82].

In Tables 5.7 and 5.8 we note that aging and enhancement both improve the “stand-alone” per-

formance of Veriﬁnger and the COTS latent matcher. Although our high-resolution minutiae ex-

tractor does not improve the stand-alone performance of Veriﬁnger (“HR Minutiae” in Table 5.7),

it does help when fusing Veriﬁnger with the other matchers (as shown in parenthesis). The reason

for this is because the Veriﬁnger minutiae extractor performs worse than our HR minutiae extractor

on low quality, noisy ﬁngerprints, but better than our minutiae extractor on higher quality images.

By improving Veriﬁnger on the lower quality image pairs with our HR minutiae extractor, we can

improve the fused matching performance, since the other matchers are already sufﬁcient to hold

the matching performance on the higher quality pairs. This can be seen visually in Figure 5.15.

When extracting minutiae with Veriﬁnger (Fig. 5.15) (a)), many spurious minutiae are marked, and

155

Table 5.8 Ablated COTS Latent Matcher (LM) Performance

Algorithm

COTS LM3

COTS LM + Aging
COTS LM + Aging

+ Enhancement

0-1 months
(17 subjects)2

1-2 months
(36 subjects)

2-3 months
(83 subjects)

35.3%1
35.3%

41.2%

41.7%
44.4%

50.0%

77.1%
80.7%

84.3%

1 TAR @ FAR = 0.1% after a time lapse of 3 months from enroll-
ment age.
2 Indicates enrollment ages (authentication occurs 3 months later).
3 COTS LM does not enable using our own HR minutiae set.

Table 5.9 Ablated DeepPrint Performance

Algorithm

0-1 months
(17 subjects)2

1-2 months
(36 subjects)

2-3 months
(83 subjects)

41.0%

45.8%

27.8%

11.8%1

17.6%

22.2%

DeepPrint
DeepPrint
+ Finetuning
1 TAR @ FAR = 0.1% after a time lapse of 3 months from en-
rollment age.
2 Indicates enrollment ages (authentication occurs 3 months
later).

Veriﬁnger is unable to establish any true minutiae correspondences between the enrollment image

and the probe image. In contrast, our minutiae extractor extracts the minutiae more reliably on

this low quality ﬁngerprint pair (Fig. 5.15) (b)), enabling Veriﬁnger to establish enough minutiae

correspondences to ﬂip the example pair from a False Reject to a True Accept.

Table 5.9 shows the ablated performance of DeepPrint. Finetuning the model on infant ﬁn-

gerprints again boosts the performance. Although the performance of DeepPrint is lower than the

other matchers stand-alone, it still boosts the overall matching performance (Table 5.4) when fused

with other matchers due to the complementary texture features it extracts. We do not age ﬁnger-

prints prior to DeepPrint extraction since DeepPrint is trained on images of varying scale as a data

augmentation method during training. Furthermore, we do not enhance images prior to DeepPrint

156

(a)

(b)

Figure 5.15 Flipping a False Reject case to a True Accept by using our high-resolution minutiae
extractor. (a) Minutiae are both extracted and matched using Veriﬁnger. The signiﬁcant number
of spurious minutiae extracted by Veriﬁnger render it impossible for Veriﬁnger to establish minu-
tiae correspondences. (b) Minutiae are extracted using our high-resolution minutiae extractor and
subsequently fed into Veriﬁnger. Because our minutiae extractor is much more resistant to spuri-
ous minutiae (on infant ﬁngerprints) than Veriﬁnger’s minutiae extractor, the Veriﬁnger matcher is
able to establish enough true minutiae correpondences to ﬂip this False Reject to a True Accept.
Quantitatively speaking, the Veriﬁnger match score is improved from 23 to 48.

extraction as our goal is to have DeepPrint extract complementary textural features which may be

discarded post-enhancement.

Finally, we show in our hardware ablation study in Table 5.10 that our contact-based high-

resolution (1,900 ppi) ﬁngerprint reader enables higher infant ﬁngerprint authentication perfor-

mance than a COTS 500 ppi contact-based reader (Digital Personna). We note that there are fewer

subjects in Table 5.10 than Table 5.4. This is because Table 5.10 only considers those subjects

which were collected on both the MSU RaspiReader and the Digital Persona reader. The differ-

ence in subject counts on the MSU RaspiReader and the Digital Persona reader can be attributed

to failure to captures on the Digital Persona (often times the ergonomics of the Digital Persona

reader (Fig. 5.5 (a)) prevented us from imaging the infant’s ﬁngerprints before the infant became

too distressed).

We also show in Figure 5.16 that the contact-based RaspiReader genuine and imposter scores

are much more separated than the contactless-based RaspiReader (TAR=72.9% vs. TAR=35.6%

@ FAR=1.0%). We show score histograms (of single ﬁnger comparisons) to compare these two

157

Table 5.10 Ablated Fingerprint Reader Authentication Results

Reader

Digital Persona

(500 ppi)

MSU RaspiReader

(1,900 ppi)

0-1

months2

1-2

months

2-3

months

(12 subjects)

(31 subjects)

(73 subjects)

0%1

35.5%

52.1%

58.3%

64.5%

93.2%3

1 TAR @ FAR = 0.1% after a time lapse of 3 months from enroll-
ment age.
2 Indicates enrollment ages (authentication occurs 3 months later).
3 Differs from Table 5.4 because of a different number of subjects.

readers since we only utilized the contactless reader during our last collection session for a limited

number of subjects. Our ﬁndings of better separation between the contact ﬁngerprint pairs than

the contactless ﬁngerprint pairs contradict the study of [147] which found that high-resolution,

contactless infant ﬁngerprints outperformed high-resolution contact-based infant ﬁngerprints. We

found it very difﬁcult to match contactless infant ﬁngerprints since contactless ﬁngerprints have a

perspective deformation (certain parts of the ﬁnger are further from the camera than others), and the

contrast is lower than FTIR ﬁngerprint images. Similar observations about the difﬁculty of match-

ing contactless ﬁngerprint images have been noted in the literature [103]. In an effort to improve

the contactless matching performance, we ﬁne-tuned DeepPrint on 23, 416 contactless ﬁngerprints

from 3, 276 ﬁngers from contactless databases released in [36, 103, 110, 148, 192, 193]. We also

attempted to normalize the ridge spacing of the contactless ﬁngerprints as was done in [147]. The

ﬁne-tuning did improve the contactless matching performance, but did not bridge the gap to the

contact ﬁngerprint matching performance.

Example of failure cases (False Accept, False Reject) are shown in Fig. 5.17. These images

highlight the difﬁculty and challenges of doing accurate infant ﬁngerprint recognition over time

(moisture, distortion, small inter-ridge spacing, ﬁngerprint aging).

158

(a)

(b)

Figure 5.16 Score Histograms comparing the contact-based RaspiReader with the contactless
RaspiReader (single ﬁnger performance). Using a contact-based reader shows much better score
separation than the contactless reader (TAR=72.9% vs. TAR=35.6% @ FAR=1.0%).

5.6.5 Longitudinal Recognition

Table 5.11 Longitudinal Search Results

Time Lapse: 3 months Time Lapse: 9 months Time Lapse: 12 months

95%1, 2

90%

90%

1 Reporting Rank 1 Search Accuracy (Gallery of 315 Infants)
2 Differs from Table 5.5 because of a different number of subjects.

Table 5.12 Longitudinal Authentication Results

Time Lapse: 3 months Time Lapse: 9 months Time Lapse: 12 months

95%1,2

90%

85%

1 Reporting TAR @ FAR = 0.1%
2 Differs from Table 5.4 because of a different number of subjects.

As a ﬁnal study, we show the longitudinal search accuracy (Table 5.11) and authentication

accuracy (Table 5.12) for infants enrolled at 2-3 months. For this experiment, we selected 20

infants from our total of 315 which were present in all 4 sessions of the data collection and were

2-3 months of age at the ﬁrst time enrollment (since our earlier studies showed that 2-3 months is

the age at which recognition ﬁrst becomes feasible). Although we have more subjects at individual

159

(a)

(b)

(c)

(d)

Figure 5.17 Example Infant-Prints failure cases.
(a, b) Example of a False Accept due to the
similar friction ridge patterns, and the moisture in the enrollment image (a). (c, d) Example of a
False Reject due to the motion blur of the uncooperative infant (d). These images highlight several
of the challenges in infant ﬁngerprint recognition.

time lapses, we chose the 20 infants which were present in all 4 sessions so that we can observe

the impact that time has on the recognition performance whilst ﬁxing the subjects used in the

experiments.

Tables 5.11 and 5.12 show that the authentication and search performance stays relatively stable

over time. In particular, from 3 months of elapsed time to 9 months of elapsed time, only one infant

drops off from being properly searched or authenticated. From 9 months to 12 months, the search

accuracy remains unchanged, while only one fewer infant is unable to be authenticated.

Notably, these are the ﬁrst results to show that it is possible to enroll infants at 2 months old

and authenticate them or search them a year later with relatively high accuracy. This highlights the

applicability of ﬁngerprints to address the challenges of this chapter. Namely, can we recognize an

infant from their ﬁngerprints in order to better facilitate accurate and fast delivery of vaccinations

and nutritional supplements to infants in need.

5.7 Summary

A plethora of infants around the world continue to suffer and die from vaccine related diseases and

malnutrition. A major obstacle standing in the way of delivering the vaccinations and nutrition

needed to the infants most in need is the means to quickly and accurately identity or authenticate

160

an infant at the point of care. To address this challenge, we proposed Infant-Prints, and end-to-end

infant ﬁngerprint recognition system. We have shown that Infant-Prints is capable of enrolling

infants as young as 2 months of age, and recognizing them an entire year later. This is the ﬁrst ever

study to show the feasibility of recognizing infants enrolled this young after this much time gap.

It is our hope that this feasibility study and Infant-Prints motivate a strong push in the direction

of ﬁngerprint based infant ﬁngerprint recognition systems which can be used to alleviate infant

suffering around the world. In doing so, we believe that the work outlined in this chapter will

make a major dent in Goal #3 of the United Nations Sustainable Development Goals, namely,

“Ensuring healthy lives and promoting well-being for all, at all ages.”

161

Chapter 6

Summary

6.1 Contributions

In this thesis, we have worked to develop ﬁngerprint recognition systems which are more i) robust,

ii) secure, iii) fast, and iv) applicable to all ages. These speciﬁc contributions are listed below.

• Robust: We have improved the robustness and reliability of ﬁngerprint recognition systems
through the design and manufacturing of Universal 3D Wearable Fingerprint Targets. Our

Universal Fingerprint Targets can be imaged by all major types of ﬁngerprint sensing tech-

nologies (unlike existing 3D ﬁngerprint targets) and are thus useful for ﬁngerprint reader

interoperability studies and other operational evaluations of ﬁngerprint readers. By evaluat-

ing ﬁngerprint readers with realistic ﬁngerprint targets, as opposed to the existing trivial 2D

calibration patterns, ﬁngerprint readers can be better assessed and improved.

• Fingerprint Reader Security: We have enhanced the security of ﬁngerprint recognition
systems via our open-source, 1,900 ppi RaspiReader. RaspiReader uses a built-in spoof

(fake ﬁngerprint attack) detection algorithm based upon its simultaneously captured direct-

view and FTIR ﬁngerprint images. We have demonstrated that RaspiReader obtains state-

of-the-art levels of ﬁngerprint spoof detection accuracy. We have also demonstrated that

RaspiReader generalizes well to spoofs fabricated from materials not seen during training

162

of the RaspiReader spoof detection algorithm. A DIY video showing the assembly process

of RaspiReader (from ubiquitous components easily found on Amazon) has been published

to YouTube1, and the 3D parts and capture software for RaspiReader are open-sourced on

Github2.

• Fingerprint Template Security: In addition to securing the ﬁngerprint reader module of
ﬁngerprint recognition systems via RaspiReader, we also better secure ﬁngerprint templates

with our ﬁxed-length ﬁngerprint representation in conjunction with a fully homomorphic

encryption matching scheme.

In particular, our 192D DeepPrint representations can be

matched within the encrypted domain, without loss of accuracy, in 1.25 milliseconds, pre-

venting hackers from stealing the template in the database and also during the matching

routine. In contrast, prevailing systems must either i) unencrypt a template prior to matching

(leaving them vulnerable to hackers) or ii) sacriﬁce system accuracy to keep the templates

encrypted at all times.

• Matching Speed: By using our 192D DeepPrint representations, we not only improve ﬁn-
gerprint template security, we also enable orders of magnitude faster large scale search. In

particular, while prevailing ﬁngerprint matching algorithms utilize expensive graph matching

algorithms for comparison, DeepPrint representations can be quickly matched with simple

distance metrics. Quantitatively, we showed that a state-of-the-art commercial matcher takes

27 seconds to search a ﬁngerprint against a 1.1 million background. In contrast, DeepPrint

takes only 300 milliseconds to search, but obtains comparable levels of Rank-1 search accu-

racy (DeepPrint: 98.80% vs. COTS: 98.85%).

• Extension to all Ages: Finally, we concluded the thesis by working to extend ﬁngerprint
recognition systems to all ages. In particular, we developed a high-resolution (1,900 ppi)

infant ﬁngerprint reader and an accompanying high-resolution infant ﬁngerprint matcher.

We then showed in a study of 315 infants, conducted over a time period of 1 year, that

1bit.do/RaspiReader
2https://github.com/engelsjo/RaspiReader

163

we could enroll infants at an age of 2-months and still recognize them over a year later.

We call our infant ﬁngerprint matcher Infant-Prints. Infant-Prints could provide signiﬁcant

global good in alleviating child suffering and death around the world via better vaccination

tracking and government beneﬁts and assistance.

6.2 Suggestions for Future Work

The following are directions of ongoing and future research:

• Encrypted Fingerprint Search: Although DeepPrint enables practical encrypted authen-
tication (1.25 milliseconds per encrypted match), the time for encrypted search against large

galleries remains impractical. We are working on developing an encrypted search algorithm

that enables ﬁngerprint search in the encrypted domain in real-time [46].

• Large Scale Fingerprint Synthesis: A primary beneﬁt of DeepPrint is its ability to do fast
large-scale search. However, the largest database available to us to evaluate this search is

only 1.1 million ﬁngerprints. One possibility to further evaluate DeepPrint is to synthesize a

gallery of 1 billion ﬁngerprints [118]. These ﬁngerprints must be both realistic and unique.

Given the increased concern of privacy over publicly available ﬁngerprint data, synthesizing

realistic ﬁngerprint data could also be useful for augmenting the training data of DeepPrint.

• One-class Spoof Detection: In [45] we developed a one-class classiﬁer for ﬁngerprint spoof
detection (using images from RaspiReader) which better generalized to unseen materials.

This method can be further developed and combined with existing state-of-the-art two-class

ﬁngerprint spoof detection algorithms to obtain state-of-the-art in both the seen and unseen

material evaluations.

164

BIBLIOGRAPHY

165

BIBLIOGRAPHY

[1] Aditya Abhyankar and Stephanie Schuckers. Integrating a wavelet based perspiration live-

ness check with ﬁngerprint recognition. Pattern Recogn., 42(3):452–464, March 2009.

[2] AEP Technology. NANOMAP-500LS.

http://www.aeptechnology.com/pdfs/NanoMap-

500LS.pdf.

[3] Multi camera adapter module for raspberry pi. https://www.arducam.com/multi-camera-

adapter-module-raspberry-pi. Accessed: 2017-4-15.

[4] S. S. Arora, A. K. Jain, and N. G. Paulter. 3d whole hand targets: Evaluating slap and
contactless ﬁngerprint readers. In 2016 International Conference of the Biometrics Special
Interest Group (BIOSIG), pages 1–8, Sept 2016.

[5] Sunpreet S. Arora, Kai Cao, Anil K. Jain, and Nicholas G. Paulter. Design and fabrica-
tion of 3d ﬁngerprint targets. IEEE Transactions on Information Forensics and Security,
11(10):2284–2297, October 2016.

[6] Sunpreet S. Arora, Anil K. Jain, and Nicholas G. Paulter. Gold ﬁngers: 3d targets for
evaluating capacitive readers. IEEE Transactions on Information Forensics and Security,
pages 1–1, April 2017.

[7] W Babler. Embryologic development of epidermal ridges and their conﬁgurations. Birth

defects original article series, 27(2):95–112, 1991.

[8] Denis Baldisserra, Annalisa Franco, Dario Maio, and Davide Maltoni. Fake ﬁngerprint
In International Conference on Biometrics, pages 265–272.

detection by odor analysis.
Springer, 2006.

[9] Jennelle Baptiste.

Resistivity of gold.

https://hypertextbook.com/facts/2004/

JennelleBaptiste.shtml.

[10] Mauro Barni, Tiziano Bianchi, Dario Catalano, Mario Di Raimondo, Ruggero Donida La-
bati, Pierluigi Failla, Dario Fiore, Riccardo Lazzeretti, Vincenzo Piuri, Alessandro Piva,
et al. A Privacy-Compliant Fingerprint Recognition System based on Homomorphic En-
In 2010 Fourth IEEE International Conference on
cryption and Fingercode Templates.
Biometrics: Theory, Applications and Systems (BTAS), pages 1–7. IEEE, 2010.

[11] HC Beck, I Ezon, L Flom, C Pitchford, and L Park. Iris recognition technology in newborns.

Investigative Ophthalmology & Visual Science, 49(13):2265–2265, 2008.

[12] Bir Bhanu and Xuejun Tan. Fingerprint Indexing Based on Novel Features of Minutiae
Triplets. IEEE Transactions on Pattern Analysis and Machine Intelligence, 25(5):616–622,
2003.

166

[13] Rex Black. Managing the Testing Process: Practical Tools and Techniques for Managing

Hardware and Software Testing. Wiley Publishing, 3rd edition, 2009.

[14] Vishnu Naresh Boddeti. Secure Face Matching using Fully Homomorphic Encryption. In
2018 IEEE 9th International Conference on Biometrics Theory, Applications and Systems
(BTAS), pages 1–10. IEEE, 2018.

[15] Zinelabidine Boulkenafet, Jukka Komulainen, and Abdenour Hadid. Face spooﬁng de-
IEEE Trans. Information Forensics and Security,

tection using colour texture analysis.
11(8):1818–1830, 2016.

[16] Julien Bringer and Vincent Despiegel. Binary Feature Vector Fingerprint Representation
from Minutiae Vicinities. In 2010 Fourth IEEE International Conference on Biometrics:
Theory, Applications and Systems (BTAS), pages 1–6. IEEE, 2010.

[17] Kai Cao and Anil K Jain. Latent Orientation Field Estimation via Convolutional Neural
Network. In Biometrics (ICB), 2015 International Conference on, pages 349–356. IEEE,
2015.

[18] Kai Cao and Anil K Jain. Fingerprint Indexing and Matching: An Integrated Approach. In

IEEE International Joint Conference on Biometrics (IJCB), pages 437–445. IEEE, 2017.

[19] Kai Cao and Anil K Jain. Automated Latent Fingerprint Recognition. IEEE Transactions

on Pattern Analysis and Machine Intelligence, 2018.

[20] Kai Cao, Dinh-Luan Nguyen, Cori Tymoszek, and Anil K Jain. End-to-end latent ﬁngerprint

search. IEEE Transactions on Information Forensics and Security, 15:880–894, 2019.

[21] Raffaele Cappelli, Matteo Ferrara, and Davide Maltoni. Fingerprint Indexing based on
Minutia Cylinder-Code. IEEE Transactions on Pattern Analysis and Machine Intelligence,
33(5):1051–1057, 2010.

[22] Raffaele Cappelli, Matteo Ferrara, and Davide Maltoni. Minutia Cylinder-Code: A New
Representation and Matching Technique for Fingerprint Recognition. IEEE Transactions
on Pattern Analysis and Machine Intelligence, 32(12):2128–2141, 2010.

[23] Rich Caruana. Multitask Learning. Machine Learning, 28(1):41–75, Jul 1997.

[24] Centers for Disease Control and Prevention. Positive Parenting Tips. https://www.cdc.
[Online; accessed 15-

gov/ncbddd/childdevelopment/positiveparenting/index.html, 2018.
February-2019].

[25] Jae Young Choi, Konstantinos N. Plataniotis, and Yong Man Ro. Using colour local binary
In Proceedings of the International Conference on

pattern features for face recognition.
Image Processing, ICIP 2010, September 26-29, Hong Kong, pages 4541–4544, 2010.

[26] Tarang Chugh, Kai Cao, and Anil Jain. Fingerprint spoof detection using minutiae-based
local patches. In 2017 IEEE International Joint Conference on Biometrics (IJCB). IEEE,
2017.

167

[27] Tarang Chugh, Kai Cao, and Anil K Jain. Fingerprint spoof buster: Use of minutiae-centered
patches. IEEE Transactions on Information Forensics and Security, 13(9):2190–2202, 2018.

[28] Tarang Chugh and Anil K Jain. Fingerprint spoof detector generalization. IEEE Transac-

tions on Information Forensics and Security, 16:42–55, 2020.

[29] Paolo Cignoni, Massimiliano Corsini, and Guido Ranzuglia. Meshlab: an open-source 3d

mesh processing system. ERCIM News, 2008(73):45–46, April 2008.

[30] Catherine C Cooksey, Benjamin K Tsai, and David W Allen. Spectral reﬂectance variability

of skin and attributing factors. In Proc. SPIE, volume 9461, page 94611M, 2015.

[31] Harold Cummins and Charles Midlo. Finger Prints, Palms and Soles: An Introduction to

Dermatoglyphics, volume 319. Dover Publications New York, 1961.

[32] Xiaowei Dai, Jie Liang, Qijun Zhao, and Feng Liu. Fingerprint Segmentation via Convolu-
tional Neural Networks. In Chinese Conference on Biometric Recognition, pages 324–333.
Springer, 2017.

[33] Luke Nicholas Darlow and Benjamin Rosman. Fingerprint Minutiae Extraction using Deep
In IEEE International Joint Conference on Biometrics (IJCB), pages 22–30.

Learning.
IEEE, 2017.

[34] Explainable Artiﬁcial Intelligence (XAI).

https://www.darpa.mil/program/explainable-

artiﬁcial-intelligence.

[35] Ashim K Datta. Advances in Fingerprint Technology. CRC press, 2001.

[36] Debayan Deb, Tarang Chugh, Joshua Engelsma, Kai Cao, Neeta Nain, Jake Kendall,
arXiv preprint

and Anil K Jain. Matching ﬁngerphotos to slap ﬁngerprint images.
arXiv:1804.08122, 2018.

[37] Ofﬁce of Biometric Identity Management Identiﬁcation Services. https://www.dhs.gov/

obim-biometric-identiﬁcation-services.

[38] Yaohui Ding and Arun Ross. An ensemble of one-class svms for ﬁngerprint spoof detec-
tion across different fabrication materials. In IEEE International Workshop on Information
Forensics and Security, WIFS 2016, Abu Dhabi, December 4-7, 2016, pages 1–6, 2016.

[39] Bernadette Dorizzi, Raffaele Cappelli, Matteo Ferrara, Dario Maio, Davide Maltoni, Nesma
Houmani, Sonia Garcia-Salicetti, and Aur´elien Mayoue. Fingerprint and on-line signature
In International Conference on Biometrics, pages
veriﬁcation competitions at icb 2009.
725–732. Springer, 2009.

[40] Dow

Corning.

Sylgard

PDMS

184

Data

Sheet.

http://www.dowcorning.com/DataFiles/090276fe80190b08.pdf.

[41] Dutch Ministry of the Interior and Kingdom Relations. Evaluation Report Biometrics Trial

2b or not 2b. http://www.dematerialisedid.com/PDFs/88630ﬁle.pdf, 2004.

168

[42] Christopher Edwards and Ronald Marks. Evaluation of biomechanical properties of human

skin. Clinics in Dermatology, 13(4):375 – 380, 1995. Bioengineering of the Skin.

[43] Joshua J Engelsma, Sunpreet S Arora, Anil K Jain, and Nicholas G Paulter. Universal 3d
wearable ﬁngerprint targets: Advancing ﬁngerprint reader evaluations. IEEE Transactions
on Information Forensics and Security, 13(6):1564–1578, 2018.

[44] Joshua J Engelsma, Kai Cao, and Anil K Jain. Fingerprint match in box. IEEE Biometrics

Theory Applications and Systems, 2018.

[45] Joshua J Engelsma and Anil K Jain. Generalizing ﬁngerprint spoof detector: Learning
a one-class classiﬁer. In 2019 International Conference on Biometrics (ICB), pages 1–8.
IEEE, 2019.

[46] Joshua J Engelsma, Anil K Jain, and Vishnu Naresh Boddeti. Hers: Homomorphically

encrypted representation search. arXiv preprint arXiv:2003.12197, 2020.

[47] Joshua James Engelsma, Kai Cao, and Anil K Jain. Raspireader: Open source ﬁngerprint

reader. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018.

[48] Joshua James Engelsma, Kai Cao, and Anil K Jain. Learning a ﬁxed-length ﬁngerprint

representation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019.

[49] Joshua James Engelsma, Debayan Deb, Kai Cao, Anjoo Bhatnagar, Prem Sewak Sudhish,
and Anil K Jain. Infant-id: Fingerprints for global good. IEEE Transactions on Pattern
Analysis and Machine Intelligence, 2021.

[50] Jude Ezeobiejesi and Bir Bhanu. Latent Fingerprint Image Segmentation using Deep Neural

Network. In Deep Learning for Biometrics, pages 83–107. Springer, 2017.

[51] TJC Faes, HA Van der Meij, JC De Munck, and RM Heethaar. The electric resistivity of
human tissues (100 hz-10 mhz): a meta-analysis of review studies. Physiological Measure-
ment, 20(4):R1, 1999.

[52] Vincent Falanga and Brian Bucalo. Use of a durometer to assess skin hardness. Journal of

the American Academy of Dermatology, 29(1):47 – 51, 1993.

[53] Junfeng Fan and Frederik Vercauteren. Somewhat Practical Fully Homomorphic Encryp-

tion. IACR Cryptology ePrint Archive, 2012:144, 2012.

[54] Faisal Farooq, Ruud M Bolle, Tsai-Yang Jea, and Nalini Ratha. Anonymous and Revoca-
In 2007 IEEE Conference on Computer Vision and Pattern

ble Fingerprint Recognition.
Recognition, pages 1–7. IEEE, 2007.

[55] IAFIS FAQs. https://www.fbibiospecs.cjis.gov/Certiﬁcations/FAQ.

[56] Next Generation Identiﬁcation (NGI). https://www.fbi.gov/services/cjis/ﬁngerprints-and-

other-biometrics/ngi.

[57] NGI Fact Sheet. https://www.fbi.gov/ﬁle-repository/ngi-monthly-fact-sheet/view.

169

[58] Speedmixer. http://speedmixer.com/.

[59] F Galton. Finger prints of young children. British Association for the Advancement of

Science, 69:868–869, 1899.

[60] Francis Galton. Finger prints. Macmillan and Company, 1892.

[61] Gamry electro-chemical spectrometer. https://www.gamry.com/potentiostats/.

[62] Hand anthropometry. https://www.scribd.com/document/361450670/Hand-Anthropometry.

[63] Ed German. The History of Fingerprints. https://onin.com/fp/fphistory.html.

[64] Luca Ghiani, Paolo Denti, and Gian Luca Marcialis. Experimental results on ﬁngerprint
liveness detection. In Proceedings of the 7th International Conference on Articulated Mo-
tion and Deformable Objects, AMDO’12, pages 210–218. Springer-Verlag, 2012.

[65] Luca Ghiani, Abdenour Hadid, Gian Luca Marcialis, and Fabio Roli. Fingerprint liveness
In Biometrics: Theory, Applications
detection using binarized statistical image features.
and Systems (BTAS), 2013 IEEE Sixth International Conference on, pages 1–6. IEEE, 2013.

[66] Luca Ghiani, Gian Luca Marcialis, and Fabio Roli. Fingerprint liveness detection by local
phase quantization. In Pattern Recognition (ICPR), 2012 21st International Conference on,
pages 537–540. IEEE, 2012.

[67] Goodix live ﬁnger detection.

https://ﬁndbiometrics.com/goodix-zte-biometric-sensor-

3103187/.

[68] Government of India. AADHAR. Unique Identiﬁcation Authority of India (UIDAI). https:

//uidai.gov.in, 2019. [Online; accessed 14-February-2019].

[69] Diego Gragnaniello, Giovanni Poggi, Carlo Sansone, and Luisa Verdoliva. Fingerprint live-
In Biometric Measurements and
ness detection based on weber local image descriptor.
Systems for Security and Medical Applications (BIOMS), 2013 IEEE Workshop on, pages
46–50. IEEE, 2013.

[70] Diego Gragnaniello, Giovanni Poggi, Carlo Sansone, and Luisa Verdoliva. Local contrast
phase descriptor for ﬁngerprint liveness detection. Pattern Recognition, 48(4):1050–1058,
2015.

[71] Steven A Grosz, Tarang Chugh, and Anil K Jain. Fingerprint presentation attack detection:
A sensor and material agnostic approach. In 2020 IEEE International Joint Conference on
Biometrics (IJCB), pages 1–10. IEEE, 2020.

[72] Steven A Grosz, Joshua J Engelsma, Nicholas G Paulter, and Anil K Jain. White-box
In 2020 IEEE

evaluation of ﬁngerprint matchers: Robustness to minutiae perturbations.
International Joint Conference on Biometrics (IJCB), pages 1–10. IEEE, 2019.

[73] Lumidigm V- Series. https://www.hidglobal.com/products/biometrics/lumidigm/lumidigm-

v-series-ﬁngerprint-sensors.

170

[74] Lin Hong, Yifei Wan, and Anil Jain. Fingerprint image enhancement: Algorithm and per-
formance evaluation. IEEE Trans. Pattern Anal. Mach. Intell., 20(8):777–789, August 1998.

[75] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias
Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efﬁcient convolutional neural
networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017.

[76] IARPA ODIN program.

https://www.iarpa.gov/index.php/research-programs/odin/odin-

baa.

[77] Index Mundi. World Age Structure. https://www.indexmundi.com/world/age structure.

html, 2018. [Online; accessed 15-February-2019].

[78] Index Mundi. World Birth Rate. https://www.indexmundi.com/world/birth rate.html, 2018.

[Online; accessed 15-February-2019].

[79] Michael D Indovina, RA Hicklin, and G I Kiebuzinski. Nist evaluation of latent ﬁngerprint

technologies: Extended feature sets [evaluation# 1]. Technical report, NIST, 2011.

[80] Fingerprint matching software. https://www.innovatrics.com/idkit-ﬁngerprint-sdk/.

[81] International Standards Organization, ISO/IEC 30107-1:2016, Information Technology Bio-
metric Presentation Attack Detection Part 1: Framework. https://www.iso.org/standard/
53227.html,, 2016.

[82] Joshua J Engelsma, Debayan Deb, Anil Jain, Anjoo Bhatnagar, and Prem Sewak Sudhish.
Infant-prints: Fingerprints for reducing infant mortality. In Proceedings of the IEEE Con-
ference on Computer Vision and Pattern Recognition Workshops, pages 67–74, 2019.

[83] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial Transformer Networks.

In Advances in Neural Information Processing Systems, pages 2017–2025, 2015.

[84] Anil K Jain, Sunpreet S Arora, Kai Cao, Lacey Best-Rowden, and Anjoo Bhatnagar. Fin-
gerprint recognition of young children. IEEE Transactions on Information Forensics and
Security, 12(7):1501–1514, 2017.

[85] Anil K Jain, Kai Cao, and Sunpreet S Arora. Recognizing infants and toddlers using ﬁn-
gerprints: Increasing the vaccination coverage. In IEEE International Joint Conference on
Biometrics, pages 1–8. IEEE, 2014.

[86] Anil K Jain, Yi Chen, and Meltem Demirkus. Pores and ridges: High-resolution ﬁnger-
print matching using level 3 features. IEEE Transactions on Pattern Analysis and Machine
Intelligence, 29(1):15–27, 2006.

[87] Anil K Jain, Karthik Nandakumar, and Abhishek Nagar. Biometric Template Security.

EURASIP Journal on Advances in Signal Processing, 2008:113, 2008.

[88] Anil K Jain, Karthik Nandakumar, and Arun Ross. 50 years of biometric research: Accom-

plishments, challenges, and opportunities. Pattern recognition letters, 79:80–105, 2016.

171

[89] Anil K Jain, Salil Prabhakar, Lin Hong, and Sharath Pankanti. Fingercode: A Filterbank
for Fingerprint Representation and Matching. In Computer Vision and Pattern Recognition,
1999., volume 2, pages 187–193. IEEE, 1999.

[90] Anil K Jain, Salil Prabhakar, Lin Hong, and Sharath Pankanti. Filterbank-based Fingerprint

Matching. IEEE Transactions on Image Processing, 9(5):846–859, 2000.

[91] Anil K. Jain, Salil Prabhakar, and Arun Ross. Fingerprint matching: Data acquisition and
performance evaluation. Technical Report MSUTR99-14, Department of Computer Sci-
ence, Michigan State University, East Lansing, Michigan, March 1999.

[92] Anil K Jain, Arun A Ross, and Karthik Nandakumar. Introduction to biometrics. Springer

Science & Business Media, 2011.

[93] Herve Jegou, Matthijs Douze, and Cordelia Schmid. Product Quantization for Near-
IEEE Transactions on Pattern Analysis and Machine Intelligence,

est Neighbor Search.
33(1):117–128, 2010.

[94] Xiaofei Jia, Xin Yang, Yali Zang, Ning Zhang, and Jie Tian. A cross-device matching
ﬁngerprint database from multi-type sensors. In 21st International Conference on Pattern
Recognition (ICPR), 2012, pages 3001–3004. IEEE, 2012.

[95] Xudong Jiang, Manhua Liu, and Alex C Kot. Fingerprint Retrieval for Identiﬁcation. IEEE

Transactions on Information Forensics and Security, 1(4):532–542, 2006.

[96] Zhe Jin, Meng-Hui Lim, Andrew Beng Jin Teoh, Bok-Min Goi, and Yong Haur Tay. Gener-
ating Fixed-length Representation from Minutiae using Kernel Methods for Fingerprint Au-
thentication. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 46(10):1415–
1428, 2016.

[97] Joint Research Center of the European Commission. Fingerprint Recognition for Children.

https://bit.ly/2BKuoIj, 2013.

[98] Keyence.

Keyence Optical Digital Microscope VHX-600.

http://www1.keyence.eu/

products/microscope/microscope/vhx6002/vhx6002speciﬁcations1.php.

[99] Johannes Kotzerke, Stephen A Davis, Jodie McVernon, and Kathy J Horadam. Steps to solv-
ing the infant biometric problem with ridge-based biometrics. IET Biometrics, 7(6):567–
572, 2018.

[100] Philip Dean Lapsley, Jonathan Alexander Lee, David Ferrin Pare Jr, and Ned Hoffman.
Anti-fraud biometric scanner that accurately detects blood ﬂow, April 7 1998. US Patent
5,737,439.

[101] Rubisley P Lemes, Olga RP Bellon, Luciano Silva, and Anil K Jain. Biometric recognition
of newborns: Identiﬁcation using palmprints. In 2011 International Joint Conference on
Biometrics (IJCB), pages 1–6. IEEE, 2011.

172

[102] Ruilin Li, Dehua Song, Yuhang Liu, and Jufu Feng. Learning global ﬁngerprint features by
training a fully convolutional network with local patches. In 2019 International Conference
on Biometrics (ICB), pages 1–8. IEEE, 2019.

[103] Chenhao Lin and Ajay Kumar. Matching contactless and contact-based conventional ﬁn-
IEEE Transactions on Image Processing,

gerprint images for biometrics identiﬁcation.
27(4):2008–2021, 2018.

[104] Eryun Liu. Infant Footprint Recognition. In Proceedings of the IEEE International Confer-

ence on Computer Vision, pages 1653–1660, 2017.

[105] Eryun Liu, Heng Zhao, Jimin Liang, Liaojun Pang, Hongtao Chen, and Jie Tian. Random
Local Region Descriptor (RLRD): A New Method for Fixed-length Feature Representation
of Fingerprint Image and its Application to Template Protection. Future Generation Com-
puter Systems, 28(1):236–243, 2012.

[106] Manhua Liu and Pew-Thian Yap. Invariant Representation of Orientation Fields for Finger-

print Indexing. Pattern Recognition, 45(7):2532–2542, 2012.

[107] Dario Maio, Davide Maltoni, Raffaele Cappelli, Jim L Wayman, and Anil K Jain. FVC

Ongoing. https://biolab.csr.unibo.it/fvcongoing/UI/Form/Home.aspx.

[108] Dario Maio, Davide Maltoni, Raffaele Cappelli, Jim L Wayman, and Anil K Jain. FVC2002.

http://bias.csr.unibo.it/fvc2002/. 2002.

[109] Dario Maio, Davide Maltoni, Raffaele Cappelli, Jim L Wayman, and Anil K Jain. Fvc2004:
Third ﬁngerprint veriﬁcation competition. In Biometric Authentication, pages 1–7. Springer,
2004.

[110] Aakarsh Malhotra, Anush Sankaran, Mayank Vatsa, and Richa Singh. On matching ﬁnger-
IEEE Transactions on Biometrics, Behavior, and

selﬁes using deep scattering networks.
Identity Science, 2(4):350–362, 2020.

[111] Davide Maltoni, Dario Maio, Anil K. Jain, and Salil Prabhakar. Handbook of Fingerprint

Recognition. Springer, 2nd edition, 2009.

[112] E. Marasco and C. Sansone. On the robustness of ﬁngerprint liveness detection algorithms
against new materials used for spooﬁng. In Proc. Int. Conf. Bio-Inspired Syst. Signal Pro-
cess., pages 553–558, 2011.

[113] Emanuela Marasco and Arun Ross. A survey on antispooﬁng schemes for ﬁngerprint recog-

nition systems. ACM Comput. Surv., 47(2):28:1–28:36, November 2014.

[114] Emanuela Marasco and Carlo Sansone. Combining perspiration-and morphology-based
static features for ﬁngerprint liveness detection. Pattern Recognition Letters, 33(9):1148–
1156, 2012.

[115] Tsutomu Matsumoto, Hiroyuki Matsumoto, Koji Yamada, and Satoshi Hoshino.

Impact
of artiﬁcial gummy ﬁngers on ﬁngerprint systems. In Proceedings of SPIE, volume 4677,
pages 275–289, 2002.

173

[116] Medical Expo, THE ONLINE MEDICAL DEVICE EXHIBITION.

http://www.

medicalexpo.com/medical-manufacturer/test-phantom-2563.html. Accessed: 2017-2-07.

[117] David Menotti, Giovani Chiachia, Allan da Silva Pinto, William Robson Schwartz, H´elio
Pedrini, Alexandre Xavier Falc˜ao, and Anderson Rocha. Deep representations for iris,
face, and ﬁngerprint spooﬁng detection. IEEE Trans. Information Forensics and Security,
10(4):864–879, 2015.

[118] Vishesh Mistry, Joshua J Engelsma, and Anil K Jain. Fingerprint synthesis: Search with 100
million prints. In 2020 IEEE International Joint Conference on Biometrics (IJCB), pages
1–10. IEEE, 2019.

[119] Shimon K. Modi, Stephen J. Elliott, and Hale Kim. Statistical analysis of ﬁngerprint sensor
interoperability performance. In Proceedings of the 3rd IEEE International Conference on
Biometrics: Theory, Applications and Systems, pages 294–299, 2009.

[120] S. Moore. Latest tests of biometrics systems shows wide range of abilities. IEEE Spectrum
available at http://spectrum.ieee.org/computing/embedded-systems/latest-

Online, 2004.
tests-of-biometrics-systems-shows-wide-range-of-abilities.

[121] Valerio Mura, Luca Ghiani, Gian Luca Marcialis, Fabio Roli, David A. Yambay, and
Stephanie A. C. Schuckers. Livdet 2015 ﬁngerprint liveness detection competition 2015.
In BTAS, pages 1–6. IEEE, 2015.

[122] Abhishek Nagar, Shantanu Rane, and Anthony Vetro. Privacy and Security of Features
Extracted from Minutiae Aggregates. In 2010 IEEE International Conference on Acoustics,
Speech and Signal Processing, pages 1826–1829. IEEE, 2010.

[123] Karthik Nandakumar. A Fingerprint Cryptosystem Based on Minutiae Phase Spectrum.
In 2010 IEEE International Workshop on Information Forensics and Security, pages 1–6.
IEEE, 2010.

[124] Dinh-Luan Nguyen, Kai Cao, and Anil K Jain. Automatic Latent Fingerprint Segmentation.
In 2018 IEEE 9th International Conference on Biometrics Theory, Applications and Systems
(BTAS), pages 1–9. IEEE, 2018.

[125] Dinh-Luan Nguyen, Kai Cao, and Anil K Jain. Robust Minutiae Extractor: Integrating
Deep Networks and Fingerprint Domain Knowledge. In 2018 International Conference on
Biometrics (ICB), pages 9–16. IEEE, 2018.

[126] Van Huan Nguyen, Jinsong Liu, Thi Hai Binh Nguyen, Hakil Kim, et al. Universal ﬁnger-
print minutiae extractor using convolutional neural networks. IET Biometrics, 9(2):47–57,
2019.

[127] Shankar Bhausaheb Nikam and Suneeta Agarwal. Local binary pattern and wavelet based

spoof ﬁngerprint detection. Int. J. Biometrics, 1(2):141–159, August 2008.

[128] NIST special database 14. https://www.nist.gov/srd/nist-special-database-14.

174

[129] NIST special database 4. https://www.nist.gov/srd/nist-special-database-4.

[130] Kristin Nixon, Valerio Aimale, and Robert Rowe. Spoof detection schemes. Handbook of

Biometrics, pages 403–423, 2008.

[131] Rodrigo Nogueira, Roberto Lotufo, and Rubens Machado. Fingerprint liveness detection
IEEE Trans. Information Forensics and Security,

using convolutional neural networks.
11(6):1206–1213, 2016.

[132] Timo Ojala, Matti Pietik¨ainen, and Topi M¨aenp¨a¨a. Multiresolution gray-scale and rotation
invariant texture classiﬁcation with local binary patterns. IEEE Trans. Pattern Anal. Mach.
Intell., 24(7):971–987, July 2002.

[133] Michio Okajima. Development of dermal ridges in the fetus. Journal of Medical Genetics,

12(3):243–250, 1975.

[134] S. Orandi, F. Byers, S. Harvey, M.D. Garris, S.S. Wood, J.M. Libert, and J.C. Wu. Stan-
dard calibration target for contactless ﬁngerprint scanners, NIST patent, May 24 2016.
US9349033B2.

[135] Paciﬁc Northwest X-Ray Inc.

http://www.pnwx.com/Accessories/Phantoms/Radiology/

?Sr=Go&gclid=CPesworu59ECFcS6wAodoZ4MTQ. Accessed: 2017-2-07.

[136] Sharath Pankanti, Salil Prabhakar, and Anil K Jain. On the individuality of ﬁngerprints.
IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(8):1010–1025, 2002.

[137] Perkin elmer lambda 900 uv-vis spectrometer. http://www.perkinelmer.com/category/uv-

vis-spectroscopy-uv.

[138] Javier Preciozzi, Guillermo Garella, Vanina Camacho, Francesco Franzoni, Luis Di Mar-
tino, Guillermo Carbajal, and Alicia Fernandez. Fingerprint biometrics from newborn to
adult: A study from a national identity database system. IEEE Transactions on Biometrics,
Behavior, and Identity Science, 2(1):68–79, 2020.

[139] Precise Biometrics.

https://precisebiometrics.com/products/ﬁngerprint-spoof-liveness-

detection/sensor-vulnerability-analysis/.

[140] Zhenshen Qu, Junyu Liu, Yang Liu, Qiuyu Guan, Ruikun Li, and Yuxin Zhang. A Novel
System for Fingerprint Orientation Estimation. In Chinese Conference on Image and Graph-
ics Technologies, pages 281–291. Springer, 2018.

[141] F Rahmun and O Bausinger. Best practice ﬁngerprint enrolment standards european visa

information system. https://bit.ly/2VcyyQN, 2010.

[142] Ajita Rattani and Arun Ross. Automatic adaptation of ﬁngerprint liveness detector to new
spoof materials. In 2014 IEEE International Joint Conference on Biometrics (IJCB), pages
1–8. IEEE, 2014.

175

[143] Ajita Rattani, Walter J Scheirer, and Arun Ross. Open set ﬁngerprint spoof detection
IEEE Trans. on Information Forensics and Security,

across novel fabrication materials.
10(11):2447–2460, 2015.

[144] Arun Ross and Anil Jain. Biometric Sensor Interoperability: A Case Study in Fingerprints,

pages 134–145. Springer, 2004.

[145] Arun Ross and Rohan Nadgir. A thin-plate spline calibration model for ﬁngerprint sensor

interoperability. IEEE Trans. on Knowl. and Data Eng., 20(8):1097–1110, August 2008.

[146] Robert K Rowe. Multispectral imaging biometrics, December 2 2008. US Patent 7,460,696.

[147] Steven Saggese, Yunting Zhao, Tom Kalisky, Courtney Avery, Deborah Forster, Lilia Edith
Duarte-Vera, Lucila Alejandra Almada-Salazar, Daniel Perales-Gonzalez, Alexandra
Hubenko, Michael Kleeman, et al. Biometric recognition of newborns and infants by non-
contact ﬁngerprinting: lessons learned. Gates Open Research, 3, 2019.

[148] Anush Sankaran, Aakarsh Malhotra, Apoorva Mittal, Mayank Vatsa, and Richa Singh. On
smartphone camera based ﬁngerphoto authentication. In 2015 IEEE 7th International Con-
ference on Biometrics Theory, Applications and Systems (BTAS), pages 1–7. IEEE, 2015.

[149] JK Schneider. Quantifying the dermatoglyphic growth patterns in children through adoles-

cence. https://www.ncjrs.gov/pdfﬁles1/nij/grants/232746.pdf, 2010.

[150] Patrick Schuch, Simon-Daniel Schulz, and Christoph Busch. Deep Expectation for Estima-
tion of Fingerprint Orientation Fields. In Biometrics (IJCB), 2017 IEEE International Joint
Conference on, pages 185–190. IEEE, 2017.

[151] Stephanie AC Schuckers. Spooﬁng and anti-spooﬁng measures. Information Security tech-

nical report, 7(4):56–62, 2002.

[152] Mojtaba Sepasian, Cristinel Mares, and Wamadeva Balachandran. Liveness and spooﬁng
in ﬁngerprint identiﬁcation: Issues and challenges. In Proceedings of the 4th WSEAS Inter-
national Conference on Computer Engineering and Applications, CEA’10, pages 150–158,
2009.

[153] Akihide Shiratsuki, Emiko Sano, Masahiro Shikai, Toshiro Nakashima, T Takashima,
Masato Ohmi, and Masamitsu Haruna. Novel optical ﬁngerprint sensor utilizing optical
characteristics of skin tissue under ﬁngerprints. In Photonic Therapeutics and Diagnostics,
volume 5686, pages 80–88. International Society for Optics and Photonics, 2005.

[154] Silicone Solutions. SS-27S Technical Data Sheet. available at http://siliconesolutions.com/

media/pdf/SS-27STDS.pdf.

[155] SilkID.

Silk20-Reader.

http://www.silkid.com/wp-content/uploads/2017/02/Silk20-

Reader-Brochure-v0.7.pdf, 2019. [Online; accessed 4-May-2019].

[156] Smooth-On. Silicone Ease Release 200 Technical Data Sheet. available at https://www.

smooth-on.com/tb/ﬁles/er200.pdf.

176

[157] Smooth-On. Silicone Silc Pig Technical Data Sheet. available at https://www.smooth-

on.com/tb/ﬁles/Silc Pig Pigments.pdf.

[158] Smooth-On. Silicone Thinner Technical Data Sheet. available at https://www.smooth-on.

com/tb/ﬁles/Silicone Thinner.pdf.

[159] Dehua Song and Jufu Feng. Fingerprint Indexing based on Pyramid Deep Convolutional
Feature. In IEEE International Joint Conference on Biometrics (IJCB), 2017, pages 200–
207. IEEE, 2017.

[160] Dehua Song, Yao Tang, and Jufu Feng. Aggregating Minutia-Centred Deep Convolutional

Features for Fingerprint Indexing. Pattern Recognition, 88:397–408, 2019.

[161] Stratsys. Digital Material Technical DataSheet.

available at http://global72.stratasys.
com/∼/media/Main/Files/Material Spec Sheets/MSS PJ PJMaterialsDataSheet.pdf# ga=2.
193296616.455247396.1494258560-1158297905.1493673102.

[162] Stratsys. Stratasys Connex Printer Machine Speciﬁcations. http://www.stratasys.com/∼/

media/Main/Files/Machine Spec Sheets/PSS PJ Connex3.ashx.

[163] Yijing Su, Jianjiang Feng, and Jie Zhou. Fingerprint Indexing with Pose Constraint. Pattern

Recognition, 54:1–13, 2016.

[164] Yagiz Sutcu, Shantanu Rane, Jonathan S Yedidia, Stark C Draper, and Anthony Vetro. Fea-
ture Extraction for a Slepian-Wolf Biometric System using LDPC Codes. In 2008 IEEE
International Symposium on Information Theory, pages 2297–2301. IEEE, 2008.

[165] Yagiz Sutcu, Husrev T Sencar, and Nasir Memon. A Geometric Transformation to Protect
Minutiae-based Fingerprint Templates. In Biometric Technology for Human Identiﬁcation
IV, volume 6539, page 65390E. International Society for Optics and Photonics, 2007.

[166] Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander A Alemi. Inception-v4,
Inception-resnet and the Impact of Residual Connections on Learning. In AAAI, volume 4,
page 12, 2017.

[167] Bozhao Tan, Aaron Lewicke, David Yambay, and Stephanie Schuckers. The effect of envi-
ronmental conditions and novel spooﬁng methods on ﬁngerprint anti-spooﬁng algorithms.
IEEE Information Forensics and Security, WIFS, 2010.

[168] Yao Tang, Fei Gao, Jufu Feng, and Yuhang Liu. Fingernet: An Uniﬁed Deep Network
for Fingerprint Minutiae Extraction. In IEEE International Joint Conference on Biometrics
(IJCB), 2017, pages 108–116. IEEE, 2017.

[169] ThorLabs. https://www.thorlabs.com/thorproduct.cfm?partnumber=PS911.

[170] Mitchell Trauring. Automatic comparison of ﬁnger-ridge patterns. Nature, 197(4871):938–

940, 1963.

[171] Unique Identiﬁcation Authority of India, dashboard summary. https://portal.uidai.gov.in/

uidwebportal/dashboard.do.

177

[172] Umut Uludag, Sharath Pankanti, and Anil K Jain. Fuzzy Vault for Fingerprints.

In In-
ternational Conference on Audio-and Video-Based Biometric Person Authentication, pages
310–319. Springer, 2005.

[173] Unicef. Faces, Fingerprints, and Feet. https://data.unicef.org/resources/biometrics/. [On-

line; accessed 4-October-2020].

[174] Maneesh Upmanyu, Anoop M Namboodiri, Kannan Srinathan, and CV Jawahar. Efﬁcient
Biometric Veriﬁcation in Encrypted Domain. In International Conference on Biometrics,
pages 899–908. Springer, 2009.

[175] Maneesh Upmanyu, Anoop M Namboodiri, Kannan Srinathan, and CV Jawahar. Blind
Authentication: A Secure Crypto-Biometric Veriﬁcation Protocol. IEEE Transactions on
Information Forensics and Security, 5(2):255–268, 2010.

[176] Ton Van der Putte and Jeroen Keuning. Biometrical ﬁngerprint recognition: dont get
your ﬁngers burned. In Smart Card Research and Advanced Applications, pages 289–303.
Springer, 2000.

[177] Dayong Wang, Charles Otto, and Anil K Jain. Face Search at Scale. IEEE Transactions on

Pattern Analysis and Machine Intelligence, 39(6):1122–1136, 2016.

[178] C I. Watson, G.P. Fiumara, E. Tabassi, S.L Cheng, P.A. Flanagan, and W.J. Salamon. Fin-
gerprint vendor technology evaluation, nist interagency/internal report 8034: 2015. NIST.IR,
2014. available at https://dx.doi.org/10.6028/NIST.IR.8034.

[179] Yandong Wen, Kaipeng Zhang, Zhifeng Li, and Yu Qiao. A Discriminative Feature Learning
Approach for Deep Face Recognition. In European Conference on Computer Vision, pages
499–515. Springer, 2016.

[180] World Food Programme. WFP demands action after uncovering misuse of food relief
intended for hungry people in Yemen. https://www1.wfp.org/news/wfp-demands-action-
after-uncovering-misuse-food-relief-intended-hungry-people-yemen, 2018.
[Online; ac-
cessed 15-February-2019].

[181] World Food Programme Insight. These changes show that WFP loves us. https://insight.
wfp.org/these-changes-show-that-wfp-loves-us-247f0c1ebcf, 2018. [Online; accessed 15-
February-2019].

[182] World Health Organization. Prevention not cure in tackling health-care fraud. https://
www.who.int/bulletin/volumes/89/12/11-021211/en/, 2011. [Online; accessed 15-February-
2019].

[183] World Health Organization. Progress and Challenges with Achieving Universal Immuniza-
tion Coverage. https://www.who.int/immunization/monitoring surveillance/who-immuniz.
pdf, 2018. [Online; accessed 15-February-2019].

178

[184] Haiyun Xu, Raymond NJ Veldhuis, Asker M Bazen, Tom AM Kevenaar, Ton AHM Akker-
mans, and Berk Gokberk. Fingerprint Veriﬁcation using Spectral Minutiae Representations.
IEEE Transactions on Information Forensics and Security, 4(3):397–409, 2009.

[185] David Yambay, Luca Ghiani, Paolo Denti, Gian Luca Marcialis, Fabio Roli, and Stephanie
A. C. Schuckers. Livdet 2011 - ﬁngerprint liveness detection competition 2011. In Proc.
International Conf. Biometrics, pages 208–215. IEEE, 2012.

[186] Xi Yin and Xiaoming Liu. Multi-task Convolutional Neural Network for Pose-Invariant

Face Recognition. IEEE Transactions on Image Processing, 27(2):964–975, 2018.

[187] Soweon Yoon, Jianjiang Feng, and Anil K Jain. Altered ﬁngerprints: Analysis and detection.

IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(3):451–464, 2012.

[188] Soweon Yoon and Anil K Jain. Longitudinal Study of Fingerprint Recognition. Proceedings

of the National Academy of Sciences, 112(28):8555–8560, 2015.

[189] Matthew D Zeiler and Rob Fergus. Visualizing and Understanding Convolutional Networks.

In European Conference on Computer Vision, pages 818–833. Springer, 2014.

[190] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network
for image super-resolution. In Proceedings of the IEEE conference on computer vision and
pattern recognition, pages 2472–2481, 2018.

[191] Baicun Zhou, Congying Han, Yonghong Liu, Tiande Guo, and Jin Qin. Fast minutiae ex-

tractor using neural network. Pattern Recognition, 103:107273, 2020.

[192] Wei Zhou, Jiankun Hu, Ian Petersen, Song Wang, and Mohammed Bennamoun. Perfor-
mance evaluation of 2d to 3d ﬁngerprint recognition. In 2013 6th International Congress
on Image and Signal Processing (CISP), volume 3, pages 1736–1741. IEEE, 2013.

[193] Wei Zhou, Jiankun Hu, Song Wang, Ian Petersen, and Mohammed Bennamoun. Perfor-
mance evaluation of large 3d ﬁngerprint databases. Electronics letters, 50(15):1060–1061,
2014.

[194] Yanming Zhu, Xuefei Yin, Xiuping Jia, and Jiankun Hu. Latent Fingerprint Segmentation
based on Convolutional Neural Networks. In IEEE Workshop on Information Forensics and
Security (WIFS), 2017, pages 1–6. IEEE, 2017.

[195] Yongfang Zhu, Sarat C Dass, and Anil K Jain. Statistical models for assessing the individ-
uality of ﬁngerprints. IEEE Transactions on Information Forensics and Security, 2(3):391–
401, 2007.

179