CONTRIBUTIONS TO BIOMETRIC RECOGNITION:
MATCHING IDENTICAL TWINS AND LATENT FINGERPRINTS
By
Alessandra Aparecida Paulino

A DISSERTATION
Submitted
to Michigan State University
in partial fulﬁllment of the requirements
for the degree of
Computer Science– Doctor of Philosophy
2013

ABSTRACT
CONTRIBUTIONS TO BIOMETRIC RECOGNITION:
MATCHING IDENTICAL TWINS AND LATENT FINGERPRINTS
By
Alessandra Aparecida Paulino
Automatic recognition of a person by the use of distinctive physical and behavioral characteristics is called biometrics. Two of the important challenges in biometrics are recognition of
identical twins and matching of latent ﬁngerprints to their exemplar prints (rolled or slap prints).
The contributions of this dissertation are focused on these two topics.
Identical (monozygotic) twins are a result of a single fertilized egg that splits into two cells,
each one giving birth to one individual. Identical twins have the same deoxyribonucleic acid
(DNA), thus their genotypic features (features inﬂuenced by the genetic material) are the same.
However, some of their phenotypic features (features inﬂuenced by the fetal environment) may
be different. Thus, it is essential to determine which biometric traits (either by themselves or in
combination with other traits) have the ability to distinguish identical twins and the extent of their
ability for this discrimination.
The ﬁrst contribution of this dissertation is an evaluation of the performance of biometric systems in the presence of identical twins for the three most commonly used biometric modalities,
namely ﬁngerprint, face and iris. Identical twins are shown to be a challenge to current face
recognition systems. On the other hand, ﬁngerprint and iris matching of identical twins show performance comparable to those with unrelated persons. The fusion of different samples from the
same modality of a subject (e.g., left and right iris, ﬁngerprints of multiple ﬁngers) yields the best
matching performance for identical twins, similar to what has been shown for unrelated persons.
Biometric traits can also be used to determine whether two subjects enrolled in a biometric database
are identical twins. By using face and iris modalities together, for example, we can correctly iden-

tify 80% of such identical twin pairs, while only 2% of subject pairs who are not identical twins
will be incorrectly considered identical twins.
The second contribution of this work is focused on improving latent ﬁngerprint matching performance. Latent ﬁngerprints are partial ﬁngerprint images that typically contain only a small
area of friction ridge pattern and large non-linear distortion, are blurred or smudged, and contain
complex background noise. Due to these characteristics, latents are a particularly challenging for
matching to their mates (reference prints) in a database. Given a latent print in which minutiae have
been marked by a human expert (as is the current practice in forensics), we have proposed two approaches to improve the latent matching performance. The ﬁrst approach consists of enhancing
the latent image and fusing the matching score obtained from the enhanced latents with the score
based on manually marked minutiae. This approach outperforms a commercial ﬁngerprint matcher
on the public latent database NIST SD 27. The second approach consists of developing a latent
ﬁngerprint matcher that utilizes minutiae as well as the orientation ﬁeld information. The proposed
matching algorithm outperforms three ﬁngerprint matching algorithms on two different latent ﬁngerprint databases (NIST SD 27 and WVU latent databases).
The latent ﬁngerprint identiﬁcation accuracy generally deteriorates as the size of the ﬁngerprint
database grows. To alleviate this problem and to reduce the overall search time of latent matching, we propose to combine various level 1 and level 2 features, including minutia cylinder code,
minutia triplets, singular points, orientation ﬁeld and ridge period, to efﬁciently ﬁlter out a large
portion of the reference ﬁngerprint database. The proposed approach outperforms state-of-the-art
indexing techniques on the public domain latent database NIST SD27 against a large background
database of 267K rolled prints. The experimental results also show that the proposed ﬁltering
scheme has the desirable property of attaining improved computational efﬁciency of latent search
(20% penetration rate) while maintaining the latent matching accuracy.

To my loved ones.

iv

ACKNOWLEDGEMENTS

I consider myself a very fortunate person for having so many people to be thankful to. Some were
part of the beginning of my Ph.D. “journey”, some were there even before, some became part of it
recently, and some were there from the beginning to the point we are now. Each and all of them
helped me in one way or another and I am grateful to all of them for being there for me. There
were a few difﬁcult and frustrating moments for sure, but there was never a moment that I regretted
coming here and pursuing my Ph.D.. I am very thankful for all the wonderful people I have met
and for everything I have learned.
I should start by thanking Dr. Anil K. Jain, who has been a great advisor and mentor over the
ﬁve years I have worked under his supervision. I admire him immensely as a researcher, a teacher,
an advisor, and as a person as well. We can always go to him with a problem or a question, because
he is always willing to help. He says maybe he worries about us more than ourselves; this is just
the type of advisor he is, caring and always looking out for us and helping us achieve our best. He
always goes the extra mile to help us in any way he can; this is just the type of person he is. I have
learned so much from him that no thank you will ever be enough.
I am also grateful to have such a great thesis committee. Thank you Dr. Selin Aviyente, Dr.
Xiaoming Liu and Dr. Arun Ross for evaluating my work and providing valuable comments and
suggestions. Thank you also to Dr. Pang-Ning Tan and Dr. George Stockman for being part of the
committee in the early stages.
An important part of being able to write this dissertation is due to all members of the ﬁngerprint
group at PRIP lab. Our weekly meetings, their valuable suggestions and discussions helped me not
only with my work directly, but also with becoming a better researcher in general.
I would like to thank Jianjiang Feng, with whom we had a very active work collaboration and
from whom I also learned a lot. Thank you Heeseung, Eryun and Kai for revising this dissertation.
I am grateful for the administrative support received from the CSE ofﬁce at MSU, including Linda

v

Moore, Cathy Davison, Norma Teague and Debbie Kruch, and for the computer support received
from Kelly Climer and Adam Pitcher. Thank you so much for your patience and support.
I cannot express how much I have learned from fellow Pripies, both in academic and personal
levels. We are friends in the journey and I value all the time we spent together, all the lessons I
learned from you, and all the help you provided. Thank you Soweon, Jung-Eun, Brendan, Serhat,
Kien, Qijun, Eryun, Kai, Abhishek, Pavan, Radha, Sunpreet, Unsang, Hu, Lacey, Scott, Charles. A
special thank you to Soweon for all the discussions and help; she has been someone I could always
count on, from helping me ﬁnd a small bug in my code to discussing life. She has become a great
friend and a person who I admire and respect personally and professionally.
I would like to thank my husband for being so understanding and patient over the course of my
Ph.D.. His unconditional support has helped me go through the tough times, and his help has kept
me balanced. Having him by my side gave me strength and it is wonderful to be able to share my
life with him. I can always count on him: from cooking by himself so that I can have more time to
work, to listening and talking to me so that I can ﬁgure things out and move forward.
I also would like to thank my parents for being so supportive of everything, every decision,
every step, every journey. They have always been there for me and I am so grateful. The peace
of mind that comes from knowing that I can count on them if I need to has made my life so much
easier and smoother. I grew up in a wonderful family, and it was not easy to be far from them, but
they were always very supportive and understanding. I am grateful to my sisters who, despite the
distance, were always there to share things and laugh together. And I have missed all of them so
much, especially my nephew, who was only four years old when I came to the U.S. and now is
almost as tall as his mother. I am also grateful to my extended family (in-laws), who has been very
supportive throughout this journey, and also very understanding of my husband and I being so far
from them.
And last but not least, I would like to thank all my friends for everything. I am very fortunate
to have friends to count on, laugh with and learn together. It is always wonderful to be with them.
Thank you to everyone who was a part of this journey. Your support made this possible.

vi

TABLE OF CONTENTS

LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

x

LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

xi

CHAPTER 1 INTRODUCTION . . . . . . . . . . . .
1.1 Biometric Systems . . . . . . . . . . . . . . . .
1.2 Common Biometric Traits . . . . . . . . . . . . .
1.2.1 Fingerprint Representation and Matching
1.2.1.1 Representation . . . . . . . . .
1.2.1.2 Matching . . . . . . . . . . . .
1.2.1.3 Indexing . . . . . . . . . . . .
1.2.1.4 Latent Fingerprints . . . . . . .
1.2.2 Face Representation and Matching . . . .
1.2.2.1 Representation . . . . . . . . .
1.2.2.2 Matching . . . . . . . . . . . .
1.2.3 Iris Representation and Matching . . . . .
1.2.3.1 Representation . . . . . . . . .
1.2.3.2 Matching . . . . . . . . . . . .
1.3 Applications of Biometric Systems . . . . . . . .
1.4 Contributions . . . . . . . . . . . . . . . . . . .
1.4.1 Distinguishing Identical Twins . . . . . .
1.4.2 Latent Fingerprint Matching . . . . . . .
1.4.3 Latent Fingerprint Indexing . . . . . . . .
1.5 Dissertation Organization . . . . . . . . . . . . .

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

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

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

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

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

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

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

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

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

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

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

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

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

1
5
8
10
10
16
18
19
22
22
22
25
26
27
28
29
30
31
32
33

CHAPTER 2 BIOMETRIC TRAITS OF IDENTICAL TWINS
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 Multibiometrics . . . . . . . . . . . . . . . . . . . . . . . .
2.3.1 Fingerprint Discriminability . . . . . . . . . . . . .
2.3.2 Face Discriminability . . . . . . . . . . . . . . . . .
2.3.3 Iris Discriminability . . . . . . . . . . . . . . . . . .
2.4 Experimental Results and Analysis . . . . . . . . . . . . . .
2.4.1 Databases . . . . . . . . . . . . . . . . . . . . . . .
2.4.2 Distinguishing Identical Twins . . . . . . . . . . . .
2.4.2.1 Fingerprint . . . . . . . . . . . . . . . . .
2.4.2.2 Face . . . . . . . . . . . . . . . . . . . . .
2.4.2.3 Iris . . . . . . . . . . . . . . . . . . . . .
2.4.2.4 Multibiometric Experimental Results . . .
2.4.2.5 Discussion . . . . . . . . . . . . . . . . .
2.4.3 Finding Similarities between Identical Twins . . . .

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

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

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

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

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

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

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

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

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

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

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

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

34
34
37
43
44
44
46
47
47
49
49
51
53
54
57
60

vii

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

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

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

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

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

2.5

Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

CHAPTER 3 LATENT FINGERPRINT MATCHING . . . . . . . . .
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.1 Rolled Fingerprint Matching . . . . . . . . . . . . . . . .
3.2.2 Latent Fingerprint Matching . . . . . . . . . . . . . . . .
3.2.3 NIST Evaluation of Latent Fingerprint Technologies . . .
3.2.4 Evaluation of Latent Examiners . . . . . . . . . . . . . .
3.3 Latent Fingerprint Features . . . . . . . . . . . . . . . . . . . . .
3.3.1 Orientation Field Estimation . . . . . . . . . . . . . . . .
3.3.1.1 Minutiae based Orientation Field Reconstruction
3.3.2 Local Minutia Descriptor . . . . . . . . . . . . . . . . . .
3.3.2.1 Minutia Cylinder Code (MCC) . . . . . . . . .
3.4 Latent Matching Approaches . . . . . . . . . . . . . . . . . . . .
3.4.1 Latent Matching with Enhanced Image . . . . . . . . . . .
3.4.1.1 Rank-level Fusion . . . . . . . . . . . . . . . .
3.4.1.2 Score-level Fusion . . . . . . . . . . . . . . . .
3.4.2 Latent Matching with Descriptor-Based Hough Transform
3.4.2.1 Alignment . . . . . . . . . . . . . . . . . . . .
3.4.2.2 Similarity Measure . . . . . . . . . . . . . . . .
3.5 Databases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5.1 NIST Special Database 27 (NIST SD27) . . . . . . . . . .
3.5.2 West Virginia University Latent Database (WVU LFD) . .
3.6 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . .
3.6.1 Latent Matching with Enhanced Images . . . . . . . . . .
3.6.1.1 Fusion methods . . . . . . . . . . . . . . . . . .
3.6.2 Latent Matching with Descriptor-Based Hough Transform
3.6.2.1 Commercial Matchers . . . . . . . . . . . . . .
3.6.2.2 Alignment Performance . . . . . . . . . . . . .
3.6.2.3 Matching Performance . . . . . . . . . . . . . .
3.6.2.4 Fusion of Matchers . . . . . . . . . . . . . . . .
3.6.2.5 Effect of Fingerprint Quality . . . . . . . . . . .
3.6.2.6 Computational Cost . . . . . . . . . . . . . . .
3.6.2.7 Comparison of the two proposed methods . . . .
3.7 Summary and Conclusions . . . . . . . . . . . . . . . . . . . . .

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

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

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

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

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

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

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

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

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

68
68
74
74
75
76
77
78
78
80
84
84
86
86
88
89
90
91
95
101
101
102
104
104
105
110
110
110
114
119
123
125
125
131

CHAPTER 4 LATENT FINGERPRINT INDEXING
4.1 Introduction . . . . . . . . . . . . . . . . . . . .
4.2 Related Work . . . . . . . . . . . . . . . . . . .
4.3 Feature Extraction . . . . . . . . . . . . . . . . .
4.3.1 Reference Prints . . . . . . . . . . . . . .
4.3.2 Latents . . . . . . . . . . . . . . . . . .
4.4 Indexing Approach . . . . . . . . . . . . . . . .
4.4.1 Triplets . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

132
132
135
140
140
142
143
146

viii

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

4.5

4.6

4.4.2 Minutia Cylinder Code (MCC) Indexing
4.4.3 Orientation Field Descriptor Indexing .
4.4.4 Singular Points . . . . . . . . . . . . .
4.4.5 Ridge Period . . . . . . . . . . . . . .
4.4.6 Fusion of Indexing Scores . . . . . . .
Experimental Results . . . . . . . . . . . . . .
4.5.1 Databases . . . . . . . . . . . . . . . .
4.5.2 Indexing Results . . . . . . . . . . . .
4.5.3 Implementation Details . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . . .

CHAPTER 5 SUMMARY AND FUTURE WORK
5.1 Summary . . . . . . . . . . . . . . . . . . .
5.2 Contributions . . . . . . . . . . . . . . . . .
5.3 Future Work . . . . . . . . . . . . . . . . . .

.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

148
149
150
151
152
153
153
154
162
163

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

165
165
167
168

BIBLIOGRAPHY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171

ix

LIST OF TABLES

Table 1.1 Comparison of the characteristics of the three most commonly used biometric
traits from [1]. (H = high, M = medium, L = low) . . . . . . . . . . . . . . . . .

8

Table 1.2 False reject and false acceptance rates associated with ﬁngerprint, face and iris
veriﬁcation systems. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

9

Table 2.1 Summary of studies on the biometrics of identical twins. Sets can include
identical twin pairs as well as non-identical twin pairs. . . . . . . . . . . . . . . 38
Table 2.2 Equal Error Rate (%) for distinguishing (i) genuine vs. impostor identical
twins and (ii) genuine vs. general impostors based on different biometric features. 57
Table 3.1 Comparison of rank-1 accuracies reported in the literature for the NIST SD27
database. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
Table 3.2 Rank-1 accuracies for various subjective quality levels of latents in NIST SD27. . 123
Table 3.3 Rank-1 accuracies for various objective quality values of latents in NIST SD27
(large, medium and small refer to the number of minutiae in the latent). . . . . . 123
Table 3.4 Rank-1 accuracies for various objective quality values of latents in WVU LFD
(large, medium and small refer to the number of minutiae in the latent). . . . . . 124
Table 3.5 Rank-1 accuracies for latents grouped according to NFIQ quality values of
corresponding rolled prints in NIST SD27. . . . . . . . . . . . . . . . . . . . . 124
Table 3.6 Rank-1 accuracies for latents grouped according to NFIQ quality values of
corresponding rolled prints in WVU LFD. . . . . . . . . . . . . . . . . . . . . . 124
Table 4.1 Summary of studies on ﬁngerprint indexing for rolled, plain and latent prints. . . 137
Table 4.2 Computation time for individual indexing modules. . . . . . . . . . . . . . . . . 163

x

LIST OF FIGURES

Figure 1.1 Illustration of Bertillon system for person identiﬁcation based on numerous
body measurements, which was used in the United States and in Europe from
the end of the 19th century to the beginning of the 20th century. From left to
right and then top to bottom the ﬁgures show measurement of height, reach,
trunk, length of head, width of head, right ear, left foot, left middle ﬁnger,
and left forearm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2

Figure 1.2 Examples of intra-class variability and inter-class similarity. (a) Two different impressions of the same ﬁnger of the same subject and (b) face images of
a pair of identical twins. For interpretation of the references to color in this
and all other ﬁgures, the reader is referred to the electronic version of this
dissertation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

5

Figure 1.3 An overview of a biometric system. . . . . . . . . . . . . . . . . . . . . . . . .

6

Figure 1.4 Three types of ﬁngerprint images. . . . . . . . . . . . . . . . . . . . . . . . . . 11
Figure 1.5 Examples of ﬁngerprint features belonging to the three levels: (a) orientation
ﬁeld and singular points and (b) minutiae shown in ﬁngerprints from NIST
Special Database 4 [2], and (c) sweat pores shown on part of a rolled print
image from the WVU Latent Database [3]. . . . . . . . . . . . . . . . . . . . . 11
Figure 1.6 Examples of Level 1 features in ﬁngerprints. (a) Orientation ﬁeld and (b)
singular points (core in red and delta in green) shown in rolled ﬁngerprint
images from NIST Special Database 27 [4]. . . . . . . . . . . . . . . . . . . . 12
Figure 1.7 Example ﬁngerprints belonging to each of the ﬁve major ﬁngerprint classes
(ﬁngerprints from NIST Special Database 4 [2].) . . . . . . . . . . . . . . . . . 14
Figure 1.8 An illustration of how minutia direction is deﬁned for (a) minutia ending and
(b) minutia bifurcation [5]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
Figure 1.9 Example of matched minutiae from a ﬁngerprint pair (latent and rolled print). . 17
Figure 1.10 Examples of latent ﬁngerprints in NIST SD 27 database [4]. . . . . . . . . . . . 20
Figure 1.11 Examples of facial features at the three levels: (a) facial geometry (Level
1), (b) facial landmarks, Gabor ﬁlter, local binary pattern (LBP), and shape
(Level 2) and (c) moles (Level 3) [5]. . . . . . . . . . . . . . . . . . . . . . . . 23
Figure 1.12 An illustration of a graph model ﬁtted to a face [5]. . . . . . . . . . . . . . . . 24

xi

Figure 1.13 Schematic diagram of (a) SIFT [6] and (b) LBP [5] descriptor construction. . . 25
Figure 1.14 Examples of two irises of two different subjects from the CASIA twins database.
The image on the right (b) is not of very high quality since the iris is partially
occluded by hair. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
Figure 1.15 An overview of an iris recognition system [5]. . . . . . . . . . . . . . . . . . . 27
Figure 1.16 A pair of identical twins from CASIA Twins database Version 2. . . . . . . . . 31
Figure 2.1 Face images of (a) a pair of identical (monozygotic) twins and (b) a pair
of non-identical (dizygotic) twins from the University of Notre Dame NDTwins database [7]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
Figure 2.2 A diagram of the human eye 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
Figure 2.3 The kiosk for biometric acquisition (a) and the face acquisition device (b). . . . 48
Figure 2.4 Fingerprint images of ﬁngers 1, 2, 3, and 4 of the ﬁrst twin (a), and the four
images of the corresponding ﬁngers of the second twin in an identical twin
pair (b); similarly, (c) and (d) show ﬁngerprint images of a non-identical twin
pair. Note the similarity in ridge ﬂow pattern between identical twins. All
four corresponding ﬁngers of identical twins in (a) and (b) have the same pattern type. But for non-identical twins in (c) and (d), only two corresponding
ﬁngers (no. 1 and 3) have the same pattern type. . . . . . . . . . . . . . . . . . 50
Figure 2.5 Face images of the ﬁrst and second twin in (a) an identical twin pair, and (b)
a non-identical twin pair. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
Figure 2.6 The left and right iris images of identical ((a) and (b)) and non-identical twin
pairs ((c) and (d)). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
Figure 2.7 Identical twin, general impostor distributions and genuine distributions for
ﬁngers 3 and 4, face, and right iris. . . . . . . . . . . . . . . . . . . . . . . . . 55
Figure 2.8 Identical twin, general impostor, and genuine distributions of the 4-ﬁnger
fusion and 2-iris fusion, respectively. . . . . . . . . . . . . . . . . . . . . . . . 56
Figure 2.9 ROC curves for ﬁngerprint, iris and face, and a fusion example (face + ﬁnger
4). Due to the small number of identical twin impostor (102), FAR less than
1/102 cannot be estimated. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
Figure 2.10 Examples of face images in CASIA Twins V2 database. . . . . . . . . . . . . . 60
Figure 2.11 ROC curves for face using FaceVacs 8.6. . . . . . . . . . . . . . . . . . . . . . 61
Figure 2.12 ROC curves for twin detection using iris and face. . . . . . . . . . . . . . . . . 63

xii

Figure 2.13 ROC curves for twin detection using ﬁngerprint and face. . . . . . . . . . . . . 64
Figure 2.14 Face images of pairs of (a) identical and (b) non-identical twins with extremely high face match scores considering the range of scores is [0, 1] using
FaceVacs 8.6.) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
Figure 3.1 Three types of ﬁngerprint impressions. Rolled and plain ﬁngerprints are also
called reference ﬁngerprints. . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
Figure 3.2 An example of a tenprint card containing two types of impressions (rolled
and plain) of the ten ﬁngers. The ﬁrst portion of the card contains personal
information related to the person being ﬁngerprinted, such as name, date of
birth, place of birth, gender, race, height, weight, etc. The second portion
of the card contains impressions of the ten ﬁngers. The ﬁrst and second
rows of impressions consist of rolled impressions of the ﬁngers from the
left and right hands, respectively (from left to right: thumb, index, middle,
ring, little). The third row contains plain impressions of the left four ﬁngers
taken simultaneously, left thumb, right thumb, and right four ﬁngers taken
simultaneously. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
Figure 3.3 Examples of estimated orientation ﬁeld ((b) and (d)) by gradient-based approach [8] for a (a) latent and (c) its mated rolled print in NIST SD27. . . . . . 73
Figure 3.4 Latent ﬁngerprints of three different quality levels in NIST SD27. . . . . . . . . 76
Figure 3.5 Illustration of the orientation ﬁeld computed at a small block. . . . . . . . . . . 79
Figure 3.6 Local ridge orientation estimation based on the nearest minutiae in each sector. . 80
Figure 3.7 Comparison of orientation ﬁeld estimation methods: (a) Original latent ﬁngerprint image, (b) orientation ﬁeld estimated directly from the given image, (c) orientation ﬁeld estimated from minutiae and singular points, and
(d) manually marked orientation ﬁeld. . . . . . . . . . . . . . . . . . . . . . . 82
Figure 3.8 Comparison of orientation ﬁeld estimation methods: (a) Original latent ﬁngerprint image, (b) orientation ﬁeld estimated directly from the given image, (c) orientation ﬁeld estimated from minutiae and singular points, and
(d) manually marked orientation ﬁeld. . . . . . . . . . . . . . . . . . . . . . . 83
Figure 3.9 Sections of two cylinders associated with the two corresponding minutiae,
one in latent and other in rolled print. . . . . . . . . . . . . . . . . . . . . . . . 85
Figure 3.10 Overview of the latent matching with enhanced image approach. . . . . . . . . 87

xiii

Figure 3.11 Examples of enhanced latent images. (a) and (e) are original latent images,
(b) and (f) are the skeletons of the original latent images extracted by VeriFinger 4.2 SDK, (c) and (g) are enhanced images using reconstructed orientation
ﬁeld, and (d) and (h) are the skeletons of the enhanced images extracted by
VeriFinger. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
Figure 3.12 Overview of the proposed latent matching approach. . . . . . . . . . . . . . . . 91
Figure 3.13 Illustration of the MCC descriptor-based Hough Transform alignment.2 . . . . . 93
Figure 3.14 Final score computation scheme combining minutiae and orientation ﬁeld
information. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
Figure 3.15 Latent print for which the matching score of the genuine pair is slightly reduced when a smaller threshold value is used compared to a larger threshold
value, while impostor score is greatly reduced. (a)-(c) show the latent, the
true mate, and the rank-1 non-mate according to large threshold, respectively.
(d)-(g) show latent minutiae that were matched to rolled print minutiae in the
following cases: (d) true mate using small threshold, (e) true mate using large
threshold, (f) non-mate using small threshold, and (g) non-mate using large
threshold. In (d)-(g), the scores corresponding to each case are included.
Filled-in minutiae indicate the matching minutiae. . . . . . . . . . . . . . . . . 98
Figure 3.16 Latent print identiﬁed at a higher rank after fusing minutiae matching scores
with orientation ﬁeld matching scores. The rank of the true mate was improved from 2 to 1 after the fusion, and the rank of the highest ranked nonmate was 3 after the fusion. (a)-(c) show minutiae and the image of (a) a
latent, (b) its true mate, and (c) the highest ranked non-mate according to
minutiae matching. (d) and (f) show latent minutiae and orientation ﬁeld (in
blue) aligned with minutiae and orientation ﬁeld of the true mate. (e) and
(g) show latent minutiae and orientation ﬁeld (in blue) aligned with minutiae
and orientation ﬁeld of the rank-1 non-mate. . . . . . . . . . . . . . . . . . . . 100
Figure 3.17 A latent and its corresponding rolled print in the WVU latent database. The
NFIQ quality [9] of the rolled print is 4, which is one above the worst NFIQ
quality. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
Figure 3.18 Histograms of NFIQ values of rolled prints in NIST SD27 and WVU databases. 103
Figure 3.19 CMC curves for manually marked minutiae, enhanced image using reconstructed orientation ﬁeld, automatically extracted minutiae by Veriﬁnger from
latent image and enhanced image using manually marked orientation ﬁeld. . . . 105
Figure 3.20 CMC curves for manually marked minutiae, enhanced image using reconstructed orientation ﬁeld, highest score rank-level fusion and Borda count
fusion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

xiv

Figure 3.21 CMC curves for score level fusion for latents of all qualities. . . . . . . . . . . 107
Figure 3.22 Matched minutiae shown for manually marked, rolled and enhanced latent
images. (a) Boosted max retrieved the true mate for this latent at rank one
even though one of the retrieved ranks (enhanced image) was 2, 403 and (b)
the retrieved rank for the true mate for this latent is rank one after boosted
max was applied even though neither the manually marked minutiae nor enhanced image retrieved the true mate at rank one. . . . . . . . . . . . . . . . . 109
Figure 3.23 Alignment Accuracy: percentage of correctly aligned latents vs. misalignment threshold. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
Figure 3.24 Example in which DBHT (Descriptor-Based Hough Transform) alignment is
better than GHT (Generalized Hough Transform) alignment. (a) Latent with
manually marked minutiae, (b) corresponding rolled print with automatically
extracted minutiae, (c) rolled print with latent minutiae aligned by GHT, and
(d) aligned by DBHT. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
Figure 3.25 Example in which DBHT (Descriptor-Based Hough Transform) alignment is
better than GHT (Generalized Hough Transform) alignment. (a) Latent with
manually marked minutiae, (b) corresponding rolled print with automatically
extracted minutiae, (c) rolled print with latent minutiae aligned by GHT, and
(d) aligned by DBHT. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
Figure 3.26 Example of alignment error due to the small number of true mated minutia
pairs in the overlapping area between a latent and its mated rolled print. Note
that there is only one aligned minutiae pair here. . . . . . . . . . . . . . . . . . 114
Figure 3.27 Performance of Morpho, MCC SDK, and Proposed Matcher when the union
of manually marked minutiae (MMM) extracted from latents and automatically extracted minutiae by Morpho from rolled prints is input to the matchers. . 115
Figure 3.28 Performance comparison using manually marked minutiae (MMM) and automatically extracted minutiae from latents. . . . . . . . . . . . . . . . . . . . 118
Figure 3.29 Two latent prints from the WVU database correctly identiﬁed at rank-1 by
the proposed matcher but ranked below 20 by Morpho. . . . . . . . . . . . . . 120
Figure 3.30 Latent prints ((a) from NIST SD27 and (b) from WVULFD) whose mates
were not retrieved in the top 20 candidates by the proposed matcher but correctly identiﬁed at rank-1 by Morpho matcher. . . . . . . . . . . . . . . . . . . 121
Figure 3.31 Score-level fusion of the proposed matcher and Morpho for NIST SD27 and
WVU databases. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122

xv

Figure 3.32 Latent print mate from NIST SD 27 identiﬁed at rank 1 after score-level
fusion of Morpho and proposed matcher. The ﬁrst column shows (a) a latent,
(c) its true mate, (e) rank-1 non-mate by the proposed matcher, and (g) rank-1
non-mate by Morpho matcher. The second column shows (b) latent minutiae,
(d), (f), and (h) latent minutiae (in blue) aligned by the proposed matcher to
the rolled print minutiae shown in (c),(e) and (g). In (c), (e), and (g), the
numbers in parentheses indicate the ranks at which true mated rolled print
was retrieved by the proposed matcher and Morpho matcher, respectively. . . . 126
Figure 3.33 Latent print mate from WVU LFD identiﬁed at rank 1 after score-level fusion
of Morpho and proposed matcher. The ﬁrst column shows (a) a latent, (c) its
true mate, (e) rank-1 non-mate by the proposed matcher, and (g) rank-1 nonmate by Morpho matcher. The second column shows (b) latent minutiae,
(d), (f), and (h) latent minutiae (in blue) aligned by the proposed matcher to
the rolled print minutiae shown in (c),(e) and (g). In (c), (e), and (g), the
numbers in parentheses indicate the ranks at which true mated rolled print
was retrieved by the proposed matcher and Morpho matcher, respectively. . . . 128
Figure 3.34 Score-level fusion of the two proposed approaches: by enhancement and
using descriptor-based Hough Transform (DBHT) matching approach. By
combining the two approaches, the rank-100 accuracy is 81%. . . . . . . . . . . 130
Figure 4.1 Examples of ﬁngerprints. (a) Rolled print (Michigan State Police database),
(b) plain or slap print (FVC 2002) and (c) latent print (NIST SD27). . . . . . . 133
Figure 4.2 (a) Reference print from Michigan State Police database, (b) its skeleton extracted by a commercial matcher, (c) orientation ﬁeld extracted from skeleton
image, and (d) minutiae extracted by a commercial matcher. . . . . . . . . . . . 141
Figure 4.3 (a) Skeleton of a reference print from Michigan State Police database, (b) the
grayscale image obtained after applying image ﬁltering and (c) the extracted
orientation ﬁeld. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
Figure 4.4 (a) Latent print from NIST SD27 latent database, and some manually marked
features: (b) orientation ﬁeld, (c) singular points and (d) minutiae. . . . . . . . 143
Figure 4.5 Overview of the proposed indexing scheme. . . . . . . . . . . . . . . . . . . . 145
Figure 4.6 Illustration of triplets features. . . . . . . . . . . . . . . . . . . . . . . . . . . 147
Figure 4.7 Illustration of the key generation in the orientation ﬁeld descriptor indexing
scheme. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
Figure 4.8 Rolled prints in (a) NIST SD27, (b) NIST SD14, and (c) Michigan State
Police databases. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153

xvi

Figure 4.9 Comparison of the indexing performance of basic triplets, constrained triplets
and MCC indexing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
Figure 4.10 Comparison of indexing performance based on the proposed approach against
Yuan et al. [10] and Feng and Jain [11] on the NIST SD27 latent database.
Note that Feng and Jain [11] used only a small background database of
10,258 versus the proposed approach and [10] that use 267,258 reference
prints in the background database. Further, Feng and Jain [11] did not provide the complete hit rate vs. Penetration rate curve, but only a single point
at a penetration rate of 39% (•). . . . . . . . . . . . . . . . . . . . . . . . . . . 155
Figure 4.11 Increase in the cumulative hit rate at a ﬁxed penetration rate (20%) as we
incrementally add features for indexing (from left to right), where CTrip denotes constrained triplets, OFDesc denotes orientation ﬁeld descriptor, SP
denotes singular points, RP denotes ridge period. The plot shows that we
can achieve about 91% hit rate at 20% penetration rate when we use all ﬁve
features (CTrip, OFDesc, MCC, SP and RP) for indexing. . . . . . . . . . . . . 156
Figure 4.12 Latent matching performance of two different matchers (DBHT in-house latent matcher and Morpho matcher) before and after applying the proposed
indexing to ﬁlter out 80% of the background database. Note that after the
ﬁltering and the fusion of indexing and match scores, the overall matching
performance is improved while the matching time is reduced by a factor of
ﬁve (20% penetration rate). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
Figure 4.13 Examples of latents for which the true mate was correctly retrieved when
20% of the background database was considered are shown. Note that by
applying the indexing, the retrieval rank went from (a) 3 to 1 and (b) 142 to
49 due to the exclusion of non-mated reference prints with high match scores
to the latent. In (b), the true mate only appeared in the top-100 candidate list
after the ﬁltering. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
Figure 4.14 Latent matching performance of a commercial latent ﬁngerprint matcher before and after applying the proposed indexing and retaining one third of
the background database. The ﬁltering improves the overall matching performance of the commercial latent ﬁngerprint matching while reducing the
matching time. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
Figure 4.15 Examples of latents for which the true mate was correctly retrieved when
one third of the background database was considered are shown. Note that
by applying the indexing, the retrieval rank went from last to 1 due to the
fusion of indexing and match scores. . . . . . . . . . . . . . . . . . . . . . . . 161

xvii

Figure 4.16 Examples of latents for which the true mate was not retrieved when one third
of the background database was retained. The true mates retrieval ranks before the ﬁltering were 1, so the indexing degrades the matching performance
for these two latents. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162

xviii

CHAPTER 1
INTRODUCTION
The ﬁrst systematic capture of biometric data for identiﬁcation purpose was done by Sir William
Herschel in 1858, when he captured a handprint of each worker so that, later at payday, he could
verify each worker’s identity to avoid a worker receiving someone else’s pay [12]. In 1870,
Alphonse Bertillon [13] developed a method of person identiﬁcation based on body measurements
known as “Bertillonage” (see Fig. 1.1). This method was developed to identify criminals who have
been previously arrested, repeat offenders, and who were using a different name to avoid harsher
penalties. The Bertillon system recorded the precise measurements of various attributes of human
body such as height, length of the arm and geometry of the head, as well as a listing of marks
present on the body surface such as scars, moles and tattoos. This information was then stored and
separated into categories for fast matching and retrieval. This system was believed to be very reliable and it was indeed used in law enforcement from the late 19th century to the early 20th century.
According to [14], the French police used the Bertillon system to capture 241 repeat offenders in
the year of 1884; after that, this system was adopted by law enforcement agencies both in Europe
and in the Americas. However, it was later discovered that those measurements were not unique to
each person; furthermore, two measurements of the same subject taken by different ofﬁcers were
not always consistent. Thus, Bertillon system was considered unreliable and abandoned by the
police authorities to be replaced by ﬁngerprint identiﬁcation.

1

Figure 1.1: Illustration of Bertillon system for person identiﬁcation based on numerous body measurements, which was used in the United States and in Europe from the end of the 19th century
to the beginning of the 20th century. From left to right and then top to bottom the ﬁgures show
measurement of height, reach, trunk, length of head, width of head, right ear, left foot, left middle
ﬁnger, and left forearm.

2

A few years after Bertillon system was developed, several publications on the use of ﬁngerprints as a mean of person identiﬁcation started to appear [15, 16, 17]. According to Cummins and
Midlo [18], the ﬁrst scientiﬁc publication related to ﬁngerprint identiﬁcation appeared in Nature, in
1880 [15]. In [15], Faulds suggested that ﬁngerprints left at crime scenes could be used to identify
criminals or to exclude suspects. Soon after this article was published, a letter written by Herschel
was published in Nature [16] stating that he had been using ﬁngerprint as a method of identiﬁcation
in India for about 20 years, with different applications such as to avoid personiﬁcation. In 1888,
Sir Francis Galton introduced the idea of using minutiae features to compare two ﬁngerprints [17].
Sir Francis Galton also introduced a categorization for ﬁngerprints into three major classes (arch,
loop and whorl), and the classes were further divided into subclasses.
After reading Sir Francis Galton’s work on using ﬁngerprints as a mean of identiﬁcation, Juan
Vucetich started collecting ﬁngerprints from arrested men, along with measurements from the
Bertillon system. Juan Vucetich was a police ofﬁcial in Argentina [19], and, in 1892, he used
ﬁngerprint evidence to identify a suspect of ﬁrst-degree murder, who was then convicted. This was
the ﬁrst use of ﬁngerprint as forensic evidence.
In 1893, the acceptance of the hypothesis by the Home Ministry Ofﬁce, UK, that any two
individuals have different ﬁngerprints made many law enforcement agencies aware of the potential
of using ﬁngerprints as a mean of identiﬁcation [1]. Some of the law enforcement agencies started
collecting ﬁngerprints from offenders so that they could identify them later in case they used an
alias (changed their names) to evade harsher penalties. The ﬁngerprints collected from crime
scenes were compared to ﬁngerprints collected from previous offenders in order to identify repeat
offenders.
In 1900, Edward Henry has reﬁned the ﬁngerprint classiﬁcation scheme introduced by Galton
by adding more categories [1]. This resulting classiﬁcation scheme (referred to as Galton-Henry
classiﬁcation) was adopted by several law enforcement agencies in various countries and it has
been widely used.
The initial interest in recognition technologies were mainly by law enforcement agencies; more

3

recently, the concerns about security and identity fraud have increased, therefore creating a need for
biometric recognition technologies in non-forensic applications [1] (e.g. border control, national
ID, etc.). Over the years, several physical and behavioral characteristics have been explored, leading to the development of new recognition technologies based on face, ﬁngerprint, hand geometry
and voice that have been successfully deployed.
The use of physical or behavioral characteristics (e.g., ﬁngerprint, iris, signature, etc.) to automatically recognize a person is referred to as biometrics or biometric system. A biometric characteristic (or trait) is a measurable physical or behavioral characteristic of an individual that is
distinguishable and repeatable. A biometric trait should have four main characteristics: (i) be universal, meaning most people should have the trait; (ii) be distinctive, meaning it should be different
from person to person; (iii) be permanent, meaning it should be invariant (respective to matching
criteria) over time; and (iv) be collectable (measurable), meaning it can be measured quantitatively
[1]. These characteristics make biometrics a reliable solution to person recognition. Additionally,
biometric traits cannot be transferred, forgotten, guessed, shared or lost like other means of person
recognition such as passwords, ID-cards, etc.
Biometric recognition can be very challenging because of intra-class variability and inter-class
similarity of biometric traits [5]. Intra-class variability refers to how a biometric trait can appear
very different in multiple acquisitions of the same individual, and inter-class similarity refers to
how a biometric trait can be very similar for different persons. Intra-class variability is exempliﬁed
in Fig. 1.2a, in which two different impressions of the same ﬁnger are shown, and inter-class
similarity is exempliﬁed in Fig. 1.2b, in which two face images of a pair of identical twins are
shown. Although the two images shown in Fig. 1.2a are impressions of the same ﬁnger of the
same person (same class), the second impression in the ﬁgure contains a large amount of distortion
compared to the ﬁrst one. In Fig. 1.2b, the two face images are from different persons (different
classes) but since they are identical twins, their facial appearance is extremely similar.

4

(a)

(b)
Figure 1.2: Examples of intra-class variability and inter-class similarity. (a) Two different impressions of the same ﬁnger of the same subject and (b) face images of a pair of identical twins. For
interpretation of the references to color in this and all other ﬁgures, the reader is referred to the
electronic version of this dissertation.

1.1 Biometric Systems
A biometric recognition system involves two basic phases: enrollment and recognition [5]. In the
enrollment phase, a sensor collects the biometric data from which a set of features are extracted
(template) and stored in a database along with the individual’s identity (name, ID number, etc.). In
the recognition phase, the identity of an individual is either conﬁrmed (veriﬁcation) or determined
(identiﬁcation) by collecting the biometric data again, extracting the same features and comparing

5

them to the features stored in the database. From this comparison, a match (similarity) score is
generated and used to make a decision to whether the two sets of features came from the same
subject or not. An overview of a biometric system is shown in Fig. 1.3.

Figure 1.3: An overview of a biometric system.

A biometric system can be designed to work in two different modes: (i) veriﬁcation mode or
(ii) identiﬁcation mode. In the veriﬁcation mode, the user will claim his/her identity by using a
personal identiﬁcation number, a user name, a token, etc; then, the system will verify whether
the subject is the person he/she is claiming to be (genuine) or not (impostor) by using biometric
traits. In this case, only the stored data related to the claimed identity is compared to the data of
the person making the claim to an identity. A threshold on the match score is usually applied to
decide whether the two biometric samples are true mates. If the match score is higher than the
system threshold, then the samples are considered to be from the same source. Otherwise, they are
considered to be an impostor pair. An example of the veriﬁcation scenario occurs when you try
to use the ATM at a bank and you have to provide biometrics data along with your ATM card to
verify your identity. In this case, the owner of the ATM card is known, so the biometric system
needs to ensure that the true owner is the one who is using the card to perform the transaction.
In the identiﬁcation mode, the user does not claim any identity. The user only provides his/her
6

biometric data, and the data is compared to the stored template of every subject in the system
database. In the ATM example cited above, now the ATM account holder will not insert his ATM
card in the machine, but simply provide his biometric data for identiﬁcation. If the system can
ﬁnd a subject in the database for whom the query biometric data is close enough, then the query
is considered to be from the same subject. Otherwise, the system will output that the user is not
enrolled in the database. For example, when a ﬁngerprint impression is found at a crime scene,
forensic examiner usually does not know to whom the ﬁngerprint belongs. So, the crime scene
ﬁngerprint is compared to the reference (rolled) prints stored in the database. If a match is found,
the identity of the suspect is determined.
A match score is called a genuine score if it is computed between mated samples and an impostor score if it is computed between non-mated samples. In order to evaluate and compare the
performance of biometric systems, the most commonly used quantitative measures are Receiver
Operating Characteristic (ROC) curve in the veriﬁcation mode and Cumulative Match Characteristic (CMC) curve in the identiﬁcation mode [5].
The ROC is a plot of the Genuine Accept Rate (GAR) versus the False Acceptance Rate (FAR)
as the match score threshold is varied. The GAR refers to the portion of true mated subjects who
are correctly accepted by the system, while FAR refers to the portion of impostor mates who are
falsely accepted by the system at a speciﬁed threshold. While the desired values of GAR and FAR
are application dependent, a biometric system, operating in the veriﬁcation mode with high GAR
at low FAR is, in general, preferred.
In the identiﬁcation mode, given a query, it is desirable that the match score between this query
and its true mate be the highest score compared to the match scores between the query and any
impostor. For each query, the database can be ordered based on the match score. The identity in
the database corresponding to the highest score is considered the rank-1 match, while the identity
corresponding to the second highest score is considered the rank-2 match and so on. Ideally, we
want the true mate to be in the rank-1 position. The CMC curve is a plot of the identiﬁcation rate
versus the rank. Given a number of queries and a rank t, we can measure the percentage of the true

7

mates that are retrieved at rank-t or higher; this percentage is the identiﬁcation rate at rank-t. The
system with higher identiﬁcation rates for given queries is the most desirable system.

1.2 Common Biometric Traits
Several biometric traits have been used for recognition, such as face, ﬁngerprint, iris, palmprint,
hand geometry, voice, ear shape, signature, key stroke and gait [5]. We focus on ﬁngerprint, face
and iris, which are the most prominent traits and are used in this dissertation.
Each biometric trait has its strengths and weaknesses, and the choice of a biometric trait is
usually dictated by the application requirements. Table 1.1 summarizes advantages and disadvantages of the three biometric traits (ﬁngerprint, face and iris) [1]. For each trait, the authors in [1]
provided a perceived measure of the four characteristics: universality, distinctiveness, permanence
and collectability. For example, the high collectability of face means it is very easy to collect this
biometric trait.
It is generally known that ﬁngerprints are very distinctive, and the overall ridge structure does
not change signiﬁcantly over time (from infants to old age); even after superﬁcial cuts on the ﬁnger,
the ﬁngerprint pattern reappears after the healing process. This is supported by the high matching
performance of state of the art systems in rolled/slap ﬁngerprint matching [20]. Fingerprints are
relatively easy to acquire, but require some degree of cooperation from the subjects (collectability
is “medium” in Table 1.1). Also, some people might not have strong ﬁngerprint pattern suitable
for automatic recognition due to genetic factors, occupation, aging, etc [5].
Face is one of the most convenient biometrics in terms of acquisition and it does not even reTable 1.1: Comparison of the characteristics of the three most commonly used biometric traits
from [1]. (H = high, M = medium, L = low)
Biometric Trait

Universality

Distinctiveness

Permanence

Collectability

Fingerprint

M

H

H

M

Face

H

L

M

H

Iris

H

H

H

M

8

Table 1.2: False reject and false acceptance rates associated with ﬁngerprint, face and iris veriﬁcation systems.
Biometric
trait
Fingerprint

Face
Iris

Test

Test conditions

FVC 2006

Heterogeneous
population
including manual
workers and
elderly people
Operational
quality
Mugshots
Visa Photos
Controlled
illumination

FpVTE
2003
FRVT 2012
FRVT 2012
ICE 2006

False reject
rate (%)
4.2

False accept
rate (%)
0.1

0.6

0.01

4.0
1.0
1.1-1.4

0.1
0.1
0.1

quire cooperation from the subject. Face images can be easily acquired, even from a distance and
without the subject’s knowledge. This advantage is specially useful in surveillance applications.
However, many variations in face images such as pose, lighting, expression, and changes in appearance such as make-up and accessories, make face recognition a very challenging problem [5].
This explains the low distinctiveness of state-of-the-art face recognition systems in unconstrained
scenarios. Furthermore, face characteristics might not be stable over time (e.g., due to weight gain
or aging).
Iris image capture requires a more sophisticated and expensive sensor because the useful texture
patterns are better captured in near-infrared images; it also requires the subject’s cooperation (subject must stand at a speciﬁed distance from the iris camera) and the iris quality can be inﬂuenced
by a number of factors such as partially closed eyelids, eyelashes, contact lenses, etc. However,
studies on large scale databases suggest that iris recognition systems can achieve extremely low
error rates [21].
Table 1.2 shows false reject and false acceptance rates associated with ﬁngerprint, face and iris
veriﬁcation systems [5].
In the next section, we review the representation (feature extraction) and matching techniques
of ﬁngerprint, face and iris biometric modalities.
9

1.2.1 Fingerprint Representation and Matching
A ﬁngerprint is the impression of the friction ridge skin on a ﬁnger tip. Friction ridge skin presents
raised ridges because their function is related to grasping and gripping; this explains their presence
in the palms of our hands and sole of our feet. The characteristics of a ﬁngerprint are determined
during fetal development and its formation starts at approximately 6 or 7 weeks of gestational age
[22]. The ﬁngerprint pattern is mostly determined by the gestational environment, since minor
changes in the ﬂow of amniotic ﬂuids are responsible for the differences in the skin structures
around palm or ﬁnger tips. Thus, the ﬁngerprints of every person are different among themselves,
and different from ﬁngerprints of other persons.
There are essentially three types of ﬁngerprint images that are acquired for matching: (i) rolled,
which is obtained by rolling the ﬁnger from “nail-to-nail" either on a paper (in this case ink is ﬁrst
applied to the ﬁnger surface) or the platen of a scanner; (ii) plain, which is obtained by placing the
ﬁnger ﬂat on a paper or the platen of a scanner without rolling; and (iii) latent, which is lifted from
surfaces of objects that are inadvertently touched or handled by a person typically at crime scenes
(see Fig. 1.4). Rolled and slap ﬁngerprints are generally captured under controlled conditions and
by cooperative subjects. On the other hand, latent prints are left by criminals at crime scenes. For
this reason, rolled and plain prints (collectively referred to as reference prints) are of much better
quality than latents. The two ﬁrst types (rolled and plain) are usually collected in the form of a
tenprint card, which is a card that contains rolled and plain impressions of the ten ﬁngers of a
subject, along with their identity information (see Fig. 3.2).

1.2.1.1 Representation
The structure present in a ﬁngerprint is composed of ridges and valleys. In a rolled ﬁngerprint
image obtained by using ink (e.g. Fig. 1.4 (a)), the ridges are the dark areas corresponding to the
raised ridges in our ﬁngers and valleys are the bright areas that correspond to the space between
the raised ridges.
Fingerprint characteristics or features can be categorized into three different levels (from coarse
10

(a) Rolled

(b) Plain

(c) Latent

Figure 1.4: Three types of ﬁngerprint images.
to ﬁne): Level 1 (ridge ﬂow), Level 2 (minutiae) and Level 3 (pores, incipient ridges, dots, etc) [1].
Fig. 1.5 shows examples of features in the three levels.

(a)

(b)

(c)

Figure 1.5: Examples of ﬁngerprint features belonging to the three levels: (a) orientation ﬁeld and
singular points and (b) minutiae shown in ﬁngerprints from NIST Special Database 4 [2], and (c)
sweat pores shown on part of a rolled print image from the WVU Latent Database [3].

• Level 1 features
This is the coarsest level representation and consists mainly of the ridge orientation map and
ridge frequency map of the ﬁngerprint. The ridge orientation (ridge frequency) map consists
of the local orientation (frequency) of the ridges in the ﬁngerprint. The local orientation of
a ridge at point (x, y) in the ﬁngerprint image is deﬁned as the orientation of the line tangent
to the ridge passing through (x, y), and it is in the range [0, π ). The local ridge frequency

11

at point (x, y) is deﬁned as the average number of ridges passing through a line segment
centered at (x, y) and of unit length that is normal to the local ridge orientation.
Ridge orientation is mostly smooth in ﬁngerprints. However, there can be a few areas of
the ﬁngerprint in which the ridge orientations change abruptly. These locations are called
singular points. There are two types of singular points: core, in which a set of ridges enters
and exits from the same direction, and delta, in which three sets of ridges appear to meet.
The number of cores and deltas in a ﬁngerprint is determined by the ﬁngerprint type. The
maximum number of cores and deltas in a ﬁngerprint is 4, and a ﬁngerprint always contains
an even number of singular points because cores and deltas appear in pairs. Therefore, a
ﬁngerprint might have no singular points (arch or tented arch type ﬁngerprint), one core and
one delta (loop type ﬁngerprint), or two cores and two deltas (double loop or whorl type).
Examples of ridge orientation map and singular points are shown in Fig. 1.6.

(a)

(b)

Figure 1.6: Examples of Level 1 features in ﬁngerprints. (a) Orientation ﬁeld and (b) singular points (core in red and delta in green) shown in rolled ﬁngerprint images from NIST Special
Database 27 [4].
Ridge orientation is also referred to as orientation ﬁeld or ridge ﬂow. We will mainly use
the term orientation ﬁeld throughout this dissertation to refer to this ﬁngerprint feature. The

12

orientation ﬁeld is commonly extracted by using a gradient-based method in local neighborhoods [8]. The ﬁngerprint image is usually divided into small non-overlapping blocks
(usually 8 × 8) and one dominant orientation is computed for each block as the orthogonal
orientation of the gradient angle, the latter being the direction of the maximum intensity
change. Ridges and valleys are more than one pixel wide, so to consider only one angle at
a given pixel position makes it sensitive to noise. Therefore, an average of the gradient in
a small neighborhood (window) is computed as the gradient angle estimate at a given pixel.
This averaging is performed with following restrictions: (i) the angles are doubled so that
the circularity of the orientations is maintained — for example the average of the angles 5◦
and 175◦ is 0◦ instead of 90◦ and (ii) the sin and cos components are used instead of angle
values.
Based on the ridge structure, ﬁngerprints can be classiﬁed into ﬁve classes: left loop, right
loop, whorl, arch and tented arch [23] (see Fig. 1.7). Loops (left and right) are characterized
by the ridge ﬂow that enters from one side of the ﬁngerprint, curves and returns on the same
side from which it came; whorls are characterized by the ridge ﬂow forming a complete
circuit; arches are characterized by the ridge ﬂow that enters from one side and exits the
opposite side; and tented arches are similar to arches, but at least one ridge presents a high
curvature.
• Level 2 features
The ridges in a ﬁngerprint exhibit discontinuities in various ways. The locations where the
discontinuities occur are called minutiae (minute or small detail). The most commonly used
discontinuities are the locations where a ridge ends and the locations where a ridge bifurcates. A rolled print typically has about 100 minutiae points; the exact number depends
upon the ﬁnger surface condition and image capture process. A minutia can be represented
by its location (x and y coordinates), direction (the orientation of the ridge associated with the
minutia), and type (ending or bifurcation). Minutiae characteristics of permanence, unique-

13

(a) Left Loop

(c) Whorl

(b) Right Loop

(d) Arch

(e) Tented Arch

Figure 1.7: Example ﬁngerprints belonging to each of the ﬁve major ﬁngerprint classes (ﬁngerprints from NIST Special Database 4 [2].)
ness and ease of representation make minutiae the most commonly used feature in ﬁngerprint
matching. Minutia type is often not reliable because depending on the pressure that is applied to make the impression, a bifurcation can look like an ending and vice versa. The
minutia direction deﬁnition is illustrated in Fig. 1.8; this deﬁnition ensures that the error in
minutia type does not inﬂuence the direction estimation.
The most common method of extracting minutiae is to use binarization and thinning techniques to obtain the ﬁngerprint ridge skeleton (one pixel wide dark ridges on a white background). The binarization consists of thresholding the ﬁngerprint image so that the ridges
are separated from the rest of the image (black ridges on a white background). Then, morphological operations using the ridge orientation are applied to reduce the width of the ridges
to one pixel. From the skeleton, minutiae can be extracted by ﬁnding the endings and bi-

14

(a)

(b)

Figure 1.8: An illustration of how minutia direction is deﬁned for (a) minutia ending and (b)
minutia bifurcation [5].
furcations of the one pixel wide ridges. Usually some type of image enhancement (such as
applying Gabor ﬁlters) is necessary before the binarization and thinning procedures. The
goal of this enhancement step is to correct degradations occurring in ridge patterns that are
very noisy and/or corrupted. As an example, parallel ridges might not be well separated
because of the noise.
• Level 3 features
The Level 3 features can be viewed as micro level characteristics of a ﬁngerprint. Level 3
features include pores (sweat pores), incipient ridges, dots, dimensional attributes of ridges
such as width and shape, etc. [1, 5]. These features are not usually observed in low resolution images (less than 1000 ppi). Level 3 features are very important for latent ﬁngerprint
examiners, especially when the number of minutiae in latent is too small (e.g. less than 5),
because they are highly distinctive and easily observed in high resolution images (over 1000
ppi). However, they are not widely used in automatic ﬁngerprint matching systems because,
even in high resolution images, their extraction is both computationally demanding and not
as reliable as level 1 and level 2 features. In [24], the use of level 3 features is supported
for latent print matching because forensic experts often rely on additional information, especially when the number of minutiae in latents is small, to compare latents to rolled prints.
Fig. 1.5 (c) shows examples of sweat pores, which are regularly spaced along the ridges.

15

1.2.1.2 Matching
Most published ﬁngerprint matching algorithms are based on minutiae. A few techniques are
based on image correlation methods (e.g., [25, 26]). In the latter, the features are mostly the pixel
intensities. In this case, the correlation between two ﬁngerprint images is computed for alignment,
which can be performed globally or locally. The images of different impressions of the same ﬁnger
might appear very different due to pressure variations (ridge thickness, contrast, global structure,
etc), which largely affects the correlation between two images. Also, the methods included in this
category might be computationally very expensive [1].
There are other methods that are feature-based, where the features include singular points,
level 3 features, texture information, etc. (e.g., [27, 28]). These methods sometimes are used in
combination with minutiae to improve the matching performance and they are especially useful in
cases where it is very difﬁcult to extract minutiae or the number of minutiae is small, as often is
the case with latents. However, these non-minutiae features are not as distinctive as minutiae, so
they can only be used in conjunction with minutiae.
As mentioned before, the most common ﬁngerprint matching algorithms are minutiae-based.
There are three main steps in ﬁngerprint matching using minutiae: alignment, pairing and score
computation. Alignment refers to estimating the parameters between two minutiae sets that can be
used to transform one of the sets to the same coordinate system as the second set; pairing refers to
ﬁnding corresponding minutiae; and score computation refers to assigning a match score to a pair
of ﬁngerprints usually based on the number of corresponding minutiae.
Two common ﬁngerprint alignment methods are based on (i) pairs of matched minutiae and (ii)
Generalized Hough Transform. In the ﬁrst method, a pair of matched minutiae is found using, for
example, a minutia descriptor. A minutia descriptor is a structure that contains information about
the local neighborhood of the minutia. This information can be based on neighboring minutiae, or
the texture around minutia, etc. The local minutiae descriptors are matched and a small number of
most similar minutiae candidate pairs are used for initial alignment. The rotation in the alignment
is derived from the direction difference between the minutiae pair, and translation parameters are

16

estimated as the distance in x and y coordinates between the two minutiae.
Alignment based on the Generalized Hough Transform [29] consists of ﬁnding the peaks in the
parameter space associated with the rigid transformation between the two sets. In this case, each
pair of minutiae (one in each ﬁngerprint) will vote for a speciﬁc set of translation and rotation parameters computed as mentioned above. Then, the peak in the parameter space (Hough Transform)
is obtained — usually more than one peak is selected for robustness.
After the ﬁngerprints are aligned, the pairing consists of ﬁnding the minutiae correspondences
between the two ﬁngerprints. The simplest way of ﬁnding the correspondences is to consider a pair
of minutiae as matched minutiae if the distance between them and their directional difference are
smaller than some pre-speciﬁed thresholds (e.g., 15 pixels in translation and 20◦ in rotation). Fig.
1.9 shows an example of minutiae correspondences in a ﬁngerprint pair (latent and its true mate).

Figure 1.9: Example of matched minutiae from a ﬁngerprint pair (latent and rolled print).
Score computation can be done in a number of ways. However, the most inﬂuential feature
in the scoring process is the number of matched minutiae. Other major features used in match
score computation include ridge ﬂow, ridge frequency, etc. One simple way of generating a score
between two ﬁngerprints is to count the number of matched minutiae, and to normalize this number by the average number of minutiae in both ﬁngerprints. In cases where the overlapping area
between the two ﬁngerprints is small, the score will be excessively penalized by this normalization
scheme; so, the number of minutiae in the overlapping area should be used in the averaging instead

17

of the total number of minutiae in the ﬁngerprints. Some other features might also be useful in the
scoring, such as the quality of minutiae and the similarity of local minutiae descriptors.

1.2.1.3 Indexing
In the identiﬁcation mode, a search ﬁngerprint is compared to the reference prints in the background database to ﬁnd the true identity of the subject, if present. When the background ﬁngerprint database is extremely large (say, tens of millions), ﬁngerprint matching becomes a challenge
in terms of computational load. A common approach to alleviate this problem is to quickly exclude
a large number of reference ﬁngerprints with low similarity to the search ﬁngerprint before performing the more detailed one to one matching. This process of ﬁltering out a large percentage of
the background database quickly (e.g., based on level 1 features) before ﬁne level matching is referred to as database ﬁltering, ﬁngerprint indexing or ﬁngerprint retrieval. Indexing usually refers
to an approach that provides continuous classiﬁcation rather than exclusive classes or categories,
while the terms ﬁltering and retrieval can be used in both cases.
Approaches that have been proposed for ﬁngerprint indexing can be mainly divided into three
categories: minutiae-based, orientation ﬁeld-based, and based on other features (e.g. SIFT [30]).
In the minutiae-based indexing case, descriptors based on minutiae are utilized (e.g. triplets,
minutia cylinder code). The most common approach is based on minutiae triplets, or the triangles
formed by sets of three minutiae in the ﬁngerprint. Rotation and translation invariant features
such as the side lengths of the triangles, the difference between minutiae direction and one side
of the triangle, handedness of the triangle, minutiae type, etc., are quantized into bins to generate
a unique key or index for each triangle. Given a search or query ﬁngerprint, triplets in the query
are generated and quantized in the same way to obtain the keys. Then, triplets in the background
database are retrieved as matched triplets if they have the same key as the search triplets (e.g.
[31, 32]). The number or matched triplets between the search and each ﬁngerprint in the database
can be used to retrieve the ﬁngerprints that are most similar to the search print.
Orientation ﬁeld-based indexing usually involves aligning the ﬁngerprints with respect to some

18

reference point (e.g. singular points), and then clustering the background database so that the orientation ﬁeld of the search print only needs to be compared to a limited number of representative
orientation ﬁelds (e.g. [33]). All ﬁngerprints in the closest cluster center are retrieved. Another
approach involves representing the orientation ﬁeld by coefﬁcients of an orientation ﬁeld model,
using the coefﬁcients to search the background database, and retrieving the ﬁngerprints with orientation ﬁeld coefﬁcients similar to the ones of the search print [34].
Indexing performance is usually measured by computing the hit rate at various penetration
rates. A hit rate at a given penetration rate p refers to the portion of the search prints for which the
true mates are retrieved within p% of the background database (penetration rate). The desirable
outcome for ﬁngerprint indexing is high hit rates at low penetration rates, which means that only a
small portion of the database needs to be searched (or ﬁne matched) in order to ﬁnd the true mate
for a high percentage of the given search prints; high computational efﬁciency is also desired so
that the overall search time is reduced compared to ﬁne matching the search print to every reference
print in the background database.
In order to be useful, indexing must be much faster than the more detailed one-to-one matching. However, it still needs to have a good performance so that the matching performance is not
degraded. There is a tradeoff between the penetration rate and the hit rate. If a very small penetration rate is chosen, some true mate ﬁngerprints might be excluded from the ﬁner matching, thus
the identiﬁcation performance will drop. On the other hand, if a large penetration rate is chosen,
the true mate is more likely to be in the retained background database, but a large portion of the
database will need to be searched, thus making the total gain in speed very small.

1.2.1.4 Latent Fingerprints
The sweat and the contact of a ﬁnger touching a surface leaves an impression of the ridges on these
surfaces. This type of ﬁngerprint impressions are called latents. Although latents can refer to any
impressions lifted from surfaces touched by a person, we usually use this term for impressions of
ﬁngers lifted at crime scenes. The moisture in the ﬁngers that is transferred to object surface is

19

not usually visible. Thus, some processing methods, typically chemical [35], must be applied to
make them visible so that a copy of the impression can be lifted from the objects and scanned.
Alternatively, a digital photograph can be taken directly to generate a latent ﬁngerprint image.
Latent prints can be lifted from a variety of objects and materials such as metal, paper, plastic, human skin, glasses, etc. Some of the characteristics of latent ﬁngerprints that makes latent
matching a very difﬁcult problem compared to reference to reference print matching include: (i)
the uneven pressure of the ﬁnger and the non-ﬂatness of the surface can generate a distorted impression of the ﬁnger; (ii) the area of the ﬁngerprint ridge structure contained in a latent impression
is usually very small; (iii) the way the object was touched by the ﬁnger might generate a blurred
or smudged impression; and, (iv) in the case of a digital photograph of the latent, the image might
contain a very noisy background. All these factors heavily inﬂuence the quality of the latent, and
pose a major challenge to latent matching. Some examples of latent ﬁngerprints are shown in Fig.
1.10.

(a)

(b)

(c)

(d)

Figure 1.10: Examples of latent ﬁngerprints in NIST SD 27 database [4].

Due to the aforementioned factors, latent ﬁngerprint images are generally of very poor quality
compared to rolled and plain ﬁngerprint, which makes it very difﬁcult to automatically extract reliable features from them. Thus, features in latents are usually manually marked by latent examiners.
Features that are commonly marked by latent examiners are minutiae, singular points and region
of interest (ROI). Furthermore, because latents capture only a partial impression of the ﬁnger, the
number of minutiae in latents compared to plain or rolled prints is small. For example, in NIST SD
27 [4], the average number of minutiae is 21 in the latents and 106 in the mated rolled prints. Since

20

latent prints can be lifted from a variety of object surfaces in uncontrolled environments, latents
are also more likely to contain large amount of distortion. The feature extraction in latents and
the matching of latent prints to reference prints are challenging and they require more specialized
algorithms than the ones available for matching reference ﬁngerprints (rolled to rolled or slap to
slap).
NIST evaluation of reference ﬁngerprint matching technology reports excellent performance.
In the Fingerprint Vendor Technology Evaluation (FpVTE) [20] multiple ﬁngerprint recognition
systems were evaluated on different types of data and scenarios. The experiments were divided into
three main groups: large-scale test (LST), medium-scale test (MST) and small-scale test (SST).
The best performing ﬁngerprint recognition system, evaluated on a medium-scale test and using
single ﬁngerprints of operational quality, reached a true accept rate of 99.4% at 0.01% false acceptance rate. Latent ﬁngerprint matching technology has also been evaluated. In Phase I of the
Evaluation of Latent Fingerprint Technologies (ELFT), the average reported rank-1 identiﬁcation
accuracy (100 latents against 10, 000 rolled prints) was 67%. In Phase II, the best reported rank-1
identiﬁcation accuracy (835 latents against 100, 000 rolled prints) was 97.2%. It is important to
clarify that in Phase II, the latent images selected were of very good quality compared to Phase I,
which explains the impressive performance of Phase II results over Phase I results.
A tentative benchmark was established for ﬁngerprint indexing, but no systematic evaluation is
yet available [36]. Several researchers have reported the performance of their indexing approaches
on publicly available databases such as NIST SD4 [2] and FVC databases [37]. For example, at
20% penetration rate, the hit rate is around 98% for NIST SD4 and as high as 100% for FVC
databases [38]. The indexing performance for latents is not as good as for the reference prints. The
reported hit rates for latents in NIST SD27 (the only public domain latent database) are: 97.3% at
39% penetration rate [11], and 92.7% at 40% penetration rate [10]. The results reported in [10]
could not be veriﬁed by us since by applying the algorithm in [10] (code was provided to us by
the authors) to our background database of similar size, the hit rate was only around 82% at 40%
penetration rate.

21

1.2.2 Face Representation and Matching
Face is composed of the skull characteristics and the musculature and associated soft tissue [22].
These structures inﬂuence the variation among human faces, along with gender and age. Some
challenges to face recognition systems include matching in the presence of variations in pose,
lighting, expression, occlusion, weight changes, hair style, etc.

1.2.2.1 Representation
The ﬁrst step in many face-related applications is face detection, in which the face location in the
sensed image is determined. This process is usually done by placing masks or ﬁlters of varying
size in the image and solving the two-class classiﬁcation problem: face vs non-face for each patch.
After the face is detected and localized inside the image, facial features can be extracted. Face
images are usually aligned based on the eye positions and normalized (w.r.t. size and illumination)
prior to matching.
Similar to ﬁngerprint features, facial features can be divided into three levels [39]. Level 1
features consist of gross facial characteristics such as general geometry of the face. These coarse
features are usually the global face features. They can be easily obtained even from low resolution
images, and they can be used, for example, to quickly distinguish between an elongated and a
round face. Level 2 features consist of more localized face characteristics. Some examples of level
2 features include the structure of facial components (e.g. mouth), the spatial relationship between
facial components, etc. As in ﬁngerprints, level 2 features are the most important features for face
recognition. Level 3 features consist of the ﬁnest details of the face, which include scars, moles,
freckles, etc. [5]. These different levels of facial features are illustrated in Fig. 1.11.

1.2.2.2 Matching
There are three main approaches to match face images: (i) appearance-based, (ii) model-based and
(iii) texture-based approaches [5].

22

(a)

(b)

(c)

Figure 1.11: Examples of facial features at the three levels: (a) facial geometry (Level 1), (b) facial
landmarks, Gabor ﬁlter, local binary pattern (LBP), and shape (Level 2) and (c) moles (Level 3)
[5].
Appearance-based techniques refer to mapping a face image into a low dimensional sub-space.
The set of representative vectors is learned based on a training set of face images. Examples of
approaches in this category includes Principal Component Analysis (PCA) [40] and Linear Discriminant Analysis (LDA) [41]. In both of these approaches, a new face image can be represented
in terms of the learned basis vectors (as a weighted sum of the basis vectors). The weights associated with a test image can be compared to the weights of reference images in a database by using
the Euclidean distance, a measure of the dissimilarity between the two face images. The difference
between PCA and LDA representation is that LDA incorporates class information in the training
stage (supervised technique) while PCA does not (unsupervised technique). PCA projects the data
so that the overall variance is maximized while the LDA projects the data so that the ratio of the
inter-class variance to the intra-class variance is minimized.
Model-based approaches refer to the the use of face models. One example is graph matching,
where the face is represented based on a model graph. A model graph, in which ﬁducial points
(landmarks) of the face are associated with the nodes in the graph (see Fig. 1.12), is ﬁtted to a
face to generate a representation of that face. The model graph contains several local descriptors
(bunch) at each ﬁducial point to account for the variations in the local neighborhoods in the face
image. At each ﬁducial point of a query face image, a local descriptor is extracted and compared to

23

the descriptors in the stored model; the descriptor of the ﬁducial point in the query image is chosen
as the most similar descriptor in the bunch.

Figure 1.12: An illustration of a graph model ﬁtted to a face [5].
Texture-based approaches use local features such as Local Binary Patterns (LBP) and Scale
Invariant Feature Transform (SIFT). Both these local descriptors can be extracted at pre-speciﬁed
points in the image (ﬁxed if we assume the faces are roughly pre-aligned using the eyes) and feature vectors can be generated for both descriptors. The feature vectors can then be compared to
generate a similarity or dissimilarity. SIFT is a histogram of gradient orientations in a neighborhood, whereas LBP is a representation of the relationship among the intensities of neighboring
pixels. Fig. 1.13 shows schematic diagrams of SIFT and LBP features.
In the last decade, advances in face recognition technology have been reported by several third
party evaluations, for example, Face Recognition Grand Challenge (FRGC) [42] and Multiple Biometrics Evaluation (MBE) [43]. In FRGC, the performances of several face recognition systems
were measured in the veriﬁcation mode. The best veriﬁcation rate at a ﬁxed false acceptance rate
(FAR) of 0.001 was 99% for frontal face images in a controlled environment (studio lighting).
However, the veriﬁcation rate dropped to ∼ 80% at the same FAR of 0.001 when frontal face
images from uncontrolled environment were used as queries. MBE evaluated face recognition
systems in an identiﬁcation mode. The results from this evaluation showed a rank-1 accuracy of
24

(a) Scale Invariant Feature Transform

(b) Local Binary Patterns
Figure 1.13: Schematic diagram of (a) SIFT [6] and (b) LBP [5] descriptor construction.
92% on a background database (gallery) of 1.6 million criminal records, and a rank-50 accuracy
of 97%.

1.2.3 Iris Representation and Matching
The human iris is an annular shaped part of the eye that controls the amount of light entering the
eye through the pupil [5]. The iris texture pattern is formed and stable after the eighth month
of the gestational period. It is commonly believed that this pattern is mostly determined by the
gestational environment and not the genetic factors, thus the iris pattern of left and right eyes of
the same person are different. It should be noted that the visual pattern of a human iris includes
both color and texture. However, iris color has very limited discriminating power for recognition.
So gray-level iris images, captured under near infrared illumination, are used to record iris texture
pattern for person identiﬁcation. Fig. 1.14 shows two examples of irises captured using nearinfrared lighting.

25

(a)

(b)

Figure 1.14: Examples of two irises of two different subjects from the CASIA twins database. The
image on the right (b) is not of very high quality since the iris is partially occluded by hair.
In the next section, we present a brief overview of iris features and matching1.

1.2.3.1 Representation
The ﬁrst method to match two irises was designed and presented by John Daugman [44]. This
method is well known and many of the commercial iris recogniton systems still use its algorithms
[5].
The ﬁrst step in the extraction of iris features is the segmentation of the iris. Segmentation
aims to detect the inner and outer boundaries of the iris. These boundaries are usually represented
as circles or ellipses. The segmentation also generates a mask of the valid regions of the iris. The
second step is the normalization of the segmented iris region. This step is accomplished by using a
transformation that maps the cartesian coordinates of the iris to a pseudo-polar coordinate system.
Each concentric region of the iris is mapped to a row in a rectangular image of the iris. In other
words, the distance of a point in the iris image to the center of the iris is associated with a row in
the rectangular representation, and the angles formed with the horizontal axis are associated with
the columns. This mapping is also called “unwrapping of the iris".This normalization process is
necessary to account for the variations in the size of the iris texture. For example, if a pupil is
1 For a

more detailed description, please refer to Chapter 4 of [5]

26

dilated, the captured texture area will be small, while if the pupil is contracted, the captured texture
area will be large.
After the normalization step, two dimensional Gabor wavelets are convolved with the normalized image. Then, a demodulation is performed by taking the sign of the real and imaginary parts
of the convolution output, thus yielding a two-bit representation for each point in the normalized
image. This two-bit representation is basically a representation of which quadrant the phase falls
into. The size of the normalized image is usually 1024 bits, so the resulting binary output is a
vector of size 2048 and this binary vector is called the iris code.

1.2.3.2 Matching
Two iris codes can be compared by using the Hamming Distance [5]. Basically, each valid bit in
the iris code is compared to its corresponding valid bit in the other iris, and the number of bits
that differ normalized by the total number of valid bits in both codes is considered as the distance
between the two iris codes. Fig. 1.15 shows an example of an iris recognition system.

Figure 1.15: An overview of an iris recognition system [5].

27

NIST has evaluated 95 iris recognition algorithms in the identiﬁcation scenario; iris images
from over 2 million subjects were used [21]. When only one eye was used, the false negative error
rates (the percentage of cases in which the true mate was not identiﬁed) are at 1.5% or higher;
when both eyes were considered, this number dropped to 0.7%. These error rates are extremely
low and are mostly due to poor quality images, abnormal irises, ground truth errors, etc.. When
a threshold was set to produce no more than 25 false matches (false positive errors) in every 1013
comparisons, the false positive rate was still low (below 2.5%).

1.3 Applications of Biometric Systems
The increasing concern with security and identity fraud has supported the growth in the use of
biometric recognition technology for applications other than forensic ones. Also, due to the great
advancements in technology, nowadays it is more reliable and much less expensive to incorporate biometric recognition into several applications, such as border control, ATM access, national
identiﬁcation, voting access, health care access, criminal and victim identiﬁcation, etc. Biometric
recognition can be used in small and large scale applications. For example, one can use ﬁngerprint
to protect one’s laptop or other device; on the other hand, in the Unique Identiﬁcation Authority of
India (UIDAI) program, biometric data (ﬁngerprint, face and iris) were collected from hundreds of
millions of subjects and this number keeps growing [45].
Other applications of biometric systems include:
• Health care access
According to [46], estimates indicate that, in 2006, approximately 20% of the cost incurred
by health insurance providers involves some type of fraud. Also, the cost of health insurance
could go down as much as 3.5% if the persons who have the health insurance and are authorized to receive the health insurance beneﬁts were the only ones to actually beneﬁt from
it. To avoid fraud by an unauthorized person receiving health care under another person’s
health insurance, automated ﬁngerprint veriﬁcation is used at the health care providers.

28

• Elections
In Brazil, every adult must vote in all elections (federal, state and city levels). In 2000, the
election was fully automated. The next step was to include ﬁngerprint technology to efﬁciently and correctly verify whether the person who was trying to vote was the actual person
registered to vote. It is expected that all the ballot boxes will have biometric recognition
systems by 2020 [47].
• Finger vein recognition
A ﬁnger vein recognition system was commercialized by Fujitsu Ltd. to be placed in ATMs
in Japan. The bank client still has an ATM card, but instead of inputting a personal identiﬁcation number (PIN), the client will place his/her hand over the ﬁnger vein sensor [48]; the
collected data is compared to the data stored in the card, and the access is granted based on
the veriﬁcation result. The entire veriﬁcation process (insert the ATM card and verify ﬁnger
veins) takes about the same time as if someone inserts the ATM card and input a PIN.

1.4 Contributions
The contributions of this dissertation are as follows:
1. Biometric-based identiﬁcation of identical twins
• An analysis of the discriminative ability of the three most common biometric modalities (face, ﬁngerprint and iris) of identical twins is conducted, including results of
combining different biometric traits.
• An analysis of using multiple biometrics to quantitatively determine whether two subjects are identical twins is performed.
2. Latent ﬁngerprint matching

29

• To improve latent ﬁngerprint matching performance by the enhancement of the latent
image.
• To improve latent ﬁngerprint matching performance by using a descriptor-based Hough
Transform.
3. Latent ﬁngerprint indexing
• An indexing approach for latent ﬁngerprints that reduces the total search computational
time by a factor of 5 while still maintaining the matching accuracy.

1.4.1 Distinguishing Identical Twins
One of the challenges in biometric identiﬁcation is related to distinguishing twins, especially
identical twins. There are two types of twins: monozygotic (or identical) and dizygotic (or nonidentical) [49]. Monozygotic twins are a result of a single fertilized egg that splits into two cells,
each one giving origin to one individual. Dizygotic twins are a result of two different fertilized
eggs. Monozygotic twins have the same deoxyribonucleic acid (DNA), thus their genotypic features (features inﬂuenced by the genetic material) are the same since they share the same genetic
material, while some phenotypic features (features inﬂuenced by the environment) may be different. Therefore, identical twins are more likely to have biometric features with somewhat higher
degree of similarity compared to non-identical twins or unrelated persons. Because of this, an
analysis of the ability of a biometric system to distinguish identical twins is essential. In recent
years, the Federal Bureau of Investigation (FBI) has shown an interest in further investigating the
biometrics of twins by supporting the collection of twins’ biometric data. A pair of face images of
identical twins is shown in Fig. 1.16.
In this dissertation, we study the performance of biometric systems in the presence of identical twins for the three most commonly used biometric modalities, namely ﬁngerprint, face and
iris, as well as an analysis of the performance of multimodal systems, which use a combination
of biometric traits to make a decision. Face recognition of identical twins is shown to be a chal-

30

(a)
Figure 1.16: A pair of identical twins from CASIA Twins database Version 2.
lenge to current face recognition systems, while ﬁngerprint and iris mathcing results for identical
twins show performance comparable to unrelated persons. The fusion of different instances from
the same modality (e.g., multiple ﬁngers and left and right iris) yields the best performance for
distinguishing identical twins as well as for unrelated persons. We also provide an analysis of how
multiple biometrics can be used to determine whether two subjects are identical twins. The similarities between the biometrics of identical twins can be very useful to advance the state-of-the-art
biometric systems.

1.4.2 Latent Fingerprint Matching
Fingerprints have been successfully used for person recognition for over a century. Latent ﬁngerprints constitute a valuable source of evidence in law enforcement agencies to help solve crimes.
However, a majority of the processing (e.g. feature marking) involving latent prints is done manually by forensic (latent) experts. While progress has been made to automate this process, according
to the latest evaluations by NIST [50], the state-of-the-art technology for fully automated latent
matching (lights out processing) is not mature. The current practice involves latent examiners to
manually mark the features in the latent and then input it to the system for automatic matching
with reference prints. The matcher returns a list of candidates that are manually checked by a
latent examiner who then makes the ﬁnal identiﬁcation. The practice of manually marking features in latents is labor intensive, but it is nevertheless currently more reliable than state-of-the-art

31

systems.
Latents are a particularly challenging to match because they are usually of poor quality, contain
a small friction ridge area, present large non-linear distortion, can be blurred or smudged, and
usually contain complex background noise. All these challenges make it very difﬁcult to reliably
extract features in latents; unreliable features lead to low matching performance. Given that the
minutiae are the most commonly used features in ﬁngerprint matching, latents contain signiﬁcantly
fewer number of minutiae than a reference ﬁngerprint. This makes rolled-to-rolled ﬁngerprint
matching algorithms unsuitable for latent matching. In this dissertation, our efforts are focused
on improving latent ﬁngerprint matching performance, given reliable manually marked features in
latents.
The ﬁrst method consists of enhancing the latent image and fusing the matching score of the
enhanced latents with the score based on manually marked minutiae. This ﬁrst method outperforms
a commercial ﬁngerprint matcher on the publicly available latent database NIST SD 27.
The second method consists of developing a new latent ﬁngerprint matcher based on manually marked minutiae that uses minutiae and orientation ﬁeld information in the matching. The
proposed matching algorithm outperforms three ﬁngerprint matching algorithms on two different
latent ﬁngerprint databases (NIST SD 27 and WVU Latent databases).

1.4.3 Latent Fingerprint Indexing
The identiﬁcation of a person requires that his/her ﬁngerprint be compared to all the ﬁngerprints in
the database [1]. If the database is very large, matching a query ﬁngerprint to the entire database
might become computationally unfeasible. Thus, strategies to quickly ﬁlter out a large portion of
the database without degradation in the matching performance are very useful.
Given manually marked features in latents, our indexing approach consists of using singular
points, minutiae, orientation ﬁeld and frequency to signiﬁcantly reduce the background database
size. By applying our indexing approach, we are able to ﬁlter out 80% of the reference database
while maintaining the latent matching accuracy.

32

1.5 Dissertation Organization
This dissertation is organized as follows. In Chapter 2, we study the problem of distinguishing
identical twins based on unimodal (face, ﬁngerprint and iris trait alone) and multimodal (based on
multiple traits) biometric systems. also analyze the possibility of using multibiometrics to provide
a quantitative measure to determine whether two subjects enrolled in the database are identical
twins. In Chapter 3, two different approaches to improve latent ﬁngerprint matching performance
using minimal amount of manually marked features in latents are presented and discussed. The
proposed techniques perform better than some of the commercial ﬁngerprint matchers on two different latent ﬁngerprint databases. In Chapter 4, an indexing approach is presented to ﬁlter out
a large portion of the background database, thus greatly speeding up the search while maintaining comparable matching accuracy. Finally, summary of our research and some ideas for future
research are presented in Chapter 5.

33

CHAPTER 2
BIOMETRIC TRAITS OF IDENTICAL
TWINS
2.1 Introduction
A twin is “one of two offspring produced in the same pregnancy” [49]. Twins can be categorized
into two types with respect to the number of eggs fertilized: monozygotic (or identical) and dizygotic (or non-identical) [49]. Monozygotic twins are a result of a single fertilized egg that splits
into two cells, each one giving origin to one individual; thus, monozygotic twins have the same
deoxyribonucleic acid (DNA). Dizygotic twins are a result of two different fertilized eggs, and
therefore they do not have the same DNA. Face images of a pair of identical twins and of a pair
of non-identical twins are shown in Fig. 2.1. Three offspring produced in the same pregnancy
are called triplets, four are called quadruplets, and so on. In this dissertation we are primarily
interested in the ability of biometric traits to distinguish identical twins.
In the year 2009, approximately 153 of every 100,000 births in the United States were triplets
or more [51]. This birth rate increased 400% during the 1980s and 1990s [52], and it has been
ﬂuctuating since 1999, with a slight downward trend [51]. However, the rate in the year 2009 represented a signiﬁcant increase from 2008 (4%) compared to the ﬂuctuations in this rate in previous

34

(a)

(b)

Figure 2.1: Face images of (a) a pair of identical (monozygotic) twins and (b) a pair of nonidentical (dizygotic) twins from the University of Notre Dame ND-Twins database [7].
years [51]. The rapid increase of multiple births in the 1980s and 1990s has been associated with
the increase in the use of fertility therapies and “the older age at childbearing", the latter because
“women in their 30s are more likely than younger women to conceive multiples spontaneously
[52]."
The twin birth rate (33.2 per 1000 births in the United States in 2009) increased at an average
rate of 3% per year between 1990 and 2004 [51]. Then, the average rate of increase slowed to
less than 1% per year over the period 2005-2009 and increased to 2% from 2008 to 2009 [51].
Although the average rate of increase has decreased, the overall twin birth rate in the United States
has increased by 76% since 1980 [51]. The number of dizygotic twin births varies depending on
the ethnic group [53], while the number of monozygotic twin births is believed to be constant
worldwide. According to Nora et al. [54], the frequency of identical twins ranges from 0.35% to
0.4% among all births. This, in turn, has created a requirement for biometric identiﬁcation systems
to accurately determine the identity of a person who has an identical twin.
Accurately distinguishing identical twins has important legal ramiﬁcations. In 1985, a woman
was accused of making fraudulent money transfers from several banks by simply making telephone
calls. She was acquitted of those charges after claiming that her sister — who was her identical
twin but disappeared before the trial — committed the fraud. In the end, both twins were convicted.
35

Based on an analysis of their voice samples, a speech scientist and voice examiner testiﬁed that
indeed both women had placed the fraudulent telephone calls [55]. In this case, an ability to
distinguish identical twins based on the voice biometric was essential to the convictions.
In another incident, a rape suspect won two mistrials because there was no way of determining
whether the key DNA evidence came from him or his identical twin. However, in the third trial held
in 2006, the accused was convicted after the prosecutors provided evidence that he had committed
a series of sexual assaults over time and had attempted sexual assaults with characteristics similar
to the rape for which he was being tried [56]. A failure to distinguish the twins was the reason for
the charges being dropped in the ﬁrst two mistrials.
The similarity in facial appearance of identical twins may also give them a greater incentive to
commit fraud than an average person. Imagine a scenario where you have an identical twin sibling,
who is covered by health insurance while you are not. If you fall sick and need medical treatment,
you need health insurance. While the health care providers need to establish your identity using a
photo identiﬁcation, you could use your twin’s health insurance without being caught. But, suppose
the health insurance company requires that you need to be identiﬁed using ﬁngerprints (this is the
case for some of the health insurance companies in Brazil) [57]; if the ﬁngerprint recognition
system can distinguish identical twin ﬁngerprints, then you will not be able to get health care using
your twin’s photo ID and insurance policy.
The above facts and scenario indicate that the ability of biometric systems to identify identical
twins is necessary. Since monozygotic twins have the same DNA, they cannot be distinguished
using DNA alone [58]. Thus, it is necessary to use other forms of identiﬁcation for monozygotic
twins.
Recognition using biometric traits is now a well accepted and proven method. A biometric system relies on the distinctiveness of the biometric characteristics to perform the recognition. While
many biometric techniques are extremely accurate, some variations in sensing data, noise, etc. can
cause the matching performance to drop signiﬁcantly. One could argue that it is more difﬁcult to
discriminate identical twins than unrelated persons because of their genetic similarity. Although

36

identical twins cannot be distinguished from each other using DNA, some of the biometric modalities, such as ﬁngerprints, iris, and palmprints, can still be used to distinguish them [59]. Some
studies have shown that face and voice can also be used to distinguish identical twins [60, 61].
Due to the difﬁculty in obtaining a large biometric database of identical twins, most experiments
reported in the literature are based on small databases.
This chapter explores the ability of unimodal and multimodal biometric systems to distinguish
identical twins. In Section 2.2 of this chapter, we present studies on biometric twin data reported in
the literature. In Section 2.3, we present the underlying characteristics of the three biometric traits
that will be used in our experiments, namely ﬁngerprint, face, and iris. In Section 2.4, we analyze
the experimental results on matching individual biometric traits as well as various combinations of
modalities. In Section 2.5, we end this chapter by presenting our conclusions.

2.2 Related Work
In order to design a robust and efﬁcient biometric system, the system must be able to handle a
variety of situations like noisy data, limitations of the sensors, environmental conditions, and the
presence of identical twins. Due to the similarity of their biometric characteristics, identical twins
are more likely to pose a challenge to a biometric system. Therefore, it is important to address this
problem when designing a biometric system. Table 2.1 summarizes the studies on discrimination
of identical twins that have been reported in the literature; these studies are also discussed below.
Daugman and Downing [62] assessed the distinctiveness of iris pattern as a biometric identiﬁer.
The authors compared unrelated irises (irises from different persons), and genetically identical
irises (irises that came from the same DNA), for example pair of irises of the same person or the
irises of identical twins. They observed that the iris patterns of genetically identical eyes were
as uncorrelated as the patterns of unrelated eyes. For example, the similarity of the six pairwise
comparisons they performed between identical twins showed the same mean as for eyes that were
not genetically related.
1 It is

not clear in this work how many pairs of identical twins were used.
37

Table 2.1: Summary of studies on the biometrics of identical twins. Sets can include identical twin
pairs as well as non-identical twin pairs.

STUDY

YEAR

BIOMETRIC TRAIT

DATABASE SIZE

Daugman and Downing [62]

2001

Iris

1 set

Jain et al. [58]

2002

Fingerprints

94 sets

Kodate et al. [60]

2002

Face

10 sets

Han et al. [63]

2004

Fingerprints

66 sets

Patil and Basu [64]

2004

Voice

12 sets

Bronstein et al. [65]

2005

Face (3D)

1 set

Kong et al. [59]

2006

Palmprints

53 sets

Srihari et al. [66]

2008

Fingerprints

298 sets of twins

Ariyaeeinia et al. [61]

2008

Speech

49 sets

Sun et al. [67]

2010

Face, Fingerprints, Iris

51 sets

Hollingsworth et al. [68]

2011

Iris

76 sets

Phillips et al. [7]

2011

Face

126 sets

Pruitt et al. [69]

2011

Face

126 sets

Biswas et al. [70]

2011

Face

186 subjects1

Klare et al. [71]

2011

Face

126 sets

Vijayan et al. [72]

2011

Face (3D)

107 sets

Srinivas et al. [73]

2012

Face

126 sets

Tao et al. [74]

2012

Fingerprints

83 sets

In a study of identical twins’ palmprints, Kong et al. [59] used 1,028 palmprint images from
53 pairs of identical twins. They performed two different twin matches. In the ﬁrst experiment,
they matched the palmprints of identical twins, which they called real twin match. In the second
experiment, they matched the left and right palmprints of the same person, which they called
virtual twin match. Note that in both the experiments the palmprints shared the same genetic
information. The matching algorithm was based on the angular differences of orientation ﬁelds (of
the palmprint ridge pattern) in the two palmprints being compared. The authors observed that while
these genetically equivalent palmprints have correlated features, they can still be distinguished
based on features extracted from palmprint patterns.
38

The above observation concerning palmprints is also true for ﬁngerprints, as a study based on
94 pairs of identical twin ﬁngerprints showed [58]. Jain et al. showed that ﬁngerprint veriﬁcation
systems can be used to distinguish identical twins, even though their ﬁngerprints are generally more
correlated than ﬁngerprints coming from two genetically unrelated persons. As an example, at a
False Rejection Rate (FRR) of 1.05%, the twin-twin matching had a False Acceptance Rate (FAR)
of 8.51%, while the twin-nontwin matching had a FAR of 2.20%. For another threshold value (FRR
of 3.49%), the twin-twin FAR dropped to 1.06% and the twin-nontwin FAR dropped to 0.29%. In
another analysis of ﬁngerprints from 66 pairs of twins [63], it was also concluded that ﬁngerprints
can be used to identify identical twins with an insigniﬁcant drop in the matching performance:
the Equal Error Rate (EER) reportedly increased by only 1-2% compared to nontwin impostor
matching. The authors also extended their studies to assess the similarities among families of
nontwins (52 families of four persons — parents and two children). They observed that the largest
similarity occurred between identical twins, followed by between two siblings, between parents
and their children, and between unrelated persons.
Srihari et al. [66] analyzed the similarity between twins’ ﬁngerprints in a study using ﬁngerprint images from 298 pairs of twins (74 fraternal twins and 224 identical twins). The authors
analyzed this similarity based on Level 1 and Level 2 features (that is, pattern of the ridge ﬂow,
and minutiae, respectively). With the level 1 features, they observed that twin ﬁngers with the same
label occurred approximately 55% of the time and approximately 32% for non-twins, which means
twins’ ﬁngers are much more likely to have the same pattern type than non-twins’ ﬁngers. With
the level 2 features, they concluded that the similarity between twin ﬁngers is higher than between
two arbitrary ﬁngers (with identical twin ﬁngers similarity being not signiﬁcantly different from
fraternal one), but twins can still be distinguished using ﬁngerprints.
Although it is believed that it is difﬁcult for face and voice, along with hand geometry, to
distinguish between identical twins [58], some researchers have obtained encouraging results using
face and voice to distinguish monozygotic twins [60, 65, 61]. A veriﬁcation experiment with 10
sets of identical twins was performed using a 2D face recognition system; the experiment consisted

39

of enrolling one of the twins, and asking the other to try to log into the system. On this small
database, the rejection threshold was always satisﬁed, leading to a rejection of all the impostor
twins [60].
Another face recognition experiment for twins was based on 3D face images [65]. The recognition task was to distinguish between two identical twins; the authors tested three different algorithms: a 2D algorithm based on eigenfaces, a 3D algorithm based on rigid surface, and another 3D
algorithm based on canonical forms. The test consisted of enrolling one of the twins and matching
the enrolled subject to the other twin. All the other subjects served as impostors. For the ﬁrst
algorithm (2D eigenfaces), the rank-1 accuracy was 70.59% when enrolling the ﬁrst twin and 75%
when enrolling the other twin. In the second experiment, the rank-1 accuracy was 82.36% and
100.0%, respectively, while the third algorithm achieved perfect matching performance. According to the Web site Digital World Tokyo [75], the Japanese company Sagawa Advance has invented
an infra-red based face recognition technology that is able to distinguish identical twins.
Ariyaeeinia et al. [61] performed recognition experiments using voice data from 49 pairs of
identical twins. The authors performed basically two different experiments: a general experiment,
in which any two persons in the dataset were considered impostors, and the twin experiment,
in which the impostor tests consisted of comparing a person and his/her twin. The Equal Error
Rate reported was 1.0% for the twin experiment using short interval voice pattern (each person
saying his/her date of birth), and 0.5% for the general conﬁguration. Other authors have tried to
distinguish identical twins based on voice in a multilingual environment [64]. Using a database
of 12 twins, Patil and Basu reported the highest success rate as being 100% for a particular size
(60 seconds) of the training speech, and particular size (in seconds) of the test speech. They also
observed that the majority of errors were due to matching the actual speaker with his/her twin
brother/sister.
However, in most of the previous studies, the identical twin biometric database is very small and
an in-house biometric matcher instead of COTS matchers were used. This affects the reliability
of conclusions. Furthermore, no previous study was conducted to compare the performance of

40

different biometric traits and fuse multiple biometric traits of identical twins.
A number of studies and methods related to biometrics of twins has been published since our
work in multibiometrics of twins appeared [67]. They are summarized below.
Hollingsworth et al. [68] suggested that genetically identical irises (from identical twins or
left and right irises of a person) might be more similar than genetically unrelated irises. They
conducted experiments on iris texture similarity in which humans viewed pairs of iris images.
There were basically two experiments: in the ﬁrst experiment, volunteers had to provide their
opinion on whether the pair of irises they were viewing was from the same person (left and right)
or from different persons; in the second experiment, volunteers had to provide their opinion on
whether the pair of irises they were viewing was from identical twin pair or from unrelated persons.
The results showed that humans were successful in these two tasks in more than 80% of the cases.
These results imply that, although iris biometrics can distinguish between identical twins, some
similarity is still present. It should be noted that humans are not necessarily good at matching iris
textures.
Tao et al. [74] analyzed the performance of two state-of-the-art ﬁngerprint matchers on identical twins by using a database of 83 identical twin pairs, 4 ﬁngers per person, and 6 impressions
per ﬁnger. Their conclusion that ﬁngerprint matchers can distinguish between identical twins with
a small drop in performance agrees with previous studies. They also analyzed and compared the
probability of ﬁngerprints from identical twins having the same pattern. This probability was found
(as in [58]) to be much higher compared to the probability of ﬁngerprints from unrelated persons
having the same pattern – 0.7440 vs. 0.3215. Another conclusion of this study is that all ﬁngers
have the same probability of being of the same ﬁngerprint type in identical twins.
Phillips et al. [7] conducted matching experiments using face images of 126 pairs of identical
twins. They used three state-of-the-art face recognition systems, and the experiments were conducted on face images captured under two different lighting conditions, indoor (or studio lighting)
and outdoor (or ambient lighting), and with two facial expressions, neutral and smile. Their experiments led to the conclusion that the changes in ambient lighting and expression largely affect the

41

performance of the systems; the studio to studio lighting matching yielded the best performance,
while the ambient light conditions matching yielded the worst performance. As an example, at a
False Acceptance Rate of 0.01, the veriﬁcation rate of studio to studio lighting with neutral expression is about 0.85, while for the ambient lighting also with neutral expression is about 0.05.
Their experiments also showed that, if we ﬁx the lighting to studio conditions, the Equal Error
Rate (EER) in the case of expression changes is signiﬁcantly higher for the three algorithms. In
summary, recognition of identical twins in studio lighting with neutral expression is promising,
while when the lighting or the expression change, the performance is drastically reduced.
Pruitt et al. [69] conducted experiments similar to [7]. They used the same database used in [7],
and also analyzed the inﬂuence of different expressions and lighting using four face recognition
systems, and reaching the same conclusions as in [7]. In addition, they analyzed the inﬂuence of
the presence of eyeglasses, and they found that the performance is about the same.
Biswas et al. [70] also concluded that uncontrolled environment affects the ability to distinguish between identical twins, but their experiments analyzed the human ability. Their experiments
were conducted using face images from a database of 186 subjects; 23 volunteers were asked to
look at pairs of face images for a limited time and to decide whether the pair of face images came
from the same person or from identical twins. Humans could correctly distinguish between identical twins with an average accuracy of almost 79%. Another interesting result was that humans
used moles/scars/freckles as the most important features in their correct decisions.
Park and Jain [76] applied their facial marks detection and matching algorithm to a few identical twins face images, and showed that facial marks helped in distinguishing between twins. More
recently, Srinivas et al. [73] presented an analysis and algorithm to distinguish between identical
twins by using facial marks. In their work, they analyzed and matched manually marked and automatically extracted facial marks. Their results showed that facial marks can help in distinguishing
between identical twins; however, it is more difﬁcult to distinguish identical twins than distinguishing between unrelated persons. Their results also suggested a correlation in the distribution
of facial marks in twins.

42

Klare et al. [71] analyzed the effect of different features in the identical twin recognition based
on face images in the database used in [7]. They analyzed the importance of different components
of the face, the improvement using facial marks, and the differences in performance when the systems were trained on images of identical twins compared to unrelated persons. Their conclusions
include: (i) all facial components seem to have the same importance in identical twins matching
scenario as in the standard matching one, (ii) the use of facial marks improves the discriminability
of identical twins without reducing the accuracy on unrelated persons, and (iii) regarding discriminative learning, it appears that training on twins does not help the overall system, but this might
be only because the number of twins available for training is not sufﬁciently large.
Vijayan et al. [72] presented the results of applying four state-of-the-art 3D face recognition
algorithms to an identical twins dataset (3D Twins Expression Challenge) containing 3D scans of
neutral and smiling expressions of 107 pairs of identical twins. They conducted experiments on
different cases based on the expressions of the probe and gallery scans of each of the twins. In the
two scenarios in which the gallery had one expression and the probe had a different expression, the
best performing algorithm presented an Equal Error Rate (EER) of 0.2% and 0.5% and a rank-1
identiﬁcation rate of 98.1% for both scenarios. In the scenarios with uncontrolled expressions (one
twin in the gallery has the same expression as the other twin in the probe, and vice versa), the EER
was 0.8% for both the cases, and the rank-1 identiﬁcation rate was 91.6% and 93.5%, respectively.
Their results argue that 3D face recognition remains an open problem if different facial expressions
are present.

2.3 Multibiometrics
A multibiometric system uses multiple sources of biometric information in order to recognize an
individual. For example, a multibiometric system may use ﬁngerprint and face, the left and right
iris, or a fusion of different ﬁngerprint matching algorithms to recognize a person. In the next
subsections, we focus our attention on the distinctiveness of ﬁngerprints, face, and iris for identical
twins. These three modalities will then be used in our multibiometric experiments.
43

2.3.1 Fingerprint Discriminability
A ﬁngerprint is the impression of the friction skin on a ﬁnger. The individual characteristics of
friction ridge skin are determined during fetal development. Their formation is similar to the
formation of blood vessels or capillaries during the growth of the fetus in the uterus. The ﬁngerprint
formation starts at approximately 6 or 7 weeks of gestational age and it is due to the ﬂow of
amniotic ﬂuids in a micro-environment [22]. A minor change in this ﬂow and in the position
of the fetus in the uterus cause the minute skin structures around palm, ﬁnger tips and soles to
differentiate. Friction ridge skin can be distinguished from the skin of the rest of the body due to
a variety of factors, such as the presence of raised ridges, increased sensory abilities, absence of
hair and sebaceous glands, and a thicker and more complex epidermis. Friction ridges are useful
for grasping and gripping objects, which explains their presence in our hands and feet.
As mentioned in Chapter 1, a ﬁngerprint pattern has some details that are present in each individual ﬁngerprint, like ridge endings, the point where a ridge ends abruptly, or a ridge bifurcation
or trifurcation, where the ridges are divided into different branches. However, collectively, these
details (minutiae) are supposedly different in every ﬁngerprint, even in prints of identical twins,
since a very small difference in micro-environment is sufﬁcient to change the process of cell formation, causing minutiae points to be different. As a result, ﬁngerprint is considered very reliable
in terms of biometric identiﬁcation because of its distinctiveness. Another property of ﬁngerprint
that makes them useful in biometric identiﬁcation is that ﬁngerprints do not change signiﬁcantly
over time, an essential characteristic of a biometric modality since a biometric system is typically
meant to be used to identify a person over a long period of time.

2.3.2 Face Discriminability
Face is composed of the skull characteristics and the musculature and associated soft tissue. To
study the variation among human faces, it is necessary to study these structures. The facial skeleton
serves as the bony framework for the mimetic musculature. Since these muscles are stretched
across the facial skeleton like a mask, the variation in facial appearance is caused mostly because
44

of the form of the facial bones. A majority of individuals can be divided in three categories in
terms of facial appearance: (i) a long, narrow head, (ii) a proportional length to width head and
(iii) a short, wide head; these differences in facial shapes are due to the form of the cranial base
[22].
Facial form is inﬂuenced by gender. Males are more likely to have larger faces because of their
usually larger bodies and their need for more air in order to support larger muscles and visceras,
causing them to also have a larger nose. Males also usually have a more protrusive forehead.
Besides the gender inﬂuences in the face of an individual, a face also changes when a person ages.
Infant faces tend to be wide and short because of the development of the brain; over time, the
face develops and this wide and short face of the baby tends to change. Other effects of aging are
dehydration of the dermis and reabsorption of subcutaneous fat deposits, which result in a decrease
in the facial volume and wrinkling [22].
The muscles may vary in their presence (not everyone has all the muscles that could be in a
face), form, location, and control. These factors inﬂuence the kind of facial movement that an
individual can create. Furthermore, the facial movements of an individual change his/her face as
he/she ages. With aging, the elasticity of the skin decreases and the face then is marked with the
expressions that occur frequently, becoming relatively permanent features. This fact may explain
why identical twin faces are more likely to be distinguished as they are older than when they are
infants [22].
There are a number of factors that inﬂuence the performance of a facial recognition system. Besides those already cited above, there are differences in pose, illumination, expression, occlusion,
accessories like glasses, weight changes, hair style changes, etc. All these variations make facial
recognition systems not as accurate as some other biometrics, like ﬁngerprint and iris. Identical
twins present a particularly difﬁcult situation for face recognition systems, since they are usually
extremely similar in facial appearance.

45

2.3.3 Iris Discriminability
The iris is an annular shaped part of the eye between the pupil and the sclera (see Fig. 2.2) that
regulates the amount of light entering the eye through the pupil. The iris is physically small in size
(about 11 mm) but well-designed optical systems can magnify a human iris into a high-resolution
image that is 200 to 300 pixels in diameter. There are many minute features such as freckles,
coronas, stripes, furrows and crypts, etc. randomly distributed in the iris region, which constitute
the unique iris texture for each eye.

Figure 2.2: A diagram of the human eye 2 .
The iris texture pattern is formed and becomes stable after the eighth month of gestation. It is
commonly believed that the formation of iris pattern is determined by the gestation environment,
i.e. iris is a phenotypic biometric trait [62]. So even identical twins can be discriminated using
suitable iris features; even the irises of left and right eyes of the same person are different.
2 Figure from http://www.nei.nih.gov/photo/.

46

Random processes involved in iris development determine the irregular shape and random spatial layout of micro anatomical structures (MAS). The MAS in the iris surface may exhibit different
reﬂectance properties in the near infrared light, leading to sharp intensity variations across iris image region. We can regard each iris region as a piece of 2D surface in a 3D coordinate system.
So the surface shape from valley to peak resembles an odd Gabor ﬁlter and the shape from valley
to peak then valley resembles an even Gabor ﬁlter, and vice versa. Daugman [62] proposed to
match each iris region with even or odd Gabor ﬁlters and then encode positive correlation as 1 and
negative correlation as 0. Thereby, an iris image can be represented by 256 bytes, called the iris
code. The iris codes of twins (or non-twins) have only 50% chance of being matched because their
corresponding regions independently have an equal probability to be either 1 or 0 in iris feature
coding. In contrast, multiple iris images of the same eye have a much higher probability than 50%
to be matched in their iris codes, even though noise may perturb some parts of iris codes. A more
general explanation of the effectiveness of iris recognition for twins is based on ordinal measures
[77].

2.4 Experimental Results and Analysis
In this section, we ﬁrst describe the databases used in our experiments, followed by an analysis
of the ability of biometric systems in distinguishing between identical twins and their ability in
determining whether two subjects are identical twins based on their biometric traits, namely face,
ﬁngerprint and iris.

2.4.1 Databases
The ﬁrst version of the CASIA Multimodal Biometrics Database of Twins (CASIA-TwinsV1) was
collected in October 2, 2007 at the Beijing Chaoyang Park during the Fourth Annual Festival of
Beijing Twins Day. Figure 2.3a shows the kiosk where the biometric acquisition was performed
and Figure 2.3b shows the face acquisition device. This database includes face, iris, and ﬁngerprint

47

images from 92 pairs of twins and 2 sets of triplets. All the images were captured indoors (inside
a tent) on the same day (i. e., single data capture session).

(a)

(b)

Figure 2.3: The kiosk for biometric acquisition (a) and the face acquisition device (b).
Not all of the 94 sets of twins or triplets provided images for all the modalities. Since some of
our experiments involved combinations of units/modalities, we considered only those individuals
who have a complete set of images (face, two irises and four ﬁngerprints) in the database, and
whose twin’s images were also present in the database as a complete set. As a result, the total
number of subjects used in our experiments consisted of 134 subjects (66 families, including two
families of triplets – 51 pairs of identical twins and 15 pairs of non-identical twins). For all biometric modalities (four ﬁngers, two irises, and face), the number of genuine matches performed was
134, the number of identical twin impostor matches was 102, and the number of general impostor
matches was 17,720.
The second version of the CASIA Multimodal Biometrics Database of Twins (CASIA-TwinsV2)
was collected in 2009. This second version also includes face, iris, and ﬁngerprint images from 59
pairs of twins and 1 set of triplets. Not all of the 60 families provided images for all the modalities.
This version is more challenging than the ﬁrst one since the images were captured indoor and outdoor, and the face images contain variations such as lighting, pose and expression. By themselves,
these variations already pose a great challenge to face recognition.
Most of the twins are identical (or monozygotic twins), but some are non-identical (or dizygotic) twins. This information was not recorded, so we derived this information based on observing
48

whether the face images of a set of twins were very similar or not 3 . In the FBI’s collection of biometric data from twins, the information whether the twins were identical or not was revealed by
the subjects, i.e., the subjects were asked whether they were identical or non-identical twins. Both
methods are not perfect and subject to errors. For example, non-identical twins can have very
similar appearance which would cause them to be classiﬁed as identical twins based on facial similarity; or twins might report identical or non-identical twinning based on their guesses or similar
appearance rather than based on DNA testing. We divided our database into two groups, identical
twins and non-identical twins. Most of the subjects in the database are children, but there are some
adults as well. The subject age ranges from 5 to 65 (5 to 63), with the average age being 16.8 (16.2)
in the ﬁrst (second) version of the database. There are 12 families for which their biometrics were
collected in both versions. We used matching scores from iris and face images to ﬁnd the subjects
who were present in both collections.
In the following sections, we present more detailed descriptions of the two databases separated
by the biometric modality, the experiments performed, and experimental results on the two topics:
distinguishing identical twins and ﬁnding similarities between identical twins.

2.4.2 Distinguishing Identical Twins
2.4.2.1 Fingerprint
The ﬁngerprint images were captured using Symwave sw6888 [78], a sweep sensor, with a resolution of 500 dpi. This dataset contains images from four different ﬁngers and the number of images
per ﬁnger is not ﬁxed, but varies from 6 to 31. Because of this large variability and the poor quality
of many images, we selected one image per ﬁnger as the template and another image from the
same ﬁnger as the query image that contain the largest number of minutiae.
Figure 2.4 shows some ﬁngerprint images from the twin dataset. Figure 2.4a shows images of
the four ﬁngers 1, 2, 3, and 4 for the ﬁrst twin of an identical twin pair; Figure 2.4b shows the
3 The ideal

method to distinguish identical or non-identical twins should be based on DNA.

49

ﬁngerprints of the corresponding ﬁngers for the second twin of the pair. Figures 2.4c and 2.4d
follow the same scheme for a non-identical twin pair. The similarity between ridge ﬂow patterns
of corresponding ﬁngers of identical twins is evident. While all four pairs of corresponding ﬁngers
of identical twins have the same ﬁngerprint pattern type (left loop, left loop, right loop, right loop
for ﬁngers 1 to 4, respectively), only two pairs of corresponding ﬁngers of non-identical twins have
the same pattern type (left loop, right loop for ﬁngers 1 and 3, respectively).

(a)

(b)

(c)

(d)

Figure 2.4: Fingerprint images of ﬁngers 1, 2, 3, and 4 of the ﬁrst twin (a), and the four images
of the corresponding ﬁngers of the second twin in an identical twin pair (b); similarly, (c) and
(d) show ﬁngerprint images of a non-identical twin pair. Note the similarity in ridge ﬂow pattern
between identical twins. All four corresponding ﬁngers of identical twins in (a) and (b) have the
same pattern type. But for non-identical twins in (c) and (d), only two corresponding ﬁngers (no.
1 and 3) have the same pattern type.
A minutiae based commercial matcher, VeriFinger 4.2 [79], was used to obtain the matching
scores in the ﬁngerprint experiments. In our experience with Veriﬁnger 4.2 match scores, there are
no impostor scores greater than 300, so scores greater than 300 were set to this number, and then
the scores were normalized from [0, 300] to the range [0, 1].
Three different match score distributions were generated in order to analyze the results of the
experiments. The genuine distribution was obtained by matching the gallery image of one modality
50

of one individual to the probe image of the same modality and the same individual. Identical twin
impostor distribution was generated by matching each image of one person to his/her identical
twin. The general impostor distribution was generated by all the impostor matches except the
identical twin impostor matches described above. This scheme was used for all the experiments.
Figures 2.7a and 2.7b show the genuine, identical twin, and the general impostor distributions
of ﬁnger no. 3, which had the worst performance among the four ﬁngers, and ﬁnger no. 4, the
best performing ﬁnger, respectively. We can see that, although identical twin impostor distribution
of ﬁnger no. 3 is similar to the general impostor distribution of this same ﬁnger, there are some
peaks in the right tail of the identical twin impostor distribution that differentiates them. This is an
indicator of the larger similarity between ﬁngerprints of identical twins compared to the similarity
between ﬁngerprints of unrelated persons. This is also the case for ﬁnger no. 4. The KolmogorovSmirnov 2-sample test [80] indicates that the two samples (identical twin and general impostors)
do not come from the same population with a signiﬁcance of 0.05.
Another indicator of this large similarity between identical twin ﬁngerprints is the number of
matched minutiae. For example for ﬁnger no. 4, the mean of the number of matched minutiae for
genuine pairs is 22.00 (± 9.50), for identical twin pairs, 6.44 (± 3.97) and for general impostor
pairs, 4.22 (± 2.82). For the other three ﬁngers, these numbers of matched minutiae are about
the same. Although the number of matched minutiae for identical twin pairs is much smaller
than for genuine pairs, it is still larger than the number of matched minutiae for general impostor
pairs, which indicates the similarity between identical twins’ ﬁngerprints is larger than between
unrelated persons. The distribution of the number of matched minutiae between ﬁngerprints of
identical twins is signiﬁcantly different from the distribution of the number of matched minutiae
between general impostors (based on the Kolmogorov-Smirnov test).

2.4.2.2 Face
The face images were captured in color, all of them from a USB camera. The image size is 480 ×
640 pixels, but the face area in the image varies from 280 × 300 to 300 × 400 pixels. The images
51

contain non-uniform background, and some variations in illumination. The images were captured
in a sequence and over a short time interval. There are about 20 face images per subject. Some face
image examples are shown in Figure 2.5. The ﬁrst two images show face images of an identical
twin pair (Figure 2.5a), while the other two are face images of a non-identical twin pair (Figure
2.5b).

(a)

(b)
Figure 2.5: Face images of the ﬁrst and second twin in (a) an identical twin pair, and (b) a nonidentical twin pair.
The face subset used in the experiments contained 134 subjects, each one having around 20
images. The commercial matcher used to perform the facial matching was FaceVACS 7.1 [81],
and the scores from this matcher are in the range [0, 1]. We considered two face images per person
(the template and the query) in our experiment, mainly because the pictures were taken over a very
short time interval, which makes them very similar. The ﬁrst and the last images of each person
were used, since they are expected to be the least similar. Figure 2.7c shows identical twin and
general impostor distributions, along with the genuine distribution for the face experiments. In

52

the genuine distribution, we can see that there is almost no tail and that about 90% of the genuine
match scores are of more than 0.95. This is due to the high degree of similarity between the two
face images of the same person, since they were taken in sequence and in a very short time interval.
Also, we can observe that the identical twin impostor distribution is more similar to the genuine
distribution than to the general impostor distributions, meaning identical twins are a real challenge
to face recognition systems.

2.4.2.3 Iris
The iris images were captured using an IKEMB-100 camera produced by IrisKing [82]. The size
of the images is 480 × 640 pixels, but the approximate iris diameter is 200 pixels. IKEMB-100
is an embedded dual-eye iris camera and has an LCD displaying real time iris images for user
convenience. The iris images were captured in sequence and over a short time interval. There are
10 images for most of the subjects, with a few individuals having a smaller number of images.
Some examples of iris images are shown in Figure 2.6. Figures 2.6a and 2.6b show iris images of
an identical twin pair, where the two images in 2.6a are the left and right iris images of the ﬁrst
twin, and the two images in 2.6b are the left and right iris images of the second twin. Similarly,
Figures 2.6c and 2.6d show iris images of a non-identical twin pair.
The iris feature representation method based on ordinal measures [77] is used to test the performance of iris recognition for twins. The match scores range from 0 to 1. Two iris images were
randomly selected as probe and gallery for each eye. Figure 2.7d shows the genuine, identical twin,
and general impostor distributions of the right iris, which performed slightly better than the left iris.
The identical twin impostor distribution is very similar to the general impostor distribution. However, the peaks that are present in the identical twin impostor distribution tail may indicate that the
irises of identical twins have some correlation. The Kolmogorov-Smirnov statistical test indicates
that the identical twin and general impostor distributions are signiﬁcantly different (signiﬁcance of
0.05).

53

(a)

(b)

(c)

(d)

Figure 2.6: The left and right iris images of identical ((a) and (b)) and non-identical twin pairs ((c)
and (d)).
2.4.2.4 Multibiometric Experimental Results
In our multibiometric experiments, we ﬁrst combined units of the same modality, like the two irises
or all the four ﬁngers of a person. We performed a simple fusion using scores from the four ﬁngers
by summing them. The identical twin and general impostor distributions, along with the genuine
distributions, are shown in Figure 2.8a. A fusion using both the irises was also performed, and
the distributions are shown in Figure 2.8b. We can observe from these ﬁgures that the matching
results are extremely good; genuine and impostor distributions are well-separated. This indicates
that multimodal biometric systems can work very well even in the presence of identical twins in
the biometric database.
Two different Receiver Operating Characteristic (ROC) curves were generated for each experiment. Identical twin impostor ROC curve means the impostor matches considered were just
identical twin impostor matches, while a general impostor ROC curve means we considered all the
impostor matches, except the identical twin impostor matches. Figure 2.9a shows the ROC curves
for ﬁngers nos. 3 and 4, the best and worst performing ﬁngers, respectively, and the ROC curves
54

0.4
0.3

0.5

Genuine
Identical Twin Impostor
General Impostor

0.4

Genuine
Identical Twin Impostor
General Impostor

0.3
0.2
0.2
0.1
0
0

0.1
0.5
Score

0
0

1

0.5
Score

(a) Finger 3

1
0.8

1

(b) Finger 4

0.2

Genuine
Identical Twin Impostor
General Impostor

0.15

Genuine
Identical Twin Impostor
General Impostor

0.6
0.1

0.4
0.05

0.2
0
0

0.5
Score

1

(c) Face

0
0.5

0.6

0.7
0.8
Score

0.9

(d) Right Iris

Figure 2.7: Identical twin, general impostor distributions and genuine distributions for ﬁngers 3
and 4, face, and right iris.
of the 4-ﬁnger fusion. The performance of the ﬁngerprint identiﬁcation system in distinguishing
genuine matches from general impostor matches is always better than the performance in distinguishing genuine matches from identical twin impostor matches. It is important to note that the
4-ﬁnger fusion had the highest performance among the experiments involving ﬁngerprints.
The iris matching experiment results show that the iris biometric system can distinguish identical twins as much as it can distinguish any two different persons who are not identical twins, as
shown in Figure 2.9b. Iris experiments presented the best performance among all the experiments

55

0.4
0.3

0.2

Genuine
Identical Twin Impostor
General Impostor

0.15

0.2

0.1

0.1

Genuine
Identical Twin Impostor
General Impostor

0.05

0
0

0.5
Score

1

(a) 4-ﬁnger fusion

0
0.5

0.6

0.7
0.8
Score

0.9

(b) 2-iris fusion

Figure 2.8: Identical twin, general impostor, and genuine distributions of the 4-ﬁnger fusion and
2-iris fusion, respectively.
that used just one biometric characteristic (one ﬁnger, one iris, or face). This may be because irises
from identical twins are more uncorrelated than ﬁngerprints from identical twins, or may be due to
the fact that iris images in our dataset have a very good quality due to automatic rejection of poor
quality images by quality control software, while the quality of many ﬁngerprint images is very
poor due to improper sweep operation by some of the children, and the image area is very small.
Also, the 2-iris fusion showed improvements compared to the performance of each iris alone.
We also fused face and ﬁnger no. 4, where the ﬁngerprint scores were appropriately normalized, and the results are shown in Figure 2.9d. We did not combine iris with another biometric
modality because iris results are already really good and the database is not large enough for measuring lower error rate. The ROC curves of the face experiments are shown in Figure 2.9c. The
face experiment shows that the presence of identical twins in the face database causes the face
recognition performance to drop. Although the performance was good based on the general impostors (no identical twins in the data involved), we can only get a True Acceptance Rate (TAR) for
the identical twin data greater than zero at a false acceptance rate of over 10% for identical twins.
Compared to face matcher, the fusion of face and ﬁnger no. 4 has a better performance in dealing
with identical twin and general impostors. Compared to ﬁngerprint matcher, this fusion degraded

56

the performance on identical twin impostors, but improved the performance on general impostors.
It should be noted that the inferior performance of ﬁngerprint matcher compared to face matcher
in our experiments is due to the speciﬁc ﬁngerprint sensor (swipe sensor) used for data collection
and the difﬁculty of obtaining good quality ﬁngerprint images of younger subjects using the swipe
sensor. On the other hand, there is very small intra-class variation for face images.
Table 2.2 shows the equal error rates for the worst and best performing ﬁngers (ﬁnger numbers
3 and 4), 4-ﬁnger fusion, left and right irises, iris fusion, face, and face and ﬁnger 4 fusion; the
equal error rates are shown based on the identical twin impostors and general impostors separately.
It can be observed that all the modalities have a better performance on the general population than
on the identical twin population.
Table 2.2: Equal Error Rate (%) for distinguishing (i) genuine vs. impostor identical twins and (ii)
genuine vs. general impostors based on different biometric features.

Modality

Finger

4-ﬁnger

Left

Right

2-iris

3
Identical

Finger

Face

Face+

4

fusion

Iris

Iris

fusion

13.95

6.79

0.49

1.35

0.86

0.49

13.67

7.65

10.71

4.40

0.00

0.75

0.75

0.00

3.79

2.48

Finger 4

twins
General
impostors

2.4.2.5 Discussion
In the previous sections, we presented experimental results based on state-of-the-art biometric
recognition systems available in 2009. In this section, we discuss more recent results to indicate
the development of the systems overtime. In Fig. 2.11, we show ROC curves for face recognition
using FaceVacs 8.6, applied to both versions of CASIA Twins database (see Fig. 2.10 for examples
of face images in CASIA Twins V2). It can be noted that face recognition performance shows
an overall improvement given the controlled environment of CASIA Twins V1. However, the
challenge of the illumination, pose and expression variations is evidenced in Fig. 2.11b, which
57

(a) Fingerprint

(b) Iris
Figure 2.9: ROC curves for ﬁngerprint, iris and face, and a fusion example (face + ﬁnger 4). Due
to the small number of identical twin impostor (102), FAR less than 1/102 cannot be estimated.
shows the true acceptance rates at various false acceptance rates in the CASIA Twins V2 database.
In this unconstrained environment, it is even more difﬁcult to distinguish between identical twins.
In the iris and ﬁngerprint cases, we repeated the experiments using VeriEye [83] and Veriﬁnger
6.3 [79]. The ROC curves considering identical twin impostors and general impostors are very

58

Fig. 2.9 (cont’d)

(c) Face

(d) Multimodal Fusion
similar for iris, as it was the case in our previous experiments; for ﬁngerprints, the small drop in
the performance is still present when identical twins impostors are considered compared to general
impostors.

59

Figure 2.10: Examples of face images in CASIA Twins V2 database.
2.4.3 Finding Similarities between Identical Twins
In the previous section, we discussed the ability of biometric systems to distinguish between identical twins to answer the following question: can biometrics be used to distinguish identical twins?
A complementary question of interest is: can biometrics be used to detect identical twins? The
answer to this question is not known. The goal of this section is to provide quantitative measures
that will help in answering this second question.
Identical twins are the most genetically similar persons since they share the same DNA. Studies
on the characteristics of identical twins can generally lead to improvements in the overall knowledge about unrelated persons. For example, the knowledge about the inﬂuence of the genetics in
the occurrence of a certain disease can lead to the discovery of environmental factors in addition
to genetic factors. This type of knowledge can also be useful in biometric recognition. In [68], the
authors discussed the similarity between irises of identical twins. In their paper, they found that
humans can envision some type of texture similarity between irises of identical twins, and therefore to determine genetic relations from iris images might be possible. This, in turn, could help in
identifying victims without any form of identiﬁcation, by ﬁnding genetically related persons.
In general, studies on the similarities between biometrics of twins can lead to a broader understanding of the similarities between biometrics of the general population. Starting with twin
similarities, it can possibly be extended to ﬁnd other blood (familial) relations. In addition, the
protection of biometric data should be taken seriously since by stealing a template, one might be

60

(a) CASIA Twins V1

(b) CASIA Twins V2
Figure 2.11: ROC curves for face using FaceVacs 8.6.
able to ﬁnd genetically related persons. In the general research scenario, identical twins are of
great interest. If biometrics can be used to correctly determine whether two subjects are identical
twins, it might be much less expensive and quicker than DNA comparison.
Suppose we have N genuine biometric sample pairs, N identical twin impostor pairs, and N
general impostor pairs. Among these pairs, how well can we determine the identical twin pairs?

61

By using face and iris modalities together, we are able to correctly detect 80% of the identical
twins while falsely accepting 2% and 10% of genuine or general impostor pairs as identical twins in
database CASIA Twins V1 and V2, respectively (see Fig. 2.12). Due to the different characteristics
of the databases, it can be said that, in conditions closer to frontal face images and indoor lighting,
we can correctly determine 80% of the identical twin pairs, while falsely accepting 2% of general
subject pairs as identical twins. When the lighting, pose and expression conditions vary, the false
acceptance rate goes up to 10%.
We considered the following scheme to detect identical twins: if iris score between the subjects
is more than a threshold, the pair is considered a genuine pair, and the measure for identical twin
determination is set to zero. If iris score is less than some threshold, then the pair is considered
an impostor pair (note that identical twins fall under this category). Because of the high similarity
between faces of identical twins, the measure for identical twin determination is then set to the
similarity between the faces of the subjects. For this experiment, we randomly choose two face
images and two images of the same iris (right) for each subject. When using ﬁngerprints to determine genuine pairs instead of iris, the performance is about the same for the CASIA Twins V2,
while the performance drops for CASIA Twins V1 mainly because the quality of the ﬁngerprint
images in V1 is worse than in V2 (see Fig. 2.13).
Fig. 2.14(a) shows face images of an identical twin pair for which the identical twin evidence is
extremely high due to the similarity in their facial appearance and dissimilarity in their iris textures.
However, the twin pair shown in Fig. 2.14(b) who was classiﬁed by us as a non-identical twin pair
also shows an extremely high evidence of being identical twins according to face and iris.
Our work focused on answering the question: given a pair of subjects, can we determine
whether they are identical twins by means of biometrics? Another more general way of approaching the problem is to come up with a likelihood that a given pair of subjects is an identical twin
pair based on their biometrics.

62

(a) CASIA Twins V1

(b) CASIA Twins V2
Figure 2.12: ROC curves for twin detection using iris and face.

63

(a) CASIA Twins V1

(b) CASIA Twins V2
Figure 2.13: ROC curves for twin detection using ﬁngerprint and face.

64

(a) Match score: 0.99954

(b) Match score: 0.99951
Figure 2.14: Face images of pairs of (a) identical and (b) non-identical twins with extremely high
face match scores considering the range of scores is [0, 1] using FaceVacs 8.6.)

65

2.5 Summary and Conclusions
We have presented a study of the distinctiveness of biometric characteristics of identical twins
(ﬁngerprint, iris and face). The discriminability of these three biometric traits is supported by
anatomy and the formation process of the biometric characteristics, as discussed in Section 2.3.
We have assessed the capacity of state-of-the-art commercial biometric matchers in distinguishing
identical twins based on ﬁngerprint, iris, and face.
The unimodal face biometric system can distinguish two different persons who are not identical twins much better than it can distinguish identical twins. More recent studies support this
conclusion, and it becomes even more challenging when facial expression changes or different
lighting conditions are involved during the image capture process. Some efforts have been made to
improve the face recognition performance for identical twins, for example, by using facial marks.
As described in Section 2.3, face undergoes change over time based on usual facial expressions,
environment, diet, etc. At the same time, face recognition systems are evolving, so, in future, they
are expected to have a better performance.
Although the unimodal ﬁngerprint biometric system also can discriminate two different persons
who are not identical twins better than it can discriminate identical twins, this difference is not as
large as for the face biometric system. In the ﬁngerprint experiments, the identical twin impostor
distribution is shifted to the right, getting closer to the genuine distribution. This suggests a higher
correlation between ﬁngerprints of identical twins compared to ﬁngerprints of unrelated persons.
Previous studies have shown that the ﬁngerprint type is much more likely to be the same in twins
than in unrelated persons, and more recent studies conﬁrm this.
The iris matching experiment results show that the performance of the biometric system for the
identical twin data and for the general data are very similar, which means the iris biometric system
can distinguish identical twins to the same extent as it can distinguish any two persons who are not
identical twins. However, the shift in the identical twin impostor distribution for iris also suggests
a higher level of similarity between irises from identical twins than from unrelated persons.
Among all the unimodal biometric systems considered here, the iris performed the best. Again,
66

this may be due to the fact that irises from identical twins are more uncorrelated than ﬁngerprints
from identical twins, or may be due to the fact that iris images in our dataset have a very good
quality, while the quality of many ﬁngerprint images is relatively poor and the friction ridge area
is small due to the age of the subjects.
Multimodal biometric systems that combine different units of the same biometric modality
(fusion of 4 ﬁngers or 2 irises) lead to an almost perfect separation between genuine and impostor
distributions. For both the general population and identical twins, multimodal biometric systems
that combine different modalities, one being face, show improvements in the performance on the
general population compared to individual traits. However, the performance of multibiometric
systems drop on the identical twin data compared to the highest performance modality that is
being combined. This is because in the fusion of ﬁnger and face matchers, the performance of the
face matcher for identical twins is extremely poor.
Biometric traits can also be used to determine whether two subjects are identical twins. By using face and iris modalities together, for example, we can correctly determine 80% of the identical
twin pairs as such, while only 2% of subject pairs who are not identical twins will be incorrectly
considered identical twins.
To our knowledge, although multibiometric databases for twins have been collected before (the
ten ﬁngers and/or the two palmprints), no previous study has combined different units/modalities
to study the multibiometric matching performance. We have performed the multibiometric experiments and showed that the performance of a multibiometric system that uses different units of the
same modality is signiﬁcantly better compared to unimodal systems, approaching almost perfect
accuracy on our database. Also, there have not been any previous studies of identical twin irises on
a database this large. In addition, we have used commercial matchers for the face and ﬁngerprint
experiments, which are usually more accurate than in-house matchers – used in previous studies.
Based on our experiments on this relatively small twin database, we can conclude that the presence
of identical twin data poses a real challenge to commercial face recognition systems.

67

CHAPTER 3
LATENT FINGERPRINT MATCHING
3.1 Introduction
Fingerprint identiﬁcation was a completely manual approach until 1970s. Due to growing demands
on ﬁngerprint matching, and the large size of the ﬁngerprint database of criminals, research was initiated to automate ﬁngerprint recognition, which led to the development of Automated Fingerprint
Identiﬁcation Systems (AFIS). These systems are now used worldwide not only by law enforcement agencies but also in many other government applications, including background check of
certain employees (e.g., those serving in the military), border control and national ID cards. The
use of ﬁngerprint recognition in civilian applications (e.g. logical/physical access control) is also
gaining more widespread acceptance [84].
There are essentially three types of ﬁngerprints encountered in law enforcement applications
(see Fig. 3.1): rolled, plain and latents. Rolled prints are obtained by rolling the ﬁnger from nailto-nail, while plain (or slap) ﬁngerprints are obtained by placing the ﬁnger ﬂat on a surface. It
is common practice in law enforcement agencies to collect ﬁngerprints from all the ten ﬁngers of
a subject. An example of a tenprint card is shown in Fig. 3.2, in which the ﬁrst two rows show
rolled prints of the ten ﬁngers of a subject, while the last row shows plain ﬁngerprints also from
the ten ﬁngers of a subject. Plain ﬁngerprints are collected by ﬁrst placing the four ﬁngers of
one hand, then the four ﬁngers of the other hand, followed by the capture of the impressions of
68

the two thumbs (4-4-2 process). Latents are impressions of ﬁngers lifted from surfaces of objects
that are inadvertently touched or handled by a person typically at crime scenes. Lifting of latents
may involve a complicated process, and it can range from simply photographing the print to more
complex dusting or chemical processing [35].

(a) Rolled

(b) Plain

(c) Latent

Figure 3.1: Three types of ﬁngerprint impressions. Rolled and plain ﬁngerprints are also called
reference ﬁngerprints.
Two types of matching are of interest to law enforcement agencies: tenprint-to-tenprint and
latent-to-tenprint. Tenprint-to-tenprint matching is used in border control, background checks,
etc., while latent-to-tenprint matching is used to identify suspects from impressions lifted from
crime scenes.
Rolled prints contain the largest amount of information about the ridge structure on a ﬁngerprint since they capture the largest ﬁnger surface area; latents usually contain the least amount of
information for matching or identiﬁcation because of their small size (they capture only a subset
of the complete friction ridge pattern) and inherent noise. Compared to rolled or plain ﬁngerprints,
latents are smudgy and blurred and have large nonlinear distortion due to pressure variations. Due
to their poor quality and small area, latents have a signiﬁcantly smaller number of minutiae compared to rolled or plain prints (the average number of minutiae in NIST Special Database 27 (NIST
SD27) [4] images is 21 for latents versus 106 for their mated rolled prints). These characteristics
make the latent ﬁngerprint matching problem very challenging.

69

Figure 3.2: An example of a tenprint card containing two types of impressions (rolled and plain)
of the ten ﬁngers. The ﬁrst portion of the card contains personal information related to the person
being ﬁngerprinted, such as name, date of birth, place of birth, gender, race, height, weight, etc.
The second portion of the card contains impressions of the ten ﬁngers. The ﬁrst and second rows of
impressions consist of rolled impressions of the ﬁngers from the left and right hands, respectively
(from left to right: thumb, index, middle, ring, little). The third row contains plain impressions
of the left four ﬁngers taken simultaneously, left thumb, right thumb, and right four ﬁngers taken
simultaneously.
Fingerprint examiners who perform manual latent ﬁngerprint identiﬁcation follow a procedure
referred to as ACE-V (analysis, comparison, evaluation and veriﬁcation) [85]. Because the ACE-V

70

procedure is quite tedious and time consuming for latent examiners, latents are usually matched
against reference prints of a small number of suspects identiﬁed by other means, such as eye witness description or M.O. (mode of operation). With the availability of AFIS, ﬁngerprint examiners
are now able to match latents against a large ﬁngerprint database using a semi-automatic procedure
that consists of following stages: (i) manually mark the features (region of interest, minutiae and
singular points) in the latent, (ii) launch an AFIS search, and (iii) visually verify the top-N (N is
typically 50) candidate ﬁngerprints returned by AFIS. The accuracy and throughput of this latent
matching procedure is still not satisfactory. It certainly does not meet the “lights-out mode”1 of
operation desired by the FBI and which is the goal of the Next Generation Identiﬁcation [86].
It is our opinion that research efforts in latent ﬁngerprint identiﬁcation should focus on improving the matching accuracy based on existing mark-ups by latent experts, rather than on completely
eliminating manual input, or asking examiners for too much input. This opinion is supported by
the following facts: (i) latent matching accuracy is still the major concern of law enforcement
agencies, (ii) manually marking extended features is very labor extensive, and (iii) state of the
art “lights-out” latent identiﬁcation systems cannot yet offer satisfactory accuracy for most latents
of casework quality. We believe latent matching will not reach the same level of performance of
tenprint-to-tenprint matching due to some of the characteristics of the latents: (i) in tenprint acquisition, if the quality of the ﬁngerprint image is not good, the ﬁngerprint can be re-captured; in the
latents case, this option is not available; (ii) for some latent prints, there might not be enough information to make an identiﬁcation, for example, if the friction ridge area is extremely small and the
ridges are not prominent or faded. Our goal of improving latent ﬁngerprint matching performance
aims to reduce the manual labor involved in identifying a latent. This is achieved by providing
higher retrieval accuracy at top ranks, thus reducing the number of latent and rolled print pairs that
need to be manually veriﬁed.
1 Lights-out

identiﬁcation refers to a system requiring minimal or zero human assistance
in which an image is presented as input, and the output consists of a short candidate list
(deﬁnition from http://biometrics.nist.gov/cs_links/latent/workshop09/
proc/DefineLPlightsout.pdf).

71

Our efforts to improve latent ﬁngerprint matching performance are concentrated on two different directions: (i) enhance the latent image and fuse the matching score of the enhanced latent with
the score based on manually marked minutiae [87]; (ii) develop a new latent ﬁngerprint matcher
based on manually marked minutiae [88]. In the current practice, latent examiners are required
to mark minutiae and optionally mark singular points (core/delta). In the enhancement approach,
manually marked features in the latents include minutiae, region of interest, and singular points; in
the latent matching approach, we reduce the manual input to only the manually marked minutiae.
Orientation ﬁeld information is critical in the ﬁngerprint enhancement process. Orientation
ﬁeld can usually be reliably estimated from the ﬁngerprint image itself in the case of rolled or plain
ﬁngerprints of good quality. In latent ﬁngerprint images, orientation ﬁeld estimation is not very
reliable because of their poor quality. Fig. 3.3 shows the estimated orientation ﬁeld of a latent and
its mated rolled print in NIST SD27 latent database using gradient-based approach [8]. Manually
marked orientation ﬁeld, although reliable, is not easy to obtain; it requires a lot of effort and prior
training. Therefore, for the enhancement process, we reconstruct the orientation ﬁeld based on
manually marked features (minutiae, singular points, and region of interest) [89]. Gabor ﬁlters
are then used to enhance latent images, which are automatically matched to the reference print
database. We show that the performance of manually marked minutiae matching can be improved
by combining scores from both the matching of manually marked minutiae and of automatically
extracted minutiae from enhanced latents. It is important to point out that the performance of
fully automated minutiae extraction and matching based on the input image is very poor. The
experiments were conducted on a public domain ﬁngerprint database, NIST SD27, consisting of
258 latents along with their rolled mates. To make the conclusion more reliable, the gallery size
was increased by including 27, 000 rolled images from NIST SD14 [90].
For ﬁngerprint matching, there are two major problems which need to be solved. The ﬁrst is to
align the two ﬁngerprints to be compared and the second is to compute a match score between the
two ﬁngerprints. Alignment between a latent and a rolled print is a challenging problem because latents often contain a small number of minutiae and undergo large skin distortion. To deal with these

72

(a)

(b)

(c)

(d)

Figure 3.3: Examples of estimated orientation ﬁeld ((b) and (d)) by gradient-based approach [8]
for a (a) latent and (c) its mated rolled print in NIST SD27.
two problems, we propose the descriptor-based Hough transform (DBHT), which is a combination
of the generalized Hough transform and a local minutiae descriptor, called the Minutia Cylinder
Code (MCC) [91]. The MCC descriptor improves the distinctiveness of minutiae while the Hough
transform method can accumulate evidence as well as improve the robustness against distortion.
Match score computation between a latent and a rolled print is also challenging because the number of mated minutiae is usually small. To address this issue, we further consider orientation ﬁeld
as a factor in computing the match score. Since we only consider manually marked minutiae for
latents, a reconstruction algorithm based on minutiae alone is used to estimate the orientation ﬁeld
[89]. The proposed matcher was tested on two latent ﬁngerprint databases, NIST SD27 database
73

and West Virginia University latent ﬁngerprint database (WVU LFD). Two COTS matchers and
a state-of-the-art non-commercial ﬁngerprint matching algorithm (MCC SDK) were evaluated on
these two databases. Our algorithm was found to perform better than these three matchers on both
the databases. Extensive experiments on fusion of matchers and effect of ﬁngerprint quality were
also conducted and reported here.
The rest of this chapter is organized as follows: in Section 3.2, related work is reviewed; in
Section 3.3, the features used in the matching experiments are described; in Section 3.4, the two
approaches for improving latent matching performance are presented; in Section 3.5, the databases
used are described; in Section 3.6, experimental results based on the two proposed approaches are
presented and discussed; in Section 3.7, our work on latent ﬁngerprint matching is summarized.

3.2 Related Work
In this section, we review related work in four areas: (i) published research on rolled ﬁngerprint
matching2, (ii) published research on latent ﬁngerprint matching, (iii) NIST evaluation of latent ﬁngerprint technologies (ELFT), and (iv) evaluation of latent examiners. Rolled ﬁngerprint matching
technology is very advanced, and there exist commercial matchers with excellent matching performance; however, there is very little information available about the details of these algorithms and,
therefore, we do not discuss them here.

3.2.1 Rolled Fingerprint Matching
The majority of the algorithms developed for ﬁngerprint matching are based on minutiae. Although
minutiae carry a great amount of discriminatory information, additional features may be necessary
to achieve the desired level of accuracy. Most proposed algorithms for ﬁngerprint matching that use
non-minutiae features also use minutiae. For example, some algorithms combine ridge orientation
with minutiae information either at feature level by including ridge orientation information in local
2 See Chapter

4 in [1] for a more comprehensive review of this topic.

74

minutiae descriptors [27, 92] or at score level by combining scores from minutiae matching and
global orientation ﬁeld matching [92, 93].
Several recent studies on ﬁngerprint matching have focused on the use of local minutiae descriptors [91, 27, 92, 94, 95, 96, 97]. In most of these studies, the initial step consists of using
local minutiae descriptors to obtain the alignment between two ﬁngerprints by considering the
most similar minutiae pair; then, a global consolidation step is performed to obtain a better matching performance. Since these algorithms are usually tuned and evaluated using FVC databases
(plain ﬁngerprints) or NIST Special Database 4 (rolled ﬁngerprints), their performances on latent
ﬁngerprints are unknown.

3.2.2 Latent Fingerprint Matching
Recent research and development efforts on latent ﬁngerprints can be classiﬁed into three streams
according to the manual input required from ﬁngerprint examiners: consistent with existing practice, increasing the amount of manual input, or reducing the amount of manual input. Because
of large variations in latent ﬁngerprint quality and speciﬁc requirements of practical applications
(crime scenes, border crossing points, battle ﬁelds), each of the three streams has its value.
Improved latent matching accuracy has been reported by using extended features, which are
manually marked for latents [28, 98, 99, 11]. However, marking extended features (orientation
ﬁeld, ridge skeleton, etc.) in poor quality latents is very time-consuming and might be only feasible
in rare cases when there are none or very few minutiae. Thus, some studies have concentrated on
latent matching using a reduced amount of manual input, such as manually marked region of
interest (ROI) and singular points [100, 101], or no manual input [102]. However, the reported
performance of these approaches is not very good. Hence our proposed matcher takes manually
marked minutiae as input and, therefore, it is consistent with existing practice in forensics. There
have also been some studies on fusion of multiple matchers [103] and use of multiple latent prints
[104].

75

3.2.3 NIST Evaluation of Latent Fingerprint Technologies

(a) Good

(b) Bad

(c) Ugly

Figure 3.4: Latent ﬁngerprints of three different quality levels in NIST SD27.
NIST has been conducting a multi-phase project on Evaluation of Latent Fingerprint Technologies (ELFT) to evaluate latent feature extraction and matching techniques [105]. Since all
participating algorithms in ELFT are proprietary, no detailed information about these algorithms
is available. The purpose of ELFT-Phase I was to assess the feasibility of latent ﬁngerprint identiﬁcation systems using Automated Feature Extraction and Matching (AFEM), while the purpose
of ELFT-Phase II was to actually measure the performance of state-of-the-art AFEM technology
and evaluate whether it was viable to have those systems in operational use to reduce the amount
of time needed by latent examiners to manually mark latents thereby increasing the throughput.
In Phase I, latent images were selected from both operational and non-operational scenarios.
The average reported accuracy at rank-1 was of 67% (100 latents matched with 10, 000 rolled
prints) [106]. In Phase II, latent images were selected from only operational environments. The
rank-1 accuracy of the most accurate system was 97.2% (835 latents matched with 100, 000 rolled
prints) [50]. The Phase I and Phase II accuracies cannot be directly compared since the Phase I
and Phase II evaluations used different latent databases. Furthermore, the quality of latents used
in Phase II is much better compared to Phase I. As shown in Fig. 3.4, the quality of operational
latents can vary signiﬁcantly.
The impressive matching accuracy reported in ELFT, specially Phase II, does not support that

76

the current practice of manually marking minutiae in latents should be changed. Although latents
in Phase II were selected from operational scenarios, they represent successful identiﬁcations in
actual case examinations using existing AFIS technology. In the ACE-V process, when the examiner analyzes the latent image he decides whether the latent has (i) value for exclusion only, (ii)
value for individualization or (iii) no value. If a latent is classiﬁed as of “no value”, no comparison
is performed. If the latent is classiﬁed in one of the other two categories, then a comparison is carried out and the examiners can make a decision on individualization and exclusion, or determine
the comparison to be inconclusive. So the latents which are successfully identiﬁed constitute only
a small part of all latents, which are of reasonable quality. For this reason, in the ELFT-Phase II
report [50] the authors concluded that only a limited class of latents can beneﬁt from the AFEM
technology.
NIST has conducted another evaluation of latent ﬁngerprint technologies using extended feature sets (EFS) manually marked by latent examiners, called ELFT-EFS [107]. In this evaluation,
the purpose was to investigate the matching accuracy when (i) latent images and/or (ii) sets of
manually marked latent features were provided to the matcher. This evaluation suggested that the
highest accuracy was obtained when the input included both the latent image and manually marked
latent features.

3.2.4 Evaluation of Latent Examiners
A latent examiner can be viewed as a very accurate “matcher”. But, there is the issue of subjectivity
in the human decision making process along with low throughput. Because of the time consuming
process of manual decision making, quantitatively estimating the accuracy of latent examiners is
not easy. Hence the numbers of ﬁngerprint pairs used in several “black box” studies of latent
examiners are not large [108, 109, 110]. Although the exact numbers reported in these studies may
not reﬂect the real practice, the qualitative conclusions are very useful. It was found that latent
examiners’ conclusions are not always in agreement, especially in the case of poor quality latents
[108]. In addition, the same examiner can change his/her conclusions on the same ﬁngerprint pair

77

at a later time [109]. These inconsistences may even increase under bias [110].
These issues associated with including latent examiners in the latent identiﬁcation process will
only be resolved when the automatic matcher can outperform latent examiners in accuracy. No
matter how successful the application of automatic ﬁngerprint recognition technology might be,
we cannot say ﬁngerprint matching is a “solved problem” before we can reach the goal of outperforming latent examiners.

3.3 Latent Fingerprint Features
In both of our approaches to improve latent matching performance, minutiae and orientation ﬁeld
are extracted from latent as well as rolled prints. Minutiae are marked by latent examiners in
the latent, and automatically extracted using commercial matchers in the rolled print. Orientation
ﬁeld in the latents is reconstructed from minutiae location and direction, as proposed in [89], and
orientation ﬁeld is automatically extracted from the rolled print images by using a gradient-based
method [8]. In the matching using descriptor-based Hough Transform, local minutia descriptors
are used. In this section, the orientation ﬁeld reconstruction algorithm is described, followed by
the description of the local minutia descriptors used in this work.

3.3.1 Orientation Field Estimation
The value of an orientation ﬁeld at a given pixel is the angle that the ﬁngerprint ridges form with
the horizontal axis in a small neighborhood around that pixel (see Fig. 3.5). The simplest and the
most natural approach for computing orientation ﬁeld is based on the gradient values. Because of
the computational effort and the presence of noise, the orientation ﬁeld is usually computed in a
small neighborhood instead of at each pixel [1].

78

Figure 3.5: Illustration of the orientation ﬁeld computed at a small block.

Another approach to compute the orientation ﬁeld that differs from this image-based approach
is the model-based approach (e.g. zero-pole model [111, 112, 113]), which uses the location of
singular points to estimate the orientation ﬁeld based on a model.
In the latent ﬁngerprint case, since the images are usually of poor quality, it is very difﬁcult to
estimate the orientation ﬁeld based only on the image itself. Also, in many cases, the friction ridge
area of the latent ﬁngerprint is small and does not contain singular points, which makes the modelbased orientation ﬁeld unreliable. A combination of image-based and model-based approaches can
be very useful for latents. In [100], such a combination was used to enhance the latent ﬁngerprint
images and it was shown to improve the matching performance compared to the performance based
on latent image alone.
In [89], the authors proposed a method to reconstruct the orientation ﬁeld assuming (i) only
the minutiae information (location and direction) is given, and (ii) both minutiae and singular
point information is given. The advantages of this approach over the approaches described earlier
[111, 112, 113] are: it can be used in cases where no singular points are available (small area
latents) and it uses minutiae information that is reliable. Our experiments show that when this
reconstructed orientation ﬁeld is used for latent enhancement, the matching performance of the
enhanced image is comparable to the matching performance of images enhanced by manually
marked orientation ﬁeld. This reconstructed orientation ﬁeld is also useful for orientation ﬁeld

79

matching. Therefore, we adopt this approach which is described in detail below.

3.3.1.1 Minutiae based Orientation Field Reconstruction
In this section, we describe the method proposed in [89] to reconstruct the orientation ﬁeld based
on minutiae and singular points, if available. Here we assume that a latent expert has marked
the available features (minutiae, singular points, region of interest). Also, in cases where an odd
number of singular points were marked in the region of interest, paired singular points outside
of the region of interest are guessed (referred to as virtual singular points [100]) to satisfy the
constraints of singularity number, which states the total number of singular points is even, and the
cores and deltas appear in pairs.
Because of the reasons mentioned in Section 3.3.1, the orientation ﬁeld is also computed in
non-overlapping blocks of a predeﬁned size (e.g., 8 × 8 or 16 × 16). Now, consider lines passing
through the non-overlapping blocks, that divides the image into 8 equally spaced sectors as shown
in Fig. 3.6. Then, a local orientation ﬁeld estimate is obtained, for each block, based on the
direction of the nearest minutia in each of the 8 sectors.

Figure 3.6: Local ridge orientation estimation based on the nearest minutiae in each sector.
Let {xn , yn , αn }, 1 ≤ n ≤ N, be a set of N ﬁngerprint minutiae marked by a latent expert, where
(xn , yn ) is the location and αn is the direction of the nth minutia. Then, by doubling the minutia
direction αn , which means taking 2αn instead of αn as the minutia direction, it becomes equivalent
to αn + π (this is necessary since ﬁngerprint ridges are not oriented). For the K minutiae selected
in eight sectors (in our experiments, one minutia per sector), cosine and sine components can be

80

computed and summed as follows:
K

u=

∑ cos (2αk )wk ,

(3.1)

∑ sin (2αk )wk ,

(3.2)

k=1
K

v=

k=1

where wk is a weighting function based on the Euclidean distance between the block center and the
kth minutia; this makes the closest minutia direction dominates the ridge orientation of neighboring
blocks.
The orientation at block (m, n) is then computed as:
D(m, n) =

v
1
arctan .
2
u

(3.3)

If we consider that marked singular points are also available in the latent ﬁngerprints, then the
direction ﬁeld of Ns singular points is given by the Zero-Pole model [111], [112]:
Ns

Ds (m, n) =

n − ns

∑ tsi arctan m − msi ,

(3.4)

i

i=1

where msi , nsi , and tsi (core: 1, delta: −1) are the location and type of the ith singular point.
Orientation ﬁeld is ﬁrst estimated using minutiae whose direction is subtracted by Ds . The
reconstructed orientation ﬁeld is then given by
O(m, n) =

2D(m, n) + Ds(m, n)
.
2

(3.5)

The orientation ﬁeld is reconstructed simply as D(m, n) (Eq. 3.3) in the enhancement case in
which there are no marked singular points, and in the matching using Descriptor-Based Hough
Transform in which we consider only minutiae are marked. The orientation ﬁeld is only reconstructed in the region of interest (ROI). In the ﬁrst approach (enhancement) we assume region of
interest is manually marked. In the second approach, we estimate the region of interest as the
convex hull of minutiae.
Figs. 3.7 and 3.8 show some examples of reconstructed orientation ﬁeld. In both ﬁgures,
(a) shows the original latent image, (b) shows the orientation ﬁeld estimated directly from the
81

gray scale image, (c) shows the reconstructed orientation ﬁeld and (d) shows manually marked
orientation ﬁeld. We can observe that it is not easy to estimate the orientation ﬁeld from the latent
image. But, the reconstructed orientation ﬁeld from minutiae and singular points is quite reliable,
although it is not as smooth as manually marked orientation ﬁeld.

(a)

(b)

(c)

(d)

Figure 3.7: Comparison of orientation ﬁeld estimation methods: (a) Original latent ﬁngerprint
image, (b) orientation ﬁeld estimated directly from the given image, (c) orientation ﬁeld estimated
from minutiae and singular points, and (d) manually marked orientation ﬁeld.

82

(a)

(b)

(c)

(d)

Figure 3.8: Comparison of orientation ﬁeld estimation methods: (a) Original latent ﬁngerprint
image, (b) orientation ﬁeld estimated directly from the given image, (c) orientation ﬁeld estimated
from minutiae and singular points, and (d) manually marked orientation ﬁeld.

83

3.3.2 Local Minutia Descriptor
Local descriptors have been widely used in ﬁngerprint matching (e.g. [94, 27, 95, 11, 91]). Feng
and Zhou [114] evaluated the performance of local descriptors associated with ﬁngerprint matching
in four categories of ﬁngerprints: good quality, poor quality, small common region, and large
plastic distortion. They also coarsely classiﬁed the local descriptors as image-based, texture-based,
and minutiae-based descriptors. Their results show that the minutiae-based descriptor, Minutia
Cylinder Code (MCC) [91], performs better in three of the four categories (good quality, poor
quality and large plastic distortion), and texture-based descriptor performs better for the small
common region category.

3.3.2.1 Minutia Cylinder Code (MCC)
A minutia cylinder records the neighborhood information of a minutia as a 3D function. A cylinder
contains several layers and each layer represents the density of neighboring minutiae along the
corresponding direction. The cylinder can be concatenated as a vector, and therefore the similarity
between two minutia cylinders can be efﬁciently computed. Fig. 3.9(b) shows the sections of
two valid cylinders associated with the two corresponding minutiae (in the latent and in the rolled
print) indicated in Fig. 3.9(a). A more detailed description of the cylinder generation and of the
similarity between two cylinders can be found in [91].
A cylinder is divided into sections (height), and each section is divided into cells (base). Each
cell in the cylinder receives contribution from two different sources: the ﬁrst is related to the
distance of neighboring minutiae to the center of the cell, and the second is related to the directional
difference between the neighboring minutiae and the center minutiae (the main minutiae for which
the cylinder is being built). Sections correspond to directional differences, and a cell receives a
contribution only from those neighboring minutiae that have a directional difference to the center
minutia similar to the directional difference corresponding to the section where the cell is located.
A cell is only valid if it is inside the cylinder and if it is inside the convex hull of minutiae (with
some tolerance for the last condition). Smoothing functions (e.g. sigmoid functions) are used
84

(a) Latent and corresponding rolled print with a mated minutiae pair indicated.

(b) Sections of the cylinder corresponding to the minutia indicated in the
latent (ﬁrst row) and in the rolled print (second row).
Figure 3.9: Sections of two cylinders associated with the two corresponding minutiae, one in latent
and other in rolled print.
to compute directional differences between minutiae and the center minutia, and to compute the
distance among neighboring minutiae and cell center points. The contributions of all neighboring
minutiae to a cell are combined under another smoothing function to limit the contribution of
noisy regions, which usually contain a large number of minutiae. This contribution is compared to
a threshold so that the ﬁnal cylinder is a binary representation (bit-based implementation).
Only valid cylinders and cells are used in the similarity computation. If ca and cb are two cylinders represented as vectors, cab indicates which cells are valid in both cylinders, ca|b = ca AND cab ,
ˆ
ˆ
and cb|a = cb AND cab , then the similarity between the two cylinders in the bit-based implementaˆ
85

tion case is given by


 1 − ||ca|b X OR cb|a || if c and c are matchable

a
b
||ca|b|| + ||cb|a||
s(ca , cb ) =


 0
otherwise,

(3.6)

where ca and cb are matchable if (i) the directional difference between the two minutiae is less
than some predeﬁned threshold (maximum rotation allowed), (ii) there is a minimum number of
corresponding elements in the two cylinder (minimum size of the valid intersection), and (iii) the
denominator is different from zero.

3.4 Latent Matching Approaches
In this section, we present our two proposed approaches to improve latent ﬁngerprint matching
performance. The ﬁrst approach accomplishes that by using enhancement of the latent images,
while the second approach improves the performance by using a descriptor-based Hough Transform alignment in the matching algorithm.

3.4.1 Latent Matching with Enhanced Image
The feature extraction process in latent ﬁngerprints usually does not yield reliable features because
of the poor quality of the latents. One way of improving the latent feature extraction process is by
enhancing the latent image so that more reliable features can be extracted and a better matching
performance can be achieved. In [115], the authors proposed to enhance ﬁngerprint images using
Gabor ﬁlters to improve the clarity of ridges and valleys. This enhancement is performed in local
blocks of the image, and, for each block, an orientation is required. In this approach, orientation
ﬁeld reconstructed from minutiae and singular points (if available) as described in Section 3.3.1
is used to enhance the latent image. Then, a commercial matcher is used to extract minutiae and
perform the matching. The scores from manually marked minutiae and from the automatically
extracted minutiae from the enhanced images are then combined to improve the latent matching

86

performance. The overall scheme of the latent matching using enhanced image is illustrated in Fig.
3.10.

Figure 3.10: Overview of the latent matching with enhanced image approach.

Fig. 3.11 shows some examples of enhanced latent images, along with their skeletons, the
corresponding enhanced images and their enhanced image skeletons. It can be noted from the
ﬁgure that the clarity of the ridges improves and the noise is greatly reduced.
Another approach that is often used to improve the matching accuracy consists of combining
different sources of information. Information fusion can be implemented at a variety of levels, such
as feature level, score level, and rank level. Feature level fusion in our context would involve both
manually marked minutiae and automatically extracted minutiae for matching. In our preliminary
experiments, feature level fusion did not show promising results, so we focus on the other two
fusion levels, rank and score.

87

(a)

(b)

(c)

(d)

(e)

(f)

(g)

(h)

Figure 3.11: Examples of enhanced latent images. (a) and (e) are original latent images, (b) and
(f) are the skeletons of the original latent images extracted by VeriFinger 4.2 SDK, (c) and (g)
are enhanced images using reconstructed orientation ﬁeld, and (d) and (h) are the skeletons of the
enhanced images extracted by VeriFinger.
3.4.1.1 Rank-level Fusion
For rank-level fusion, we considered two techniques: Borda count [116] and the highest rank
method.
Borda count is a generalization of the majority vote. In our case where we have two different
matchers (classiﬁers), one matching manually marked latent to rolled and the other matching enhanced latent image to rolled, the Borda count for a given rolled print in the database will be the
sum of the number of rolled prints that are ranked below the true mate by each matcher. Then, the
ranking is performed by sorting the rolled prints in descending order based on their Borda count.
The magnitude of the Borda count for each rolled print in the database measures how much the
two matchers agree on whether the input comes from the true mate of the latent.
In the highest rank approach, each rolled print in the database is assigned the highest rank as
computed by the two matchers [117]. If there are any ties, the rolled prints are sorted by their lower

88

rank. If both higher and lower ranks are the same, the tie is broken randomly.
Our experiments show that while Borda count does not improve the matching performance of
manually marked minutiae, the highest rank fusion shows some improvement. These results will
be discussed in Section 3.6.1.1.

3.4.1.2 Score-level Fusion
Min, max, product and weighted sum rules are some of the well-known score-level fusion rules.
Let s1 and s2 be the two match scores obtained for the same (latent, rolled) ﬁngerprint pair by using
two different matchers, called matcher 1 and matcher 2. Then, for the match score pair s1 and s2
we can compute fused scores based on the score level fusion rules mentioned above as follows:
Smin

= min (s1 , s2 )

Smax

= max (s1 , s2 )

S prod

= s1 s2

(3.7)

Swsum = w1 s1 + w2 s2 ,
where w1 + w2 = 1.
In addition to the above score fusion rules, we also considered combining the two match scores
by applying a slight modiﬁcation of the boosted max fusion approach proposed in [98]. The main
purpose of boosted max is to boost the scores of genuine matches. It relies on the assumption
that the spatial transformations among the given latent, enhanced latent and rolled ﬁngerprints are
consistent for genuine matches and inconsistent for impostor matches.
In general, given a transformation matrix T and a two dimensional vector v = (x, y), the coordinates for v, v′ = (x′ , y′ ), after the transformation is applied is given by v′ = T v:
 

 
′
 x   cos θ sin θ ∆x   x 
 

 
 y′  =  − sin θ cos θ ∆y   y  ,
 

 
 

 
0
0
1
1
1
where θ is the rotation and ∆x and ∆y are translation parameters in x and y, respectively.

89

(3.8)

The parameters θ , ∆x, and ∆y in the transformation matrix can be estimated by using pairs of
corresponding minutiae points. Note here that scaling parameter is not considered due to the fact
that the ﬁngerprints are all in the same resolution (500 ppi).
Let TMR , TER , and TME be the estimated transformations between manually marked minutiae and rolled ﬁngerprints, enhanced latent images to rolled ﬁngerprints, and manually marked
minutiae to enhanced latent images, respectively. Given TER and TME , we can compute TMER =
TER ×TME , which denotes the spatial transformation of the manually marked latent to the automatically extracted minutiae in the enhanced latent, which is then matched to the rolled ﬁngerprint image. It is expected that the transformations TMR and TMER should be similar for genuine pairs and
different for impostor pairs. This would be likely the case since the two minutiae sets (manually
marked and extracted from enhanced image) are different and are expected to match to different
parts of the impostor rolled prints, while the two sets are expected to be correctly matched to their
true corresponding minutiae in the true mate rolled print. The similarity between the two transformation matrices is measured in terms of the rotation angle and the Euclidean distance between the
translations in x and y directions of each transformation. If the difference between the two rotation
angles in the two matrices TMR and TMER is less than some threshold and the Euclidean distance
between the two translations in those same matrices is less than another threshold, then the pair is
considered consistent. In our study, the rotation and translation thresholds were set to 30 degrees
and 140 pixels, respectively. The boosted max score for a given pair of match scores s1 and s2 is


w max (s , s ) + w min (s , s ), if consistent
1
1 2
2
1 2
(3.9)
Sb =
max (s , s ), otherwise.

1 2

In Section 3.6 we discuss the results obtained by using enhanced images and different fusion
schemes.

3.4.2 Latent Matching with Descriptor-Based Hough Transform
Another way of improving matching performance is to design a matcher especially for latent ﬁngerprint matching, which takes into account the special characteristics of latents such as small
90

number of minutiae. Our goal is to develop an algorithm for latent ﬁngerprint matching that uses
as few manually marked features as possible.
Given a latent ﬁngerprint (with manually marked minutiae) and a rolled ﬁngerprint, we extract
additional features (besides minutiae coordinates) from both prints, align them in the same coordinate system, and compute a match score between them. These three steps are described in the
following subsections. An overview of the proposed algorithm is shown in Fig. 3.12.

Figure 3.12: Overview of the proposed latent matching approach.

3.4.2.1 Alignment
Fingerprint alignment or registration consists of estimating the parameters (rotation and translation) that align two ﬁngerprints. A scale parameter can be ignored here because all the images
have the same resolution. There are a number of features that may be used to estimate alignment
parameters between two ﬁngerprints, including singular points, orientation ﬁeld, ridges, and minutiae. There are also a number of methods to align two ﬁngerprints: Generalized Hough Transform

91

(e.g., [29]), local descriptors (e.g., [91]), energy minimization, etc3 .
In the latent matching case, singular points are not always present in latents, making it difﬁcult to base the alignment of the ﬁngerprint on singular points alone. As state earlier, obtaining
manually marked orientation ﬁeld is expensive, and to automatically extract reliable orientation
ﬁeld from a latent image is difﬁcult. Since manually marking the minutiae in a latent is a common
practice for latent matching, our approach to align two ﬁngerprints is based on minutiae.
Local descriptors can also be used to align two ﬁngerprints. In this case, usually the most
similar minutiae pair is used as a base for estimating the initial transformation parameters (rotation
and translation); the most similar minutiae pair is chosen based on a measure of similarity between
the local descriptors of the minutiae pair.
Ratha et al. introduced an alignment method for minutiae matching that estimates rotation,
scale, and translation parameters using the Generalized Hough Transform [29]. Given two sets of
points (minutiae), a matching score is computed for each transformation in the discretized set of
all allowed rigid transformations. For each pair of minutiae, one minutia from each image (latent
or reference), and for given scale and rotation parameters, unique translation parameters can be
computed. Each parameter receives “a vote" that is proportional to the matching score for the
corresponding transformation. The transformation that gives the maximum score is considered to
be the best one. In our approach, the alignment is conducted in a similar way, but the evidence for
each transformation parameter is accumulated based on the similarity between the local descriptors
of the two minutiae being matched, with the similarity and descriptor being the ones described in
Section 3.3.2.1.
The descriptor-based Hough transform alignment algorithm takes as input two sets of minutiae,
ML and MR , and two sets of local descriptors CL and CR , one set corresponding to the latent and
one to the rolled print. Each set contains a local descriptor for each minutia. A high level algorithm
of the proposed approach to align two ﬁngerprints given the sets of minutiae and local descriptors
is shown in Algorithm 1, and an illustrative scheme is shown in Fig. 3.13.
3 Refer

to Chapter 4 of [1] for details and published work.

92

Figure 3.13: Illustration of the MCC descriptor-based Hough Transform alignment.4
Given two sets of minutiae, one from the latent and the other from the rolled print being compared, translation and rotation parameters can be obtained for each possible minutiae pair (one
minutia from each set). Let {(xl , yl , θl )} and {(xr , yr , θr )} be the minutiae sets for latent and rolled
prints, respectively, centered at their means. Then, for each pair of minutiae, we have


 θl − θr ,
if θl − θr ≤ 180



∆θ =
θ − θr − 360, if θl − θr > 180
 l



 θ − θr + 360, if θ − θr < −180
l
l
 

 


 ∆x   xr   cos ∆θ sin ∆θ   xl 
−
,
=


yr
∆y
yl
− sin ∆θ cos ∆θ

(3.11)

(3.12)

where ∆x and ∆y are translation parameters in x and y, respectively, and ∆θ is the rotation

parameter.
Since the scale (resolution) is ﬁxed in ﬁngerprint matching (all images have the same resolution), unique translation parameters can be obtained for each pair based on the rotation difference
4 MCC

Local descriptors ﬁgure is extracted from http://biolab.csr.unibo.it/
research.asp?organize=Activities&select=&selObj=81&pathSubj=111%
7C%7C8%7C%7C81&Req=&.
93

Algorithm 1 Descriptor-based Hough Transform.
Input: {ml } = {(xl , yl , θl )} ∈ ML , {mr } = {(xr , yr , θr )} ∈ MR , CL , and CR
Output: A set of 10 rigid transformation matrices
Initialize the accumulator array A
Compute local minutiae descriptor similarity (W ) for every possible minutiae pair using CL and
CR
for all possible minutiae pairs ml , mr do
Compute their direction difference ∆θ = (θr − θl )
if ∆θ < maxθ then
Compute translation parameters (∆x , ∆y ) and increase the voting for this set of alignment
parameters:
A(∆x , ∆y , ∆θ ) = A(∆x , ∆y , ∆θ ) +W (l, r)

(3.10)

end if
end for
Smooth A using a Gaussian low-pass ﬁlter
Find the 10 highest peaks in A
for each peak k do
Compute a rigid transformation between two ﬁngerprints using minutiae pairs that contributed
to peak k and its immediate neighborhood
if the estimated rigid transformation is not reliable (not enough non-colinear points) then
Repeat the voting in peak k and its neighborhood using a reﬁned range
Find the highest peak in the small neighborhood of peak k
end if
end for
between the two minutiae in the pair. The translation and rotation parameters need to be quantized
to the closest bins. After the quantization, evidence is accumulated in the corresponding bin based
on the similarity between the local minutiae descriptors. The assumption here is that true mated
minutiae pairs will vote for very similar sets of alignment parameters, while non-mated minutiae
pairs will vote randomly throughout the parameter space. As a result, the set of parameters that
presents the highest evidence is considered the best one. For robustness, ten sets of alignment
parameters with strong evidence are considered.
In order to make the alignment computationally efﬁcient and also more accurate, we use a
two-step approach to compute the alignment parameters for a ﬁngerprint pair. The ﬁrst step is to
perform the voting using the Descriptor-based Hough Transform. If the bins are too small, the true
peak in the Hough Transform space is not likely to receive sufﬁcient votes. On the other hand, if
94

the bins are too large, they will not provide accurate alignment parameters because of the possible
non-linear deformation. The strategy we adopted is to keep the bins relatively large (10 pixels),
and to include a second step to compute reliable alignment parameters. This second step consists
of using the minutiae pairs that vote for a peak to compute a rigid transformation between the latent
and rolled ﬁngerprints. The use of voting minutiae pairs to compute the transformation gives more
accurate alignment parameters than directly using the peak parameters. In cases where a rigid
transformation matrix cannot be reliably obtained, the voting is repeated inside a neighborhood of
the corresponding peak, but with a smaller bin size (2 pixels). A peak is chosen from this reﬁned
Hough Transform space, and used as the alignment parameters.

3.4.2.2 Similarity Measure
For each of the 10 different alignments, a matching score between the latent and the rolled ﬁngerprints is computed by comparing minutiae and orientation ﬁelds. The maximum value of the
10 scores corresponding to the 10 different alignments is chosen as the ﬁnal matching score between the latent and rolled ﬁngerprints. The details for computing matching scores of minutiae and
orientation ﬁeld are given below, and the ﬁnal score computation scheme is shown in Fig. 3.14.
To compute minutiae matching score under a given alignment, we ﬁrst ﬁnd the corresponding
minutiae pairs (one in the latent, one in the rolled print). For this purpose, we align the minutiae
sets of the two ﬁngerprints and then ﬁnd an one-to-one matching5 between the two minutiae sets
using a greedy algorithm. For each minutia ml in the latent, a set of candidate minutiae in the
rolled print is found. A minutia mr in the rolled print is called a candidate if it has not yet been
matched to any minutia, and both its location and angle are sufﬁciently close to ml . The threshold
values TS for spatial distance and TA for angle distance were determined empirically to measure
the proximity between two minutiae. Among all candidates, the one closest to ml in location is
chosen as the matching minutia of ml .
5 One-to-one matching

means that each minutia in the latent is matched to at most one minutia
in the rolled print, and vice versa.

95

Figure 3.14: Final score computation scheme combining minutiae and orientation ﬁeld information.
After the corresponding minutiae are found, we compute a matching score between the latent
and the rolled print. Suppose that n pairs of matching minutiae between the latent and the rolled
print are found. The minutiae matching score SM between the two ﬁngerprints is given by
SM =

1 n
∑ s (i)s (i),
N i=1 C S

(3.13)

where sC (i) denotes the similarity between the minutia cylinder codes of the ith pair of matched
minutiae, sS (i) = 1 − dS (i)/2TS maps the spatial distance dS (i) of the ith pair of matched minutiae
into a similarity score, and N denotes the number of minutiae in the latent.
According to equation (3.13), the matching score depends on the number of matching minutiae,
which itself is affected by the distance threshold TS . However, due to large distortion present in
many latents, it is difﬁcult to choose an appropriate value for TS . While a large threshold value
will lead to more matching minutiae for distorted mated pairs, the number of matching minutiae
for non-mated pairs will increase as well. Hence, we use two different values (15 pixels and 25
pixels) for TS and for each threshold, a set of matching minutiae is found and a matching score
96

is computed using equation (3.13) at a ﬁxed TA of 20◦ . The mean of the two scores is used as
the minutiae matching score. Fig. 3.15 shows an example in which the score of the genuine pair
is slightly reduced when the smaller threshold is used compared to the larger threshold, while the
score of the latent and the rank-1 non-mate6 using the larger threshold is greatly reduced when the
smaller threshold is used.
We use a simple orientation ﬁeld matcher that basically measures the consistency of the orientation ﬁeld differences between the latent and rolled prints. If we use the Euclidean distance,
for example, to measure the orientation ﬁeld differences, a small error in the rotation alignment
between the latent and the rolled will contribute a small amount to the orientation ﬁeld difference
for every block being compared, resulting in a large overall difference or small similarity score.
In [33], the authors proposed a distance measure for orientation ﬁeld matching that can handle
small rotation alignment errors. Given the aligned latent orientation ﬁeld OL and the rolled orientation ﬁeld OR , each containing K blocks, namely OL (k) and OR (k), the similarity between the
two orientation ﬁelds is given by
SO =

| ∑K vk e j2(OL (k)−OR (k)) |
k=1
∑K vk
k=1

,

(3.14)

where vk is 1 if both corresponding blocks k are valid, and 0 otherwise.
The overall matching score between the latent and rolled prints based on minutiae matching
and orientation ﬁeld matching is given by
S = (1 − w0 )SM + w0 SO ,

(3.15)

where the weight w0 is empirically set as 0.4. Fig. 3.16 shows an example in which the fusion of
minutiae matching and orientation ﬁeld matching scores helps improve the retrieval rank7 of the
true mate due to the higher orientation ﬁeld matching score between the latent and the true mate
orientation ﬁelds, compared to the score between the latent and the non-mate orientation ﬁelds.
6 The

rank-1 non-mate refers to the non-mated rolled print whose match score with the latent
ranks ﬁrst among all the rolled prints in the database.
7 Retrieval rank of a rolled ﬁngerprint refers to its rank in the whole candidate list which is
sorted in the decreasing order of matching scores with the latent.
97

(a)

(b)

(d) 0.152

(c)

(e) 0.171

Figure 3.15: Latent print for which the matching score of the genuine pair is slightly reduced when
a smaller threshold value is used compared to a larger threshold value, while impostor score is
greatly reduced. (a)-(c) show the latent, the true mate, and the rank-1 non-mate according to large
threshold, respectively. (d)-(g) show latent minutiae that were matched to rolled print minutiae
in the following cases: (d) true mate using small threshold, (e) true mate using large threshold,
(f) non-mate using small threshold, and (g) non-mate using large threshold. In (d)-(g), the scores
corresponding to each case are included. Filled-in minutiae indicate the matching minutiae.
The retrieval rank of the true mate improved from 2 to 1 after the fusion, while the retrieval rank
of the rank-1 non-mate according to minutiae matcher was changed from 1 to 3 after the fusion.

98

Fig. 3.15 (cont’d)

(f) 0.113

(g) 0.172

99

(a)

(b)

(c)

(d)

(e)

(f)

(g)

Figure 3.16: Latent print identiﬁed at a higher rank after fusing minutiae matching scores with
orientation ﬁeld matching scores. The rank of the true mate was improved from 2 to 1 after the
fusion, and the rank of the highest ranked non-mate was 3 after the fusion. (a)-(c) show minutiae
and the image of (a) a latent, (b) its true mate, and (c) the highest ranked non-mate according to
minutiae matching. (d) and (f) show latent minutiae and orientation ﬁeld (in blue) aligned with
minutiae and orientation ﬁeld of the true mate. (e) and (g) show latent minutiae and orientation
ﬁeld (in blue) aligned with minutiae and orientation ﬁeld of the rank-1 non-mate.

100

3.5 Databases
Matching experiments were conducted on two different latent ﬁngerprint databases: NIST Special
Database 27 (NIST SD27) and West Virginia University Latent Fingerprint Database (WVU LFD).
In this section, we present some characteristics of these two latent databases.

3.5.1 NIST Special Database 27 (NIST SD27)
NIST Special Database 27 is the only publicly available database comprising latent ﬁngerprints
from operational scenarios (latents collected at crime scenes). It consists of 258 latent ﬁngerprint
images and 258 corresponding (mated) rolled prints. Both latents and rolled prints are available at
500 and 1000 ppi – we used 500 ppi for consistency with the other databases. The quality of the
latents in NIST SD27 varies, reﬂecting the operational (casework) quality.
NIST SD27 contains latent prints of three different qualities, termed “good”, “bad”, and “ugly”,
which were classiﬁed by latent examiners. Some examples of latents from those three qualities are
shown in Fig. 3.4. Although this classiﬁcation of latent prints as “good”, “bad”, and “ugly”
is subjective, it has been shown that such a classiﬁcation is correlated with the latent matching
performance [28].
Another indicator of ﬁngerprint quality that affects the matching performance is the number of
minutiae in the latent print [28]. In other words, while the latent may have clear and sharp edges,
it may have relatively few minutiae available for matching. Based on the number of minutiae n
in latents in NIST SD27, Jain and Feng [28] classiﬁed latents in NIST SD27 into three groups:
large number of minutiae (n ≥ 22), medium number of minutiae (13 < n ≤ 21), and small number
of minutiae (n ≤ 13), containing 86, 85, and 87 prints, respectively. We present our experimental
results for each of the six quality groups based on subjective quality (good, bad and ugly) and the
number of minutiae.
We use manually marked minutiae – provided with NIST SD27 – as features in latent ﬁngerprints. For rolled ﬁngerprint images, the minutiae are automatically extracted using the two

101

commercial matchers.

3.5.2 West Virginia University Latent Database (WVU LFD)
West Virginia University Latent Database8 consists of 449 latent ﬁngerprint images collected in a
laboratory environment and 4, 740 rolled prints, including the 449 mated rolled prints of the 449
latents. The latent images in this database are at 1000 ppi, and they were converted to 500 ppi for
our experiments so that all the databases used in our experiments had the same resolution. Fig.
3.17 shows a latent with its corresponding rolled print in the WVU latent database. Similar to
NIST SD27 database, manually marked minutiae were also provided with these latents. Minutiae
were automatically extracted from the rolled prints using the two commercial matchers.

Figure 3.17: A latent and its corresponding rolled print in the WVU latent database. The NFIQ
quality [9] of the rolled print is 4, which is one above the worst NFIQ quality.
There is no subjective quality value assigned to the latents in the WVU database. As mentioned
earlier, one way to assign an objective quality measure to a latent is based on the number of
minutiae in the latent, so any latent can be assigned an objective quality. If we apply the same
objective quality classiﬁcation scheme as in NIST SD27 to WVU database, we obtain 208, 80, and
161 latent ﬁngerprints in the objective qualities of large, medium, and small number of minutiae,
respectively.
8 To

request WVU latent ﬁngerprint database, please contact Dr.
Arun A. Ross
(http://www.csee.wvu.edu/∼ross/) at Integrated Pattern Recognition and Biometrics Lab
(http://www.csee.wvu.edu/∼ross/i-probe/).

102

The two latent databases used in our experiments, NIST SD27 and WVU, have different characteristics: most of the latent images in NIST SD27 contain signiﬁcant background noise, while
in WVU latent images, collected in a laboratory environment, there is a uniform background in
most latents. However, overall, based on NFIQ quality [9], the quality of the rolled prints in WVU
database is worse than the quality of rolled prints in NIST SD27. Fig. 3.18 shows the histograms
of NFIQ values [9], one of the very well recognized measure of ﬁngerprint quality, of the rolled
prints which have corresponding latents in NIST SD27 and in WVU databases (258 and 449 rolled
prints, respectively). NFIQ deﬁnes ﬁve quality levels in the range [1, 5] with 1 indicating the highest quality. The overall worse rolled print quality of WVU database compared to NIST SD 27
could be explained because in the operational database such as NIST SD27, rolled prints were
captured by experienced law enforcement ofﬁcers which, to our understanding, is not the case for
the WVU database. Further, if the rolled prints corresponding to the latents are of poor quality,
the number of mated minutiae will be small and, therefore, it would be much more challenging to
identify the true mates of the latents at rank-1.

Frequency of rolled prints

0.8

NIST SD27
WVU

0.6
0.4
0.2
0

1

2
3
4
NFIQ Quality Levels

5

Figure 3.18: Histograms of NFIQ values of rolled prints in NIST SD27 and WVU databases.

103

3.6 Experimental Results
3.6.1 Latent Matching with Enhanced Images
Matching experiments were conducted on the NIST Special Database 27, which consists of 258
latent ﬁngerprint images and 258 mated rolled ﬁngerprint images. The manually marked features in the latents in this database are region of interest (ROI), minutiae, visible singular points
(inside the ROI) and “virtual” singular points (outside the ROI). To make the matching problem
more challenging and realistic, the background database (gallery) was increased from 258 mated
rolled ﬁngerprints to 27, 258 total rolled ﬁngerprints by adding 27, 000 ﬁngerprint images from
the database NIST SD14 [90]. For the rolled ﬁngerprint images, only minutiae were needed for
matching and they were automatically extracted using Veriﬁnger [79]. For boosted max fusion, the
transformations were computed based on the matched minutiae output by Veriﬁnger for each pair
of ﬁngerprints being matched.
The latent images from NIST SD27 were enhanced using the approach described in [115]. In
addition, since latent images are usually of poor quality, ridge frequency values were estimated for
each image block [115], and the median of those values was chosen as the ridge frequency for the
entire image. We also evaluated the use of a constant ridge frequency value for all the images, and
different ridge frequency values in each block of the image. The median ridge frequency value
provided the best performance and therefore the results shown here are based on it.
The matching performance of enhanced latent images using reconstructed orientation ﬁeld
against rolled ﬁngerprints is much better than the matching performance using minutiae automatically extracted from the original latent image. While the matching performance of enhanced images is worse than the performance of manually marked minutiae, this performance is comparable
to the matching performance of enhanced images using manually marked orientation ﬁeld, which
requires signiﬁcantly more manual labor. Fig. 3.19 shows the Cumulative Matching Characteristic curves on NIST SD 27 for manually marked minutiae, enhanced images using reconstructed
orientation ﬁeld, automatically extracted minutiae from latent image, and enhanced images using

104

manually marked orientation ﬁeld. Minutiae in the latents were either manually marked (MMM)
or automatically extracted by Veriﬁnger 4.2 [79].
70

Identification Rate (%)

60
50
40

Manually Marked Minutiae (MMM)
Latent Enhanced with ROF (Veri)
Latent image (fully automatic)
Latent Enhanced with OF (Veri)

30
20
10
0

20

40

Rank

60

80

100

Figure 3.19: CMC curves for manually marked minutiae, enhanced image using reconstructed
orientation ﬁeld, automatically extracted minutiae by Veriﬁnger from latent image and enhanced
image using manually marked orientation ﬁeld.

3.6.1.1 Fusion methods
We performed experiments using different fusion levels and methods. These fusion scenarios always used manually marked minutiae-based information and enhanced images using reconstructed
orientation ﬁeld-based information (scores, ranks, etc). We found that Borda count (rank-level)
could not improve the matching performance of manually marked minutiae. This might be because Borda count method assumes all the matchers perform equally well, which is not true in
our case involving two matchers. However, the highest rank fusion improved the matching performance most of the time, as can be seen in Fig. 3.20. The highest rank fusion explores the strength
of each matcher more effectively, since being assigned a high rank from only one of the matchers
increases the likelihood of receiving a high rank after the second sorting.

105

70

Identification Rate (%)

60

50

40

30

20
0

20

Manually Marked Minutiae
Enhanced with ROF
Highest rank fusion
Borda count fusion
40
60
80
100
Rank

Figure 3.20: CMC curves for manually marked minutiae, enhanced image using reconstructed
orientation ﬁeld, highest score rank-level fusion and Borda count fusion.
Min and product fusion rules do not improve the matching performance while the Max rule
provides very minor improvement. The best results were obtained using the weighted sum rule
and boosted max fusion (Fig. 3.21). At rank 1, the improvement due to boosted max fusion is
approximately 10% compared to the two individual matchers over all latent image quality levels
as well as for each speciﬁc quality level separately. This means boosted max fusion was able to
correctly ﬁnd the mates of 25 additional latents than using manually marked minutiae alone. In
the case of latent search, since the AFIS accuracy is not sufﬁciently high, the output is a list of N
candidates for manual comparison by a latent examiner.
For good quality latent images, although the boosted max method performed better than manually marked minutiae in the top few ranks, its performance dropped for some ranks (Fig. 3.21b).
This might indicate that the best approach for good quality latents is the simple sum rule. For bad
quality images, the boosted max method outperformed manually marked minutiae by an average
of 10% at all ranks (Fig. 3.21c). In the ugly quality images, boosted max fusion shows consistent
improvement at all ranks compared to manually marked minutiae, and performs better for the ﬁrst
30 ranks compared to the sum rule (Fig. 3.21d).

106

70

Identification Rate (%)

65
60
55
50
45
40
35
0

Manually Marked Minutiae (MMM)
Weighted sum (MMM + enhanced with ROF)
Boosted max (MMM and Enhanced with ROF)
20
40
60
80
100
Rank

(a) All quality latents
85

Identification Rate (%)

80
75
70
65
60
55
0

Manually Marked Minutiae (MMM)
Weighted sum (MMM + enhanced with ROF)
Boosted max (MMM and Enhanced with ROF)
20
40
60
80
100
Rank

(b) Good quality latents
Figure 3.21: CMC curves for score level fusion for latents of all qualities.
As discussed in Section 3.4.1.2, we modiﬁed the boosted max approach to include a second
weight. Therefore, two different parameters (weights w1 and w2 ) must be speciﬁed in order to

107

Fig. 3.21 (cont’d)
70

Identification Rate (%)

65
60
55
50
45
40
35
0

Manually Marked Minutiae (MMM)
Weighted sum (MMM + enhanced with ROF)
Boosted max (MMM and Enhanced with ROF)
20
40
60
80
100
Rank

(c) Bad quality latents

Identification Rate (%)

60

50

40

30

20
0

Manually Marked Minutiae (MMM)
Weighted sum (MMM + enhanced with ROF)
Boosted max (MMM and Enhanced with ROF)
20
40
60
80
100
Rank

(d) Ugly quality latents
apply the boosted max. These two weights were empirically determined and they do not need to
sum to one because the purpose is to boost the scores of image pairs that are found to be consistent.
The computation of the transformation matrices used to decide the consistency between a pair of
108

ﬁngerprints in boosted max approach is based on the matched minutiae output by VeriFinger.
Ideally, TME should be the Identity matrix because the two sets of minutiae are extracted from the
same image. However, in practice manually markings and automatically extracted minutiae differ
in their positions, which leads to a transformation matrix near to the Identity matrix, but not exactly
the same.
We observed that in almost all the cases, if the mated rolled ﬁnger was ranked ﬁrst in one of the
matching experiments (manually marked minutiae or enhanced images), it was also ranked ﬁrst in
the boosted max approach. Fig. 3.22a shows an example of this case where the latent was ranked
1 by manually marked minutiae, ranked 2, 403 by enhanced image, and ranked 1 by boosted max.

(a)

(b)
Figure 3.22: Matched minutiae shown for manually marked, rolled and enhanced latent images.
(a) Boosted max retrieved the true mate for this latent at rank one even though one of the retrieved
ranks (enhanced image) was 2, 403 and (b) the retrieved rank for the true mate for this latent is
rank one after boosted max was applied even though neither the manually marked minutiae nor
enhanced image retrieved the true mate at rank one.

109

Boosted max ranked true mated rolled ﬁngerprints at a lower rank in only two cases that were
ranked ﬁrst by manually marked minutiae (two good quality images). It also corrected the retrieved
rank in eleven cases (4 good quality, 3 bad and 4 ugly quality latents). This means the mated rolled
was neither ranked ﬁrst in manually marked minutiae nor in enhanced images, but it was ranked
ﬁrst after boosted max fusion. Fig. 3.22b shows an example of a latent that was ranked as 268 by
manually marked minutiae, ranked 2 by enhanced image, and ranked 1 by boosted max.

3.6.2 Latent Matching with Descriptor-Based Hough Transform
We ﬁrst provide a description of the baseline algorithms to be compared with the proposed algorithm. Then we report the performances of alignment and matching. This is followed by the fusion
of matchers and the effect of ﬁngerprint quality. Then, we discuss the issue of computational cost.

3.6.2.1 Commercial Matchers
In order to compare the performance of the proposed latent ﬁngerprint matcher, we used two
commercial ﬁngerprint matchers, Veriﬁnger 4.2 [79] and Morpho as baseline. In addition, we also
used the algorithm presented in [91, 118] as a baseline, for which the SDK was provided by the
authors (MCC SDK). It should be pointed out that none of these three baseline matchers were
designed speciﬁcally for the latent matching case. But, despite our efforts, we could not ﬁnd any
latent ﬁngerprint matching SDK or a forensic AFIS that is available for evaluation purposes by a
research lab. Still, the matchers we are using in our comparative study are well known: one of
the COTS (VeriFinger) [79] has been widely used as a benchmark in ﬁngerprint publications, and
MCC is one of the best performing algorithms in FVC-onGoing [119].

3.6.2.2 Alignment Performance
In order to estimate the alignment error, we use ground truth mated minutiae pairs from NIST
SD27, which are marked by ﬁngerprint examiners, to compute the average distance between the

110

true mated pairs after alignment9 . If the average Euclidean distance for a given latent is less than
a pre-speciﬁed number of pixels in at least one of the ten best alignments (peaks in the DescriptorBased Hough Transform), then we consider it a correct alignment. This alignment performance is
shown in Fig. 3.23 for the NIST SD27 latent database. The x-axis shows the misalignment threshold10, and the y-axis shows the percentage of correctly aligned latent ﬁngerprints in at least one of
the ten top alignments. For comparison, we show the accuracy of aligning the minutiae sets based
on the peaks of the Generalized Hough Transform (GHT) and based on the most similar minutiae
pair (according to the MCC similarity)11. Two latent alignment examples are given in Figs. 3.24
and 3.25 to show the alignment results by DBHT and GHT. As we can see from these ﬁgures, the

Percentage of Correctly Aligned Latents

proposed algorithm is superior to GHT for difﬁcult latents where the number of minutiae is small.
100
95
90
85
80
75
70
65
60
55

10

Generalized Hough Transform
MCC Most Similar Pair
Descriptor−Based Hough Transform
15
20
25
30
Misalignment Threshold (in pixels)

Figure 3.23: Alignment Accuracy: percentage of correctly aligned latents vs. misalignment threshold.
9 Here

we use the term ground truth minutiae to refer to minutiae which are marked by latent
examiners by looking at the latent and the corresponding rolled print at the same time, and we use
the term manually marked minutiae to refer to minutiae which are also marked in the latent by
latent examiners, but without looking at the true mate (rolled print).
10 The alignment is deemed as incorrect if the average distance between mated minutiae pairs
after alignment is larger than this threshold.
11 In this case, each alignment is based on one of the ten most similar minutiae pairs.

111

(a)

(b)

(c)

(d)

Figure 3.24: Example in which DBHT (Descriptor-Based Hough Transform) alignment is better
than GHT (Generalized Hough Transform) alignment. (a) Latent with manually marked minutiae,
(b) corresponding rolled print with automatically extracted minutiae, (c) rolled print with latent
minutiae aligned by GHT, and (d) aligned by DBHT.
There are very few errors in alignment if we set the threshold value of misalignment as 20
pixels. One of the reasons for these failure cases is there are a very small number of true mated
minutia pairs in the overlapping area between the latent and mated rolled print. As a result, not

112

(a)

(b)

(c)

(d)

Figure 3.25: Example in which DBHT (Descriptor-Based Hough Transform) alignment is better
than GHT (Generalized Hough Transform) alignment. (a) Latent with manually marked minutiae,
(b) corresponding rolled print with automatically extracted minutiae, (c) rolled print with latent
minutiae aligned by GHT, and (d) aligned by DBHT.
many true mated pairs vote for the correct alignment parameters. The absence of true mated pairs
is due to a limited number of minutiae in latents and the error in minutiae detection in the rolled
print. One such example is shown in Fig. 3.26. Blue squares represent manually marked minutiae

113

in the latent print (left), red squares represent automatically extracted minutiae in the rolled print
(right), and the green line indicates the only true mated minutiae pair available for this (latent,
rolled) image pair.

Figure 3.26: Example of alignment error due to the small number of true mated minutia pairs in
the overlapping area between a latent and its mated rolled print. Note that there is only one aligned
minutiae pair here.

3.6.2.3 Matching Performance
In the identiﬁcation scenario, the size of the background database (or gallery) signiﬁcantly affects
the identiﬁcation accuracy. Therefore, to make the problem more challenging and realistic, we
built a large background database of rolled prints containing 258 mated rolled prints from NIST
SD27, 4, 740 rolled prints from WVU database, and 27, 000 rolled prints from the NIST Special
Database 14 [90]. Therefore, the total number of rolled prints in the background database is 31, 998
from a combination of the rolled prints in the three databases.
Minutia Cylinder Code (MCC) is used as the local descriptor for minutiae in our experiments.
The local descriptors are built using MCC SDK, which uses the bit-based implementation (binary
descriptors) [118]. The parameters used for MCC are set as suggested in [118], with the number

114

of cells along the cylinder diameter as 16 (Ns ). In our method, the local descriptor similarities are
used in both the alignment and scoring modules.
70

Identification Rate (%)

65
60
55
50
45
40
35
0

Morpho (MMM and Morpho features)
MCC SDK (MMM and Morpho features)
Proposed Matcher (MMM and Morpho features)
5
10
15
20
Rank

(a) NIST SD27
70

Identification Rate (%)

65
60
55
50
45
40
35
0

Morpho (MMM and Morpho features)
MCC SDK (MMM and Morpho features)
Proposed Matcher (MMM and Morpho features)
5
10
15
20
Rank

(b) WVU LFD
Figure 3.27: Performance of Morpho, MCC SDK, and Proposed Matcher when the union of manually marked minutiae (MMM) extracted from latents and automatically extracted minutiae by
Morpho from rolled prints is input to the matchers.
Our matcher and MCC SDK both take minutiae as input. In the latent cases, we use man115

ually marked minutiae. For the rolled prints, we used both the COTS to extract minutiae. The
performance of the proposed matcher using minutiae extracted from rolled prints using Morpho
is slightly worse on the NIST SD27 database compared to the performance using minutiae extracted using Veriﬁnger; however, for WVU LFD, using Morpho minutiae yielded a signiﬁcantly
better performance compared to the performance using minutiae extracted using Veriﬁnger. This
demonstrates that the performance of COTS can be signiﬁcantly affected by the image quality.
Overall, since minutiae extracted from Morpho resulted in a better performance, we only report
the results in which minutiae are extracted using Morpho. Fig. 3.27 shows the performance of
Morpho, MCC SDK, and the proposed matcher using manually marked minutiae in latents and
automatically extracted minutiae by Morpho in rolled prints. The proposed approach outperforms
the other ﬁngerprint matchers used in our study.
In our earlier study [120], we did not use the bit-based implementation of MCC representation
and we only evaluated the algorithm on NIST SD27. We used our own implementation of MCC
with the parameters suggested in [91]. However, the bit-based implementation and the more optimized parameters suggested in [118] yielded a much better performance on WVU LFD, and a
decrease in performance on the NIST SD27 database. We decided to use the bit-based implementation with the new parameters because of its better overall performance, and because we wanted
to use a consistent framework for both the databases.
It is worth noting that the matching performance on WVU LFD when manually marked minutiae are used is generally worse than the performance on NIST SD27. We believe this is due to
a number of factors: (i) there are 14 latents with less than 3 manually marked minutiae in WVU
LFD, while the minimum number of manually marked minutiae in NIST SD27 latents is 5; (ii)
while the genuine (latent, rolled) pairs were provided with the database, after we examined the images in the WVU database we identiﬁed some that appeared to be wrongly paired; (iii) the quality
of the mates (rolled prints) is slightly worse in WVU LFD than in NIST SD27. We did not exclude
any of the latents or (latent, rolled) mated pairs from the WVU database (from cases (i) and (ii)) to
allow for a fair comparison by other researchers with our results.

116

The performance of the COTS matchers, each using its own proprietary templates for latents
(including automatically extracted minutiae and possibly other features), is worse than using manually marked minutiae for both the databases. However, the gap between the performances of
manually marked minutiae and of proprietary template is much larger in the case of NIST SD27
than in the case of WVU latent database. This is probably due to the characteristics of these two
databases. Note that WVU is a laboratory collected database and so most of the latents in it do
not contain background noise. On the other hand, in NIST SD27 the images are of operational
casework quality and most of the latents contain a large amount of background noise, which poses
a challenge in automatic feature extraction. Fig. 3.28 shows the performance of the two COTS
matchers using both manually marked minutiae and proprietary templates (automatically extracted
minutiae) for NIST SD27 and WVU databases.
There have been several studies on latent matching reported in the literature. Almost all of them
are based on NIST SD27. Table 3.1 shows the reported results on the matching performance for
NIST SD27 database. There is no reported performance on the WVU latent database. It should be
noticed that most of the reported results cannot be directly compared mainly because of two factors:
(i) the amount of input information related to the latent ﬁngerprint, which could be automatically
extracted features, or manually marked features such as minutiae, singular points, quality map, etc,
or a combination of both; and (ii) some differences in the composition of the background databases
and their size. In Table 3.1 we show the reported rank-1 accuracy, the manually marked latent
features used in each method, and the size of the background database used. One of the results
that could be almost directly compared to our results is the reported rank-1 accuracy (34.9%) in
[28] where only manually marked minutiae was used as input, which is the same scenario as in our
proposed matcher. The proposed matcher achieves a signiﬁcantly higher rank-1 accuracy of 53.5%
with similar background database size and images as in [28]. The other approach presented here
(Paulino et al. [87]) can also be compared to this proposed matcher, although the former requires
additional manually marked input features beyond minutiae.

117

Identification Rate (%)

60
50
40
30

Verifinger (Proprietary Template)
Morpho (Proprietary Template)
Verifinger (MMM and Verifinger features)
Morpho (MMM and Morpho features)

20
10
0

5

10
Rank

15

20

(a) NIST SD27

Identification Rate (%)

60
50
40
30
20
10
0

Verifinger (Proprietary Template)
Morpho (Proprietary Template)
Verifinger (MMM and Verifinger features)
Morpho (MMM and Morpho features)
5
10
15
20
Rank

(b) WVU LFD
Figure 3.28: Performance comparison using manually marked minutiae (MMM) and automatically
extracted minutiae from latents.

Fig. 3.29 shows two examples of latent prints in WVU LFD correctly identiﬁed at rank-1 by
12 SP:

singular points, ROI: region of interest, RQM: ridge quality map, RFM: ridge ﬂow map,
RWM: ridge wavelength map.
118

Table 3.1: Comparison of rank-1 accuracies reported in the literature for the NIST SD27 database.
Work

Manually Marked

Background

Rank-1

Latent Features

Database Size

Accuracy (%)

Jain and Feng [28]

Minutiae, Skeleton, SP,
ROI, RQM, RFM, RWM12

29, 257

74.0

Proposed Matcher [88] +
Morpho

Minutiae

31, 998

57.4

Proposed Matcher [88]

Minutiae

31, 998

53.5

Paulino et al. [87]

Minutiae, SP, ROI12

27, 258

48.0

Jain and Feng [28]

Minutiae

29, 257

34.9

27, 258

26.0

27, 258

25.0

ROI12

Yoon et al. [101]

SP,

Feng et al. [102]

None

the proposed matcher. Even though the number of minutiae in these latents is small, they could
still be correctly identiﬁed. The ranks of the true mates using Morpho matcher are 1871 and 181,
respectively, and the matched minutiae shown are from our proposed matcher.
Fig. 3.30 shows examples of latent prints in NIST SD27 and in WVU LFD whose mated reference prints are not included in the top 20 candidates by the proposed matcher, but were correctly
identiﬁed at rank-1 by Morpho matcher. The ranks of these latents using the proposed matcher
are 3626 and 64, respectively, and the matched minutiae shown are from our proposed matcher. In
the ﬁrst latent, a large number of minutiae do not have mated minutiae due to missing minutiae in
the rolled print, and therefore the score is not as high as for impostor pairs in which many more
minutiae could be matched. In the second case, we can see that the minutiae marked in the latent are relatively sparse, while the minutiae automatically extracted in the rolled print are denser.
These facts make local neighborhoods (and descriptors) very different between the latent and its
true mate, leading to a low match score.

3.6.2.4 Fusion of Matchers
We noticed that the two most accurate matchers (the proposed matcher and Morpho) perform differently on different latents, meaning they are complementary to each other. This suggests that a

119

(a)

(b)
Figure 3.29: Two latent prints from the WVU database correctly identiﬁed at rank-1 by the proposed matcher but ranked below 20 by Morpho.
fusion of these two matchers would result in a better performance. We performed a score-level
fusion of these two matchers. The scores from Morpho matcher were normalized to the range
[0, 1] for each latent (local min-max normalization) because local normalization was shown to perform better than global normalization in the identiﬁcation scenario [121]. Although the proposed
matcher and Morpho matcher have similar matching accuracy, the fusion weights selected (0.8
and 0.2) were not equal because of the large range of the scores for the Morpho matcher. The
performance improvement obtained by the score-level fusion of Morpho matcher and the proposed
120

(a)

(b)
Figure 3.30: Latent prints ((a) from NIST SD27 and (b) from WVULFD) whose mates were not
retrieved in the top 20 candidates by the proposed matcher but correctly identiﬁed at rank-1 by
Morpho matcher.
matcher is shown in Fig. 3.31 for both the databases. Some latents for which the fusion of the two
matchers (Morpho and proposed matcher) improved the ranks of the true mates compared to the
retrieval ranks by the individual matchers separately are shown in Figs. 3.32 and 3.33. Note that
like those mated pairs (shown in Fig. 3.29 and Fig. 3.30) identiﬁed at rank-1 by either one the two
matchers, mated pairs (shown in Fig. 3.32 and Fig. 3.33) which both matchers failed to identify at
rank-1 also beneﬁt from the fusion. The reason is the scores of non-mated pairs given by the two
matchers are not consistent.
Improvements in matching accuracy were also obtained by combining the proposed matcher

121

75

Identification Rate (%)

70
65
60
55
50
45
40
0

Morpho (MMM and Morpho features)
Proposed Matcher (MMM and Morpho features)
Proposed Matcher + Morpho
5
10
15
20
Rank

(a) NIST SD27
75

Identification Rate (%)

70
65
60
55
50
45
40
0

Morpho (MMM and Morpho features)
Proposed Matcher (MMM and Morpho features)
Proposed Matcher + Morpho
5
10
15
20
Rank

(b) WVU LFD
Figure 3.31: Score-level fusion of the proposed matcher and Morpho for NIST SD27 and WVU
databases.
and other matchers in our study (Veriﬁnger and MCC SDK), but they are not reported here because
the fusion performance with Morpho was consistently better than the performance of Veriﬁnger
and of MCC SDK. We also performed rank-level fusion using the highest rank and Borda Count

122

Table 3.2: Rank-1 accuracies for various subjective quality levels of latents in NIST SD27.
Quality

Veriﬁnger (%)

Morpho (%)

MCC (%)

Proposed (%)

All

38.0

47.3

42.6

53.5

Good

55.7

70.5

69.3

75.0

Bad

36.5

36.5

31.8

47.1

Ugly

21.2

34.1

25.9

37.6

Table 3.3: Rank-1 accuracies for various objective quality values of latents in NIST SD27 (large,
medium and small refer to the number of minutiae in the latent).
Quality

Veriﬁnger (%)

Morpho (%)

MCC (%)

Proposed (%)

All

38.0

47.3

42.6

53.5

Large

59.3

73.3

70.9

75.6

Medium

43.5

45.9

43.5

56.5

Small

11.5

23.0

13.8

28.7

methods [122]. However, since score-level fusion showed a better performance, we only report
here results for score-level fusion.

3.6.2.5 Effect of Fingerprint Quality
In Section 3.5, we discuss how the quality of the latent ﬁngerprints can be measured subjectively
(assigned by latent experts as in NIST SD27) and objectively (based on the number of minutiae
available). Rank-1 accuracies are shown for each quality separately in Tables 3.2, 3.3, and 3.4
for both the latent databases. We can see that the matching performance is highly correlated with
the number of minutiae available in the latent prints. The performance of the proposed matcher is
consistently better over all qualities and for both the databases.
The quality of reference prints also has a large impact on the matching accuracy. In Fig. 3.18,
the histograms of NFIQ quality values for the corresponding rolled prints in each latent database
are shown. According to the NFIQ quality measure, the quality of the rolled prints in WVU
database is slightly worse than the quality of the rolled prints in NIST SD27. The NFIQ quality
measure is an integer value in the range 1 to 5, where 1 is the highest quality and 5 is the worst

123

Table 3.4: Rank-1 accuracies for various objective quality values of latents in WVU LFD (large,
medium and small refer to the number of minutiae in the latent).
Quality

Veriﬁnger (%)

Morpho (%)

MCC (%)

Proposed (%)

All

35.6

45.4

44.3

47.9

Large

63.5

73.1

74.0

74.5

Medium

28.8

45.0

37.5

45.0

Small

3.1

9.9

9.3

14.9

Table 3.5: Rank-1 accuracies for latents grouped according to NFIQ quality values of corresponding rolled prints in NIST SD27.
Quality

Veriﬁnger (%)

Morpho (%)

MCC (%)

Proposed (%)

All

38.0

47.3

42.6

53.5

NFIQ ≤ 3

42.1

54.9

49.4

60.4

30.9

34.0

30.9

41.5

NFIQ > 3

Table 3.6: Rank-1 accuracies for latents grouped according to NFIQ quality values of corresponding rolled prints in WVU LFD.
Quality

Veriﬁnger (%)

Morpho (%)

MCC (%)

Proposed (%)

All

35.6

45.4

44.3

47.9

NFIQ ≤ 3

36.9

50.0

48.0

52.4

34.4

41.1

40.6

43.3

NFIQ > 3

quality. We observed a signiﬁcant difference in the matching performance when the latents were
divided into the following two quality groups: (i) rolled prints are of good quality (NFIQ value of
1, 2 and 3), and (ii) rolled prints are of poor quality (NFIQ values of 4 and 5). The difference in
matching performance between good NFIQ and poor NFIQ qualities for all matchers ranges from
11 − 21% for NIST SD27, while it ranges from 2 − 9% for WVU database (see Tables 3.5 and 3.6).
As an example, the rank-1 accuracy of Morpho matcher on NIST SD27 is 54.9% and 34.0% for
good and poor NFIQ quality, respectively.

124

3.6.2.6 Computational Cost
The implementation of our matching algorithm is in Matlab. The average speed of our matcher
running on a PC with Intel Core 2 Quad CPU and Windows XP operating system is around 10
matches per second and it depends on the number of minutiae (the smaller the number of minutiae,
the faster the matching speed). Multi-thread capability was not utilized. The majority of the
running time (70%) is spent matching the local minutiae descriptors. In a C/C++ implementation,
this matching would be much faster than in Matlab because of the nature of the MCC descriptors
(binary). We have not optimized the code for speed.

3.6.2.7 Comparison of the two proposed methods
The manually marked input features are essentially the same for both the proposed methods: minutiae, singular points and region of interest in the enhancement, and minutiae in the descriptor-based
Hough transform approach. The performance of the latent matching approach using descriptorbased Hough Transform is better than the performance of latent matching approach by enhancement. However, the improvement by enhancement has the advantage in terms of computational
efﬁciency, because of the use of the commercial matcher in both steps (matching using manually
marked minutiae and matching using minutiae automatically extracted from enhanced images).
Furthermore, by simply combining the two approaches at the score level, the rank-100 accuracy
increases from 72% to 81% compared to the descriptor-based Hough Transform matching alone
(see Fig. 3.34).

125

(a)

(b)

(c) (5, 10)

(d)

Figure 3.32: Latent print mate from NIST SD 27 identiﬁed at rank 1 after score-level fusion of
Morpho and proposed matcher. The ﬁrst column shows (a) a latent, (c) its true mate, (e) rank-1
non-mate by the proposed matcher, and (g) rank-1 non-mate by Morpho matcher. The second
column shows (b) latent minutiae, (d), (f), and (h) latent minutiae (in blue) aligned by the proposed
matcher to the rolled print minutiae shown in (c),(e) and (g). In (c), (e), and (g), the numbers
in parentheses indicate the ranks at which true mated rolled print was retrieved by the proposed
matcher and Morpho matcher, respectively.
126

Fig. 3.32 (cont’d)

(e) (1, 30848)

(f)

(g) (3617, 1)

(h)

127

(a)

(b)

(c) (2, 42)

(d)

Figure 3.33: Latent print mate from WVU LFD identiﬁed at rank 1 after score-level fusion of
Morpho and proposed matcher. The ﬁrst column shows (a) a latent, (c) its true mate, (e) rank-1
non-mate by the proposed matcher, and (g) rank-1 non-mate by Morpho matcher. The second
column shows (b) latent minutiae, (d), (f), and (h) latent minutiae (in blue) aligned by the proposed
matcher to the rolled print minutiae shown in (c),(e) and (g). In (c), (e), and (g), the numbers
in parentheses indicate the ranks at which true mated rolled print was retrieved by the proposed
matcher and Morpho matcher, respectively.

128

Fig. 3.33 (cont’d)

(e) (1, 29408)

(f)

(g) (48, 1)

(h)

129

85

Identification Rate (%)

80
75
70
65
60
55
50
45
0

20

40

Rank

Enhancement
DBHT
Fusion by sum rule
60
80
100

Figure 3.34: Score-level fusion of the two proposed approaches: by enhancement and using
descriptor-based Hough Transform (DBHT) matching approach. By combining the two approaches, the rank-100 accuracy is 81%.

130

3.7 Summary and Conclusions
Latent ﬁngerprint matching is a very challenging problem due to the characteristics of latent ﬁngerprints (poor quality, small area, small number of minutiae, noisy background, etc). While a
fully automatic latent matching system is desired, but given the difﬁculty of the problem and the
relatively low matching performance of available AFIS, manual input is still needed. Thus, we
have presented two different methods of improving latent ﬁngerprint matching performance in the
case where manually marked features are used.
We have shown that the performance of manually marked minutiae in latents can be improved
by utilizing automatically extracted minutiae from enhanced latent images. This framework consists of the following steps: (i) reconstruct the orientation ﬁeld based on manually marked minutiae
and singular points; (ii) enhance the latent using median ridge frequency computed in small image
blocks and the reconstructed orientation ﬁeld; (iii) match enhanced latents and rolled ﬁngerprints;
(iv) combine the scores from two matchers using boosted max fusion. This framework improved
the latent matching performance with a large background database irrespective of latent quality.
We have also presented a ﬁngerprint matching algorithm designed for matching latents to
rolled/plain ﬁngerprints which is based on a descriptor-based Hough Transform alignment. A
comparison between the alignment performance of the proposed algorithm and the well-known
Generalized Hough Transform shows the superior performance of the proposed method. We also
reported matching results for two different latent ﬁngerprint databases with a large background
database of about 32K rolled prints. We compared the performance of the proposed matcher with
three different state-of-the-art ﬁngerprint matchers. Experimental results show that the proposed
algorithm performs better than the three baseline ﬁngerprint matchers used in the study across all
image qualities. Furthermore, a score-level fusion of the proposed matcher and the best performing
commercial matcher shows a further boost in the matching performance.

131

CHAPTER 4
LATENT FINGERPRINT INDEXING
4.1 Introduction
Law enforcement agencies routinely collect tenprint records of all apprehended criminals in two
forms: rolled and plain. Latents – ﬁngerprints lifted from the surfaces of objects at crime scenes
– are regarded as an extremely important source of evidence to identify suspects since they can be
compared to rolled or plain ﬁngerprints in the background database of known identities. Compared
to rolled and plain ﬁngerprints that are obtained in an attended mode (see Figs. 4.1 (a) and (b)),
latents typically have poor quality in terms of ridge clarity and complex background noise, and
contain only a small part of a ﬁnger (friction ridge pattern) and large non-linear skin distortion (see
Fig. 4.1 (c)).
Due to these characteristics of latents, the search for the source of a latent is a challenging
problem in terms of both efﬁciency and identiﬁcation accuracy [106, 50], especially when the
background ﬁngerprint database is extremely large. The size of the ﬁngerprint database maintained
by just the police department of a typical large city can be of the order of a few millions. The Integrated Automated Fingerprint Identiﬁcation System (IAFIS) maintained by the Federal Bureau of
Investigation (FBI) contains ﬁngerprint records of over 74 million subjects [123] (approximately
740 million images) and this ﬁgure keeps growing on a daily basis as more individuals are added
to the database. The identiﬁcation accuracy generally deteriorates as the size of the ﬁngerprint
132

(a) Rolled

(b) Plain

(c) Latent

Figure 4.1: Examples of ﬁngerprints. (a) Rolled print (Michigan State Police database), (b) plain
or slap print (FVC 2002) and (c) latent print (NIST SD27).
database grows. Under simplifying assumptions, the performance in the identiﬁcation mode can
be estimated based on the FAR and FRR [5]. Assuming the same threshold is used in both the identiﬁcation and veriﬁcation scenarios, and that all identities for which the match score is above the
aforementioned threshold is returned by the identiﬁcation system, we have that the False Negative
Identiﬁcation Rate (FNIR)1 is the same as the FRR. Furthermore, the False Positive Identiﬁcation
Rate (FPIR)2 relates to FAR by
FPIR = 1 − (1 − FAR)N

(4.1)

, where N is the number of subjects in the background database. Thus, when the value of N
increases, the FPIR increases and the True Positive Identiﬁcation Rate (TPIR) decreases, which
means the matching accuracy of the identiﬁcation system is degraded.
This problem of large background database can be alleviated by quickly ﬁltering out a large
number of ﬁngerprints from the database that have very low similarity to the search latent before
performing detailed one to one matching. Therefore, an effective scheme for ﬁngerprint indexing,
also referred to as retrieval or database ﬁltering, is highly desirable to reduce the search space while
maintaining or possibly even improving the identiﬁcation accuracy. Fingerprint indexing is also
1 FNIR is

the rate of subjects enrolled in the database who are not identiﬁed by the system.
is the rate of subjects not enrolled in the database who are mistakenly identiﬁed by the
system as one of the subjects enrolled in the database.
2 FPIR

133

receiving more attention in recent years due to the limitations of the traditional Henry classiﬁcation
system. The Henry system of ﬁngerprint classiﬁcation attempts to divide the background database
into a ﬁxed number of classes – usually based on the ﬁve ﬁngerprint types: right loop, left loop,
whorl, arch and tented arch). The major problem with this classiﬁcation approach lies in the fact
that 90% of the ﬁngerprints belong to only three of these ﬁve classes [124], thus the background
database to be searched is not signiﬁcantly reduced for most queries. Furthermore, some ﬁngerprints cannot be reliably assigned to only one of these ﬁve categories. Indexing is the term used in
the ﬁngerprint recognition literature to refer to a signiﬁcant reduction in the number of ﬁngerprints
in the background database to be considered in the ﬁner matching stage [125, 124], or, in other
terms, to retrieve a subset of ﬁngerprints in the background database that are most similar to the
search ﬁngerprint. Here we use the term indexing to refer to an approach that provides continuous
classiﬁcation 3 rather than exclusive classes or categories, while the terms database ﬁltering and
ﬁngerprint retrieval can be used in both cases.
As mentioned in Chapter 1, ﬁngerprint friction ridge features are generally described at three
different levels. While level 3 features, like pores and ridge contours, are frequently adopted
to assist in identiﬁcation by latent print examiners [126], it is the level 1 (i.e., orientation ﬁeld,
singular point and type ) and level 2 (i.e., minutiae) features that are widely used in Automated
Fingerprint Identiﬁcation Systems (AFIS). It is believed that minutiae are the most discriminating
and reliable features. However, the information provided by minutiae set alone is limited.
In the case of indexing involving rolled/slap ﬁngerprints, the use of one type of feature alone
is usually sufﬁcient to achieve good indexing performance (e.g. [38]), and the use of additional
feature types might not signiﬁcantly increase the indexing performance to justify the added features. However, in the case of latent ﬁngerprints, one feature type alone might not be sufﬁcient in
most cases, since the availability and reliability of features in latents vary widely depending on the
speciﬁc latent characteristics.
3 Given

the features of a search ﬁngerprint, a distance or similarity to the features of reference
prints in the database is computed and most similar ﬁngerprints are retrieved based on the distance
or similarity values.

134

We propose to use both level 1 and level 2 features to improve the indexing performance for
latents. Firstly, orientation ﬁeld in the neighborhood of each minutia is encoded into a rotationand translation-invariant ﬁxed length bit vector. The bit vectors are then indexed by means of
Locality-Sensitive Hashing (LSH) [127]. Then, a conventional minutiae triplet based indexing [31]
is boosted by incorporating rotation constraints. Orientation ﬁeld indexing and triplet indexing are
fused with ﬁngerprint indexing technique based on the Minutia Cylinder-Code (MCC) representation [38], and this fusion is further boosted by combining singular points and ridge period ﬁltering.
The proposed indexing algorithm is tested by searching 258 latent ﬁngerprints in NIST SD27
against a background database that contains 267,258 rolled ﬁngerprints (including 258 rolled ﬁngerprints in NIST SD27, 27,000 rolled ﬁngerprints in NIST SD14, and 240,000 rolled ﬁngerprints
from the Michigan State Police). The experimental results demonstrate that the proposed algorithm
outperforms the state-of-the-art indexing methods published in the literature. After ﬁltering out a
large fraction of background ﬁngerprints for the ﬁner matching stage, the overall identiﬁcation
accuracy is maintained while the computational efﬁciency is signiﬁcantly improved.
The rest of this chapter is organized as follows: Section 4.2 gives a brief review of the related
work; Section 4.3 presents feature extraction for both latents and rolled prints; Section 4.4 describes in detail the proposed indexing approach; experimental results are provided in Section 4.5;
conclusions are drawn in Section 4.6.

4.2 Related Work
Fingerprint indexing techniques can be roughly categorized into three classes based on the types of
features used: minutiae-based, orientation ﬁeld-based and techniques based on other features (see
Table 4.1).
Indexing based on some salient characteristics of the ﬁngerprint (as opposed to ancillary information such as gender, race, age, etc.) is more important in the latent ﬁngerprint case than in the
rolled print case because ancillary information is usually not available with the latents. Rolled-torolled or plain-to-plain matching is usually performed for background checks or de-duplication.
135

In this case, a large portion of the database can be ﬁltered out by using the subject’s demographic
information (e.g. gender, ethnicity, age range, date of birth). In the latent case, this information
is usually not available because the latent comes from an unknown subject. In addition, there is a
huge gap between the matching performance of rolled-to-rolled matching compared to latent-torolled matching due to the partial and noisy nature of the latents; this makes the latent matching
problem — and thus the latent indexing problem — much more challenging than the rolled print
matching problem.
Among the published works on indexing, only two of them reported the indexing performance
on latent ﬁngerprints. Since the focus of our research is latent indexing, we will brieﬂy summarize
only these two studies. Feng and Jain [11] proposed a multi-stage ﬁltering scheme whose ﬁrst stage
depends on the ﬁngerprint pattern type, followed by the use of singular points and orientation ﬁeld.
The features are manually marked for the latents, and automatically extracted in the rolled prints.
The reported performance on NIST Special Database 27 [4] against a relatively small background
of 10,258 rolled prints is 97.3% hit rate at a 39% penetration rate. This means that the true mates of
2.7% of the latents (among the 258 latents in NIST SD 27) were not present in the ﬁltered database
(39% of the 10,258 rolled prints).
Yuan et al. [10] improved the performance of triplet indexing technique by using triplets to
count the number of polygons of various sides. Like [11], they also used manually marked minutiae
in the latents and automatically extracted minutiae in the rolled prints. Yuan et al. reported an
accuracy of 92.7% hit rate at a 40% penetration rate (80.7% at 10% penetration rate) for the NIST
SD27 database against a large background database of 240,258 rolled prints. This results [10]
could not be veriﬁed by us since by applying the algorithm in [10] (code was provided to us by
the authors) to our background database of similar size, the hit rate was only around 82% at 40%
penetration rate.

136

Table 4.1: Summary of studies on ﬁngerprint indexing for rolled, plain and latent prints.
Fingerprint Approach
Features

Germain et al.
[31]
Rolled

Author(s)

Minutiae

Lumini
[124]

OF

et

al.

Fingerprint Database

HR @
PR
=
10%

1,204 queries and 1,204
templates (NIST SD4)

84%

Triplets

Bhanu and Tan
[32]

Minutiae

Triplets

2,000 queries and 2,000
templates (NIST SD4)

85.5%

Lee et al. [128]

OF + RF

Histogram

1,204 queries and 1,204
templates (NIST SD4)

88%

Liu et al. [129]

OF

Complex Filter
Responses

2,000 queries and 2,000
templates (NIST SD4)

90%

Li et al. [130]

OF

Complex Filter
Responses

1,204 queries and 1,204
templates (NIST SD4)

96%

Jiang et al. [33]

OF + RF

OF Clustering

2,000 queries and 2,000
templates (NIST SD4)

89.5%

Wang et al. [34]

OF

Fingerprint Orientation Model
based on 2D
Fourier Expansion

2,700 queries and 2,700
templates (last 2,700
pairs of NIST SD14)

98%

Gyaourova
Ross [131]

Scores

Set of Reference
Prints

2,000 queries and 2,000
templates (NIST SD4)

84% @
PR=25%

Cappelli et al.
[38]

Minutiae

MCC

2,700 queries and
24,000
templates
(NIST SD14)

95%

Cappelli [132]

OF + RF

1,000 queries and
1,000,000
templates
(generated by SFinGe
v4.1)

99.6%

Cappelli [133]

Minutiae +
OF

MCC

2,700 queries and 2,700
templates (last 2,700
pairs of NIST SD14)

98.7%

Liu and
[134]

OF

Polar Complex
Moments

2,000 queries and 2,000
templates (NIST SD4)

88%

and

Yap

OF: orientation ﬁeld, SP: singular points, SIFT: scale invariant feature transform, MCC: minutia cylinder code, RF: ridge frequency, HR: hit rate, PR: penetration rate.
*The images used as templates were randomly selected from each ﬁnger.
**Feng and Jain [11] only reported the hit rate at a single penetration rate of 39%.
***The hit rate of algorithm in [10] evaluated on the database used in this paper is 58.1% @
137
PR=10%.

Table 4.1 (cont’d)
Fingerprint Approach
Features

Fingerprint Database

HR @
PR
=
10%

Bhanu and Tan
[32]

Minutiae

Triplets

400 queries and 600
templates (collected by
FIU-500-F01 sensor)

100%

Jiang et al. [33]
Plain

Author(s)

OF + RF

OF Clustering

600 queries and 200
templates (FVC2000
DB2a & DB3a)

92.5%

Liang
[135]

Minutiae

Triplets

550 queries and 330
templates (FVC2004
DB1∗ )

99%

Wang et al. [34]

OF

Fingerprint Orientation Model
based on 2D
Fourier Expansion

Queries
and
templates not indicated (
FVC2002 DB1a)

99.9%

Shuai et al. [30]

SIFT

500 queries and 300
templates (FVC2000
DB2a∗ )

98%

Gyaourova
Ross [131]

Scores

Set of Reference
Prints

2,400 queries and 2,400
templates (WVU)

85%

Cappelli et al.
[38]

Minutiae

MCC

700 queries and 100
templates (FVC2002
DB1a)

99%

Iloanusi et al.
[136]

Minutiae

Quadruplets

400 queries 400 templates
(FVC2004
DB1a∗ )

98%

Cappelli [132]

OF + RF

500 queries and 300
templates (FVC2002
DB1a∗ )

99.9%

Cappelli [133]

Minutiae +
OF

MCC

700 queries and 100
templates (FVC2002
DB1a)

100%

Liu and
[134]

OF

Polar Complex
Moments

700 queries and 100
templates (FVC2002
DB1a)

85%

et

al.

and

Yap

138

Table 4.1 (cont’d)

Latent

Author(s)

Fingerprint Approach
Features

Fingerprint Database

HR @
PR
=
10%

Feng and Jain
[11]

Fingerprint
Type + SP
+ OF

Multi-stage ﬁltering

258 latent queries
and 10,258 templates
(NIST SD27 and NIST
SD14)

97.3% @
PR=39%∗∗

Yuan et al. [10]

Minutiae

Triplets

258 latent queries and
240,258
templates
(NIST SD27 and a
private database)

80.7%∗∗∗

Proposed Approach [137]

Minutiae +
OF + SP +
RF

Triplets + MCC +
OF Descriptor Indexing

258 latent queries and
267,258
templates
(NIST SD27, NIST
SD14 and MSP)

81.8%
(95.7%
@
PR=39%)

139

4.3 Feature Extraction
In this section we describe various features that are extracted from reference prints and latents for
indexing purposes. More speciﬁcally, the following features are extracted: minutiae, orientation
ﬁeld, singular points and ridge period. Note these features are essentially the same that are used
in latent matching. The difference between latent indexing and matching is the order in which the
features are used and the multistage nature of indexing. In fact, some will argue that an efﬁcient
ﬁngerprint matcher(both for rolled prints and latent) has the indexing scheme implicitly built in
the matcher itself. However, in this dissertation, we view indexing and matching as two separate
modules.

4.3.1 Reference Prints
For reference (rolled) prints in the background database, two commercial off-the-shelf (COTS)
matchers are used to automatically extract all the features. Based on our experience with the
characteristics and performance of these two matchers, one of the COTS matcher is used to extract
the minutiae and the other one is used to extract the skeleton image. Fig. 4.2 shows a reference
print from Michigan State Police database, along with the skeleton, orientation ﬁeld and minutiae.
The orientation ﬁeld and ridge period of a reference print are extracted from the skeleton image as
follows:
1. Morphological operations (dilation and erosion) are applied to the skeleton image to get a
region of interest (ROI) of the reference print.
2. The skeleton image is ﬁltered with an averaging ﬁlter of size 5 × 5 pixels to convert a binary
skeleton image to a grayscale image (see Fig. 4.3).
3. The gradient based algorithm proposed in [8] is used to extract orientation ﬁeld from the
ROI region of the grayscale image (see Fig. 4.3).

140

4. The X-signature proposed in [115] is used to estimate the ridge period from the ROI region
of the grayscale image.

(a)

(b)

(c)

(d)

Figure 4.2: (a) Reference print from Michigan State Police database, (b) its skeleton extracted
by a commercial matcher, (c) orientation ﬁeld extracted from skeleton image, and (d) minutiae
extracted by a commercial matcher.
From the orientation ﬁeld, the singular points are extracted, if present in the reference print, by
the complex ﬁltering approach proposed in [138]. For both core and delta points, their directions
can also be extracted.
141

(a)

(b)

(c)
Figure 4.3: (a) Skeleton of a reference print from Michigan State Police database, (b) the grayscale
image obtained after applying image ﬁltering and (c) the extracted orientation ﬁeld.
4.3.2 Latents
Feature extraction in latents is a challenging problem due to heavy background noise. Since the
objective of our study is indexing, we assume, similar to the earlier studies in [11] and [10], that
minutiae features have been manually marked in the latents by an expert. This is the case for the
NIST SD27 database, where manually marked minutiae have been provided along with singular
points. The orientation ﬁeld and ridge period information in latents were marked by the authors
in [28] and are available to us. Fig. 4.4 shows some of the manually marked features (orientation

142

ﬁeld, singular points and minutiae) available on a latent print from NIST SD27 latent database.

(a)

(b)

(c)

(d)

Figure 4.4: (a) Latent print from NIST SD27 latent database, and some manually marked features:
(b) orientation ﬁeld, (c) singular points and (d) minutiae.

4.4 Indexing Approach
We propose to combine different features and indexing techniques to improve latent ﬁngerprint
indexing performance. For full (rolled, slap) ﬁngerprints, one feature type alone (e.g. minutia
143

cylinder code) has been shown to be successful for indexing. But, in the case of latents, availability
of features is limited due to small area, distortion, etc. Therefore, the proposed indexing approach
using a combination of features is more appropriate for latents.
Our approach consists of combining (i) a constrained version of triplet indexing, (ii) MCC
indexing as proposed in [38], (iii) a new orientation ﬁeld descriptor indexing technique that uses
hash function, (iv) ﬁltering based on singular points as proposed in [11], and (v) averaged ridge
period comparison. Basic triplets indexing technique is improved by applying a rotation constraint
to the matched triplets. Orientation ﬁeld descriptor indexing is carried out ﬁrst by converting the
descriptor to a binary vector, and then by applying hash functions, similar to the approach proposed
in [38]. Each of these speciﬁc features used outputs an indexing score, which are then combined to
obtain the ﬁnal indexing score. A description of the techniques used in our approach is presented
below, and the overall scheme is shown in Fig. 4.5.

144

Figure 4.5: Overview of the proposed indexing scheme.

145

4.4.1 Triplets
Features extracted from the triangles (triplets) formed by any subset of three minutia points have
been popular for ﬁngerprint indexing [31]. Several modiﬁcations of the ﬁrst triplet-based algorithm
[31] are now available [32, 135, 10]. The basic features in [31] consisted of the length of the sides
of the triangles, the ridge count between every pair of minutiae in the triplet, and the angle between
minutiae direction and the side of the triangle. The ordering of the sides of the triangle was deﬁned
in the clockwise direction.
In our approach, we use the three side lengths of the triangle, and the difference between
minutiae direction and one side of the triangle for indexing. Ridge count between minutiae in
latents is not sufﬁciently reliable so we do not use this feature. As proposed in [32], we order
the three sides as Lmax , Lmin , Lmed based on their lengths (see Fig. 4.6). Each minutia point
mi = (xi , yi , β j ), i = 1, 2, 3, is associated with a vertex of the triangle by the following rule: P1 is the
point associated with the largest angle, which is opposite to the largest length side, P2 is associated
with the minimum angle in the triangle, which is opposite to the minimum length side, and P3 is
associated with the median angle in the triangle, which is opposite to the median length side, as
illustrated in Fig. 4.6. After the ordering of the sides and minutiae, the directional differences
(θi , i = 1, 2, 3) between each minutiae and one side of the triangle can be consistently obtained.
We also use the handedness of the triangle proposed in [32] as one of the triplet features. Let
Pi = (xi , yi ) be the point corresponding to the ith minutiae in the triplet, Zi = xi + jyi be the complex
√
number associated with mi , i = 1, 2, 3, j = −1, and Z21 = Z2 − Z1 , Z32 = Z3 − Z2 , Z13 = Z1 − Z3 .
Then, the triangle handedness φ is deﬁned as

φ = sign(Z21 × Z32 ),

(4.2)

where sign(·) ∈ {−1, 1} is the sign function and × is the cross product.
In summary, given a set of minutiae points in a latent, the feature vector extracted for every
possible triplet t with side length in the range [lmin , lmax ] — in our implementation, 20 and 150

146

P1
z

1

y

!max

Lmin

Lmed

x
Right handed

!min

!med
3

2

Lmax

P2

P3

Figure 4.6: Illustration of triplets features.
respectively — is given by
Ft = (Lmax , Lmin , Lmed , θ1 , θ2 , θ3 , φ ).

(4.3)

These triplet features are quantized into bins (bin size is 20 for both triplet side lengths and
directional differences; φ , which can be either −1 or 1, is mapped to {0, 1}). A key is generated based on the possible values of each bin and associated with each triplet. We have that each
bin j, j = 1, . . ., 7 can assume possible values vk , k = 1, . . ., M j , and these values are mapped to
1, . . ., M j , j = 1, . . . , 7. Assume b1 = 1, b j = b j−1 M j−1 , j = 2, . . ., 7 and v j is the value corresponding to the jth bin for triplet t, then the key to triplet t is computed as
7

Kt =

∑ v jb j.

(4.4)

j=1

During the indexing or search, triplets in reference prints with the same key as triplets in the
latents are considered as matched triplets.
Latent examiners can usually position or align a latent very close to its true position in the ﬁnger based on their knowledge of ﬁngerprint structure. Also, rolled prints in background databases
usually present only a small rotation compared to the ﬁnger upright position. Thus, an additional
147

step is added to further improve the indexing performance of the triplets by constraining the possible rotation between the two matched triplets. We call this as the constrained triplets approach.
The constraint is imposed by ﬁrst computing the rotation among each pair of matched triplets.
Given a matched triplets pair, the differences in the direction of the corresponding three minutiae
are averaged and taken as the rotation between that matched triplets pair. If this difference is larger
than a threshold (60 degrees in our implementation), then that triplet pair is excluded from the
set of matched triplets. The triplet indexing score between the latent and a reference print in the
background database is computed as the number of matched triplets between them.

4.4.2 Minutia Cylinder Code (MCC) Indexing
In [38], the authors proposed an indexing technique based on a local minutiae descriptor, called
Minutia Cylinder Code (MCC). MCC descriptor [91] represents the neighborhood minutiae information in the form of a 3D function (cylinder). Each level of the cylinder corresponds to a
direction difference range, and for each direction difference range, the descriptor contains spatial
information about the neighboring minutiae relative to the center minutia (refer to Chapter 3 for
details of MCC). MCC descriptor has been successfully used in both rolled-to-rolled ﬁngerprint
matching [91] and latent-to-rolled ﬁngerprint matching [88].
Hash functions are utilized to calculate an indexing score for two given ﬁngerprints (search
and reference print). The indexing score is computed by averaging the estimate of the Hamming
distance between the descriptors, without the need to explicitly perform the distance computations.
The estimate of the Hamming distance is based on the collisions between two descriptors under
the hash functions so that the indexing is more efﬁcient than directly comparing the descriptors.
More speciﬁcally, given two ﬁngerprint templates T1 and T2 , with V1 and V2 being the sets of
cylinders for each minutia in each template, the indexing score proposed in [38] is given by
p

SM (T1 , T2 ) ∼
=

∑v∈V1 (maxvj ∈V2 {CF (v, vj )}) h
p
|V1 | × (HM ) h

148

,

(4.5)

where |V1 | is the number of cylinders in the search print, HM is the number of hash tables, h is the
number of bits selected in each hash table, p is a parameter of the distance function (set to 30),
and CF represents the number of collisions over all hash functions between cylinders v and v j . We
used MCC SDK provided by the authors [38] with the default parameters, with the exception of
the number of hash tables that was increased to 64 instead of the default 32 because we found that
the former worked better for the latent indexing.

4.4.3 Orientation Field Descriptor Indexing
Given a set of minutiae M = {(xi , yi , βi)}N and orientation ﬁeld f (x, y), where N is the number
i=1
of minutiae, xi and yi are the coordinates of the ith minutia, respectively, and βi is the direction
of minutia, a set of descriptors O = {o1 , o2 , . . . , oN } is extracted, where oi = {oi,1 , oi,2 , . . . , oi,n }
is the descriptor for the ith minutia and n is the total number of sample points in the orientation
descriptor [27].
Given a minutia, L concentric circles are centered at the minutia with the jth circle of radius
R× j
, and K j sample points are equally spaced on each circle starting from the projection location
L
of the reference minutia along its direction on the circle. In this paper, the number of circles L
is 4 and the numbers of sample points from inner circle to outer circle are 10, 16, 22 and 28,
respectively, the maximum radius R for the outer most circle is 80 pixels. So, the total number
of sample points for each orientation descriptor is n = ∑L K j = 76. The feature value at each
j=1
sample point is computed as the counter-clock wise rotation angle from local orientation ﬁeld at
the sample point to the local orientation at the central minutia.
In the indexing algorithm, given a seed, a set of l pairs of sample points {(h j , b j )}lj=1 from the
orientation descriptor is randomly generated, where h j , b j ∈ {1, 2, . . ., n} and h j = b j . Then, a bit
vector r = {r1 , r2 , . . ., rl }, is generated as


 1, if o
i,h j > oi,b j ,
j = 1, 2, . . ., l.
rj =

 0, otherwise ,
149

(4.6)

The random generation of the bit vectors is repeated so that the total number of hash functions
is HO , and the number of bits in each hash function is ﬁxed (l). Each bit vector corresponds to a
decimal number, which is then used as one of the index keys. The generation of the orientation
ﬁeld descriptor indexing tables is illustrated in Fig. 4.7. Given two sets of orientation descriptors
O1 and O2 , similar to the MCC index score computation, the index score for orientation descriptor
is computed as
p

SO (T1 , T2 ) ∼
=

∑r∈O1 (maxrj ∈O2 {CF (r, rj )}) l
p

,

(4.7)

|O1 | × (HO ) l

where CF represents the number of collisions over all hash functions as in Equation 4.5.

Figure 4.7: Illustration of the key generation in the orientation ﬁeld descriptor indexing scheme.

4.4.4 Singular Points
Singular points provide useful characterization of a ﬁngerprint. Given a list of core points and
delta points with each singular point containing coordinate and direction information, three types
150

of singular point pair features can be extracted [11]: 1) core pair; 2) delta pair; and 3) core-delta
pair.
To order the singular points in a pair of singular points, each core point pair generates two pairs
with each core point being ranked ﬁrst once. For delta pairs, the delta points are ordered based on
the x-coordinate, and for core-delta pairs, the core points are ordered ﬁrst. For each singular pair,
three features are computed: 1) the distance (d) between the singular points; 2) the counter-clock
wise rotation angle (α ) from the direction of the ﬁrst singular point to the line connecting singular
points; 3) the counter-clock wise rotation angle (β ) from the direction of the second singular point
to the line connecting singular points.
The singular pair similarity score between latent and reference print is deﬁned as
N
1 SP
SSP =
∑ max s ,
NSP i=1 j i, j

(4.8)

where NSP is the number of singular pairs in latent and si, j is the similarity between the ith singular
pair in latent and the jth singular pair of the same type in reference print, which can be computed
as





1
f20,60 (∆di, j ) + f π π (∆αi, j ) + f π π (∆βi, j ) ,
(4.9)
3
,
,
6 3
6 3
where ∆di, j is the distance difference, ∆αi, j and ∆βi, j are the angle differences between the ﬁrst
si, j =

and second singular points, respectively, and fa,b (x) is a piece-wise linear function:

 1

x < a,



fa,b (x) =
0
x > b,

 b−x



otherwise.
b−a

(4.10)

4.4.5 Ridge Period
Given a ﬁngerprint (either latent or reference print), the average ridge period is computed within
a circle of radius 10 blocks centered at the core point, with each block being 16 × 16 pixels; only
foreground blocks are used. If there is no core point in the print, the center point of the foreground
151

region is used. The ridge period similarity is then computed as
|R − Rr |
,
SR = exp − l
σR

(4.11)

where Rl and Rr are the average ridge periods in the latent and the reference print, respectively,
and σR is a normalization term which is usually set to the average of |Rl − Rr |. In this paper, σR is
set to 2.

4.4.6 Fusion of Indexing Scores
The indexing scores based on different features (triplets, MCC, orientation ﬁeld descriptor, singular
points and ridge period) take values in different ranges. Thus, normalization of individual indexing
scores is done before the fusion. Let s = {s1 , s2 , . . . , sN R } be the list of scores corresponding to
a latent, and N R be the number of reference prints in the background database. The normalized
score s′ = {s′ , s′ , . . . , s′ R } is obtained as
1 2
N

s′ =
i

si − min(s)
.
max(s) − min(s)

(4.12)

The fusion is based on the saliency of individual features. Triplets, MCC and orientation ﬁeld
descriptor provide similar individual performances; thus, with proper normalization, the simple
sum rule is successful [139]. Singular points are not always present in ﬁngerprints, especially in
latents that usually contain a partial area. Therefore, the indexing score from the other features is
increased based on whether singular points are available in both the prints being compared.
Given the indexing scores from the constrained triplets (ST ), MCC indexing (SM ), and orienta′
′
′
tion ﬁeld descriptor (SO ), we obtain locally normalized scores ST , SM and SO . Given the indexing

scores from singular points (SSP ) and ridge period (SR ), the ﬁnal indexing score S is computed as
′
′
′
S = (ST + SM + SO ) × (1 + SSP) × SR .

(4.13)

The ﬁnal indexing scores are then used to order the background database such that, for each
latent, the top-k reference prints that are most similar to the latent print are retrieved.

152

4.5 Experimental Results
4.5.1 Databases
Indexing experiments were performed on the only publicly available latent ﬁngerprint database,
NIST Special Database 27 [4]. This database consists of 258 latent ﬁngerprint images from operational scenarios (i.e., latents collected from crime scenes) and 258 rolled prints corresponding to
the mates of the latent ﬁngerprints. We used 500 ppi resolution for both latents and rolled prints.
Features (i.e., minutiae) marked by latent examiners are provided with this database.
The purpose of latent ﬁngerprint indexing is to ﬁlter out a large portion of the background
database so that the speed of matching stage is enhanced. It is also important to ensure that this
gain in matching speed is not at the cost of loss in matching performance. To show the effectiveness of the proposed indexing scheme, we build a large background database of 267,258 rolled
ﬁngerprints. This background database includes the 258 mated rolled prints from NIST SD27,
27,000 rolled prints from NIST SD 14 [90], and 240K rolled prints from a database provided by
the Michigan State Police (see Fig. 4.8).

(a)

(b)

(c)

Figure 4.8: Rolled prints in (a) NIST SD27, (b) NIST SD14, and (c) Michigan State Police
databases.

153

4.5.2 Indexing Results
Indexing performance is usually measured by computing the hit rate at various penetration rates.
A hit rate at a given penetration rate p refers to the portion of the search prints for which the
true mates are retrieved within p% of the background database (penetration rate). The desirable
outcome is high hit rates at low penetration rates and high computational efﬁciency. Further, it is
also desired that the matching accuracy with indexing be no lower than matching accuracy without
indexing.
As a baseline, the hit rates at the ﬁxed penetration rate of 20% are 62.0%, 73.6%, and 74.0% for
indexing schemes based on basic triplets, MCC indexing and the constrained triplets, respectively.
The hit rates at varying penetration rates are shown in Fig. 4.9. Note that for the three methods
shown, only minutiae features are used.

100
90

Hit Rate (%)

80
70
60
50
40
30
20
0

20

Triplets Indexing
Constrained Triplets
MCC SDK Indexing
40
60
80
100
Penetration Rate (%)

Figure 4.9: Comparison of the indexing performance of basic triplets, constrained triplets and
MCC indexing.
Fig. 4.10 shows the performance of the proposed indexing approach along with the performance of two state-of-the-art approaches proposed in [11] and [10]. The results of the algorithm in

154

[10] shown here are based on an executable ﬁle provided by the authors [10], applied to the same
database and features as used in our approach. The hit rate at a penetration rate of 39% reported
in [11] is shown here for Feng and Jain 2008, which is based on a background database of only
10,258 rolled prints. Note that our approach cannot be directly compared with Feng and Jain’s approach [11] because they used a much smaller background database size compared to our approach
and Yuan et al.’s approach [10]. However, in our experiments we observed that the differences in
background database size do not signiﬁcantly affect the hit rates (usually about 2-3% difference).
Thus, if we ﬁx the penetration rate at 39%, both performances, ours and that reported in [11], are
comparable. Furthermore, we will show that the overall latent matching performance is improved
by using our approach that ﬁlters out 80% of the reference prints, compared to ﬁltering out only
61% in [11].

100
90

Hit Rate (%)

80
70
60
50
40
30
0

20

Feng and Jain 2008
Yuan et al. 2012
Proposed Approach
40
60
80
100
Penetration Rate (%)

Figure 4.10: Comparison of indexing performance based on the proposed approach against Yuan
et al. [10] and Feng and Jain [11] on the NIST SD27 latent database. Note that Feng and Jain [11]
used only a small background database of 10,258 versus the proposed approach and [10] that use
267,258 reference prints in the background database. Further, Feng and Jain [11] did not provide
the complete hit rate vs. Penetration rate curve, but only a single point at a penetration rate of 39%
(•).

155

Given that the proposed approach combines different features and techniques for indexing, we
also determine the incremental contributions to hit rates with the addition of different features at a
ﬁxed penetration rate (20%). This is shown in Fig. 4.11, where the x-axis denotes the successive
utilization of different ﬁngerprint features, with the leftmost feature CTrip being the constrained
triplets. It can be observed that the incremental addition of features steadily improves the indexing

Hit Rate at 20% Penetration Rate

performance.

90

85

80

75
CTrip

OFDesc

MCC

SP

RP

Figure 4.11: Increase in the cumulative hit rate at a ﬁxed penetration rate (20%) as we incrementally add features for indexing (from left to right), where CTrip denotes constrained triplets,
OFDesc denotes orientation ﬁeld descriptor, SP denotes singular points, RP denotes ridge period.
The plot shows that we can achieve about 91% hit rate at 20% penetration rate when we use all
ﬁve features (CTrip, OFDesc, MCC, SP and RP) for indexing.
In order to further evaluate the performance of the proposed indexing algorithm, we also analyzed the inﬂuence of the ﬁltering on the latent matching performance. For this purpose, three
ﬁngerprint matchers were used: (i) an in-house matcher (DBHT) [88], (ii) a commercial ﬁngerprint
matcher (Morpho) and (iii) a commercial latent ﬁngerprint matcher (here referred to as LCOTS).
The input to the two ﬁrst matchers consisted of manually marked minutiae in the latents and
156

automatically extracted minutiae in the rolled print, which is the same set of minutiae that was
used in the indexing experiments. For the LCOTS, the raw image was provided as input.
For this experiment, for computational efﬁciency we used a smaller background database of
27,258 rolled prints (258 from NIST SD27 and 27K from NIST SD14). Our indexing approach
consisted of not only ﬁltering out a portion of the database, but also of combining both the match
scores and indexing scores to improve the matching performance. After retaining a small portion
of the background database, we combine the indexing score with the match score by the sum rule.
Then, the overall latent matching performance is improved.
For the DBHT and Morpho matchers, the overall latent matching performance on NIST SD27
before and after ﬁltering out 80% of the background database (penetration rate of 20%) is improved, as shown in Fig. 4.12. This demonstrates that our indexing scheme has the desirable
property, namely a signiﬁcant reduction in matching time while improving the matching accuracy.
While our indexing is not perfect, most of the errors in the indexing are also errors in the latent
matching stage as well. By ﬁltering out 80% of the database, the total computational time for
the search is reduced by a factor of 5 for both the in-house matcher and Morpho, while the latent
matching performance is improved.
In the case where the indexing is applied, three scenarios are possible regarding the retrieval
rank before and after the ﬁltering: (i) the retrieval rank of the true mate stays the same, (ii) the
retrieval rank of the true mate improves, or (iii) the true mate is excluded from the background
database. The ﬁrst case occurs when none of the non-mated reference prints with match scores
higher than the true mate match score are excluded from the background database. In the second
case, the retrieval rank improves because of the exclusion of reference prints with higher match
scores to the latent compared to the true mate match score. The third scenario occurs when the
true mate is incorrectly ﬁltered out by the indexing module. The ﬁrst and second cases are a great
advantage due to the fact that the search time is greatly reduced while the true mate retrieval rank
is maintained or improved. The third scenario might reduce the matching performance, unless the
true mate was not going to be retrieved at higher ranks by the ﬁngerprint matcher.

157

75

Identification Rate (%)

70
65
60
55
50
45
0

20

40

Rank

DBHT before filtering
DBHT after filtering
Morpho before filtering
Morpho after filtering
60
80
100

Figure 4.12: Latent matching performance of two different matchers (DBHT in-house latent
matcher and Morpho matcher) before and after applying the proposed indexing to ﬁlter out 80% of
the background database. Note that after the ﬁltering and the fusion of indexing and match scores,
the overall matching performance is improved while the matching time is reduced by a factor of
ﬁve (20% penetration rate).
Two latents and their true mates are shown in Fig. 4.13. In both cases, the retrieval rank of the
true mate by the in-house matcher was improved after the ﬁltering was applied (3 to 1 and 142 to
49, respectively). In the second case, the true mate was not in the top-100 candidates before the
ﬁltering, while it became the top 49 candidate after ﬁltering. Ground truth matched minutiae are
shown for reference.
For the commercial latent ﬁngerprint matcher, the overall latent matching performance on NIST
SD27 before and after ﬁltering out two thirds of the background database (penetration rate of
33.33%) is improved, as shown in Fig. 4.14. Thus, our proposed indexing scheme improves the
matching accuracy of one of the top commercial latent matchers.
In the cases for which the commercial latent matcher provided zero score between the latent
158

(a)

(b)
Figure 4.13: Examples of latents for which the true mate was correctly retrieved when 20% of the
background database was considered are shown. Note that by applying the indexing, the retrieval
rank went from (a) 3 to 1 and (b) 142 to 49 due to the exclusion of non-mated reference prints with
high match scores to the latent. In (b), the true mate only appeared in the top-100 candidate list
after the ﬁltering.
and its true mate, we consider that the true mate is retrieved in the last rank. Fig. 4.15 shows
two latents and their true mates for which the true mate could be correctly retrieved at rank 1
after the indexing scheme proposed here was applied and the indexing scores and match scores
were combined. The true mates were both ranked last by the commercial latent matcher before the

159

86

Identification Rate (%)

84
82
80
78
76
74
72
0

20

40

Rank

LCOTS before filtering
LCOTS after filtering
60
80
100

Figure 4.14: Latent matching performance of a commercial latent ﬁngerprint matcher before and
after applying the proposed indexing and retaining one third of the background database. The
ﬁltering improves the overall matching performance of the commercial latent ﬁngerprint matching
while reducing the matching time.
ﬁltering.
In Fig. 4.16, two latents and their true mates are shown. In both cases, the true mates were
excluded by the indexing, while they were both ranked ﬁrst by the commercial latent matcher. The
manually marked minutiae and the minutiae extracted by Morpho in the rolled print provide only
a small number of corresponding minutiae, therefore the indexing approach is not successful and
the true mates are excluded during the ﬁltering.
The proposed approach is more suitable for latent indexing. For indexing of reference prints,
the state of the art indexing methods [38] already report very high hit rate at very low penetration
rate. For example, the indexing performance of MCC applied on NIST SD4 (the 2,000 second
impressions are searched against the 2,000 ﬁrst impressions and 10,000 reference prints from the

160

(a)

(b)
Figure 4.15: Examples of latents for which the true mate was correctly retrieved when one third
of the background database was considered are shown. Note that by applying the indexing, the
retrieval rank went from last to 1 due to the fusion of indexing and match scores.
Michigan State Police database) shows a hit rate of 97.8% at 5% penetration rate. Including additional features made only a minor improvement to the indexing performance in this case. Most of
the failure cases in the rolled print indexing are due to poor quality of one of the two impressions.

161

(a)

(b)
Figure 4.16: Examples of latents for which the true mate was not retrieved when one third of the
background database was retained. The true mates retrieval ranks before the ﬁltering were 1, so
the indexing degrades the matching performance for these two latents.
4.5.3 Implementation Details
Here we report the implementation details of each module used in our indexing scheme. The
triplets indexing algorithm, singular point and ridge period estimation modules are implemented in
Matlab; MCC SDK (implemented in C#) is provided by the authors of [38] and can be downloaded

162

from their webpage4 ; the orientation ﬁeld descriptor-based indexing algorithm is implemented
in C++. MCC indexing timing was obtained on a PC (Intel(R) Core(TM)2 Quad CPU Q9650
3.00GHz, 3GB memory), and the timing for all the other modules were obtained on a server with
48 cores (AMD Opteron(tm) Processor 6176, 2.3GHz, 193GB memory); no parallel computing
resource was used, thus only one core was used when measuring the computation time.
Table 4.2 reports the computation time for each search module. The modules are all independent and can be run in parallel; in this case, the total time for the indexing would be 28,667
reference prints searched per second, per latent, plus the time spent extracting features from the latent image (less than 3 seconds for all the 258 latents, given the manually marked features minutiae,
orientation ﬁeld, singular points and ridge frequency).
Table 4.2: Computation time for individual indexing modules.

Module

# of reference prints searched per second

Programming language

Basic triplets

39,509

Matlab

Constrained triplets

30,777

Matlab

MCC SDK

58,604

C#

Orientation descriptor

28,667

C++

Singular points

38,444

Matlab

Ridge period

382,960

Matlab

4.6 Summary
We have presented an indexing approach for latent ﬁngerprints that outperforms state-of-the-art indexing techniques on the public domain latent database NIST SD27 against a background database
of ∼267K rolled prints. The proposed approach combines various level 1 and level 2 features, including minutia cylinder code, minutia triplets, singular points, orientation ﬁeld and ridge period.
Improvement in the overall indexing accuracy due to the addition of each individual feature is
4 http://biolab.csr.unibo.it/

163

shown. The experimental results show that ﬁltering not only improves the computational efﬁciency of latent search, but the latent matching accuracy is maintained. For the NIST SD 27, the
matching accuracy of two different matchers, with and without ﬁltering, remains the same at 20%
penetration rate. While the proposed indexing scheme is not perfect, the indexing errors reﬂected
in the hit rate are also reﬂected in the matching errors.

164

CHAPTER 5
SUMMARY AND FUTURE WORK
5.1 Summary
This dissertation studied two very challenging problems in biometric recognition: biometric traits
of identical twins and latent ﬁngerprint matching. The primary contributions of this dissertation
are (i) to provide a better understanding of the similarities and dissimilarities of the biometric traits
of identical twins and (ii) to advance latent ﬁngerprint indexing and matching performance.
In Chapter 2, we have presented an analysis of the biometric characteristics of identical twins
(ﬁngerprint, face and iris). The discriminability of these three biometric traits is supported by
anatomy and the formation process of the biometric characteristics. We have assessed the capacity
of state-of-the-art commercial biometric matchers in distinguishing identical twins, as well as their
capacity to determine whether two subjects are identical twins or not.
Face biometric system can distinguish two different persons who are not identical twins much
better than it can distinguish identical twins. More recent studies support this conclusion, and the
face recognition becomes even more challenging when occlusion, expression changes or different
lighting conditions are involved.
Fingerprint biometric system also can discriminate two different persons who are not identical
twins only slightly better than it can discriminate identical twins, this difference is not as large as
for the face biometric system. In the ﬁngerprint matching experiments, the identical twin impostor
165

distribution is shifted to the right (indicating higher similarity) compared to the general impostor
distribution; a statistical hypothesis test shows that the difference between these two distributions
is signiﬁcant.
The iris matching experiment results show that while there is a signiﬁcant difference in the
performance of the biometric system for the identical twin data and for the general population
data, the iris biometric system can successfully distinguish identical twins.
Multimodal biometric systems that combine different units of the same biometric modality
(e.g., fusion of four ﬁngers or fusion of two irises) lead to an almost perfect separation between
genuine and impostor distributions for identical twins.
Biometric traits can also be used to determine whether two subjects are identical twins. By using face and iris modalities together, for example, we can correctly determine 80% of the identical
twin pairs as such, while only 2% of subject pairs who are not identical twins will be incorrectly
considered identical twins.
In Chapter 3, we have presented two different methods of improving latent ﬁngerprint matching
performance. It is assumed, as is the current practice, that manually marked features (minutiae,
singular points and ROI) are available for latents.
We have shown that the matching performance based on manually marked minutiae in latents
can be improved by utilizing automatically extracted minutiae from enhanced latent images. This
framework consists of the following steps: (i) reconstruct the orientation ﬁeld based on manually
marked minutiae and singular points; (ii) enhance the latent using median ridge frequency computed in small image blocks and the reconstructed orientation ﬁeld; (iii) match minutiae extracted
from enhanced latents and rolled ﬁngerprints; (iv) combine the scores from the two matchers using boosted max fusion technique. This framework improved the latent matching performance
irrespective of latent quality.
We have also presented a latent ﬁngerprint matching algorithm which is based on a descriptorbased Hough Transform alignment. Experimental results show that the proposed algorithm performs better than the three ﬁngerprint matchers (that includes two COTS matchers) used as baseline

166

across all latent image qualities. Furthermore, a score-level fusion of the proposed matcher and the
best performing COTS matcher shows a further boost in the latent matching performance.
In Chapter 4, we have presented an indexing approach for latent ﬁngerprints that outperforms
state-of-the-art indexing techniques on the public domain latent database NIST SD27 against a
large background database of ∼267K rolled prints. The proposed approach combines a number
of level 1 and level 2 features to boost the indexing performance. Contribution to the overall
indexing accuracy from each individual feature is shown. Indexing experimental results show that
ﬁltering not only improves the computational efﬁciency of latent search, but also improves the
latent matching accuracy when indexing scores are combined with match scores.

5.2 Contributions
In Chapter 2, an analysis of the biometric traits of identical twins provides:
• Capability of state-of-the-art biometric systems to successfully distinguish between identical
twins using the three most common biometric traits: ﬁngerprint, face and iris.
• Capability of state-of-the-art biometric systems to successfully determine whether two subjects are identical twins based on their biometric traits (face, ﬁngerprint, iris) alone.
In Chapter 3, we studied the latent matching problem and we presented two different approaches to improve the matching performance given manually marked features (minutiae, singular
points, ROI) in the latents. The contributions of this chapter are as follows:
• An approach that combines manually marked features with automatically extracted features
from enhanced latents to obtain a better matching performance using a COTS ﬁngerprint
matcher.
• Orientation ﬁeld reconstructed from minutiae along with singular points are used to enhance
the latent image to obtain reliable automatically extracted minutiae.

167

• A fusion scheme that combines match scores based on manually marked minutiae and automatically extracted minutiae from enhanced latents is able to increase the performance
consistently over all latent ﬁngerprint qualities in NIST SD27.
• A latent ﬁngerprint matcher is developed using a descriptor-based Hough Transform for
alignment (between latent and reference print) and manually marked minutiae from the latents as input.
• The proposed latent matcher advances the matching performance compared to three baseline
matchers (that includes two COTS matchers).
In Chapter 4, we presented an indexing approach that combines different features to advance
the performance of latent ﬁngerprint indexing. In latents, since there is no guarantee that all the
features will be present, a combination of different features for indexing makes more sense than
rolled/plain ﬁngerprints. The contributions of the proposed indexing approach are:
• An analysis of the contribution of different features for latent indexing.
• A reduction in the total search time while improving the latent matching accuracy on the
NIST SD27 database with a background of 267,258 reference prints.

5.3 Future Work
In Chapter 2, we presented an analysis of the biometric traits of identical twins. Identical twins are
the most similar persons in terms of genetics. The next group considering the degree of similarity
would be siblings. To extend the study on the similarity of biometrics of identical twins, the use
of siblings data would be the next step. For example, can biometrics be used to ﬁnd siblings of
unidentiﬁed victims? We used matching scores to derive a measure for identical twin determination. However, additional features need to be investigated for identical twin determination. Some
possible features to consider include ﬁngerprint pattern type and number of matched minutiae. It
has been shown in the literature that ﬁngerprint pattern type is more correlated between ﬁngerprints
168

of identical twins and the number of matched minutiae is also slightly higher compared to matched
minutiae from general impostor pairs.
In Chapter 3, our purpose was to improve the latent matching performance by developing a
matcher based on the availability of a limited amount of reliable features, i.e., manually marked
features. Research on extracting reliable features from the latents would be very beneﬁcial to
further improve state-of-the-art in latent ﬁngerprint matching.
In Chapter 4, our indexing technique assumes the availability of manually marked features in
latents, and it considers all the features in parallel for indexing. A better fusion scheme would take
into account the differences in the latents so that the weights assigned to different features used in
indexing can be adaptively determined. Further, different features could be used sequentially as
opposed to in parallel as implemented here.

169

BIBLIOGRAPHY

170

BIBLIOGRAPHY

[1]

D. Maltoni, D. Maio, A. K. Jain, and S. Prabhakar, Handbook of Fingerprint Recognition.
Springer-Verlag, 2nd ed., 2009.

[2]

NIST Special Database 4, “NIST 8-bit gray scale images of ﬁngerprint image groupsFIGS.”
http://www.nist.gov/srd/nist4.cfm.

[3]

Integrated Pattern Recognition and Biometrics Lab, “West virginia university latent
database.” http://www.cse.msu.edu/~rossarun/i-probe/.

[4]

NIST Special Database 27, “Fingerprint minutiae from latent and matching tenprint images.” http://www.nist.gov/srd/nistsd27.cfm.

[5]

A. K. Jain, A. A. Ross, and K. Nandakumar, Introduction to Biometrics. Springer, 2011.

[6]

D. G. Lowe, “Distinctive image features from scale-invariant keypoints,” International Journal of Computer Vision, vol. 60, no. 2, pp. 91–110, 2004.

[7]

P. J. Phillips, P. J. Flynn, K. W. Bowyer, R. W. V. Bruegge, P. J. Grother, G. W. Quinn,
and M. Pruitt, “Distinguishing identical twins by face recognition,” in IEEE International
Conference on Automatic Face and Gesture Recognition (FG), pp. 185–192, 3 2011.

[8]

A. M. Bazen and S. H. Gerez, “Systematic methods for the computation of the directional
ﬁelds and singular points of ﬁngerprints,” IEEE Trans. on Pattern Analysis and Machine
Intelligence, vol. 24, pp. 905–919, July 2002.

[9]

E. Tabassi, C. L. Wilson, and C. I. Watson, “NIST ﬁngerprint image quality,” tech. rep.,
NIST Interagency/Internal Report (NISTIR) - 7151, August 2004.

[10] B. Yuan, F. Su, and A. Cai, “Fingerprint retrieval approach based on novel minutiae triplet
features,” in Biometrics: Theory, Applications and Systems (BTAS), pp. 170–175, September
2012.
[11] J. Feng and A. K. Jain, “Filtering large ﬁngerprint database for latent matching,” in International Conference on Pattern Recognition (ICPR), pp. 1–4, December 2008.
[12] NSTC Subcommittee on Biometrics, “Biometrics history.” http://www.
biometrics.gov/documents/biohistory.pdf, August 2006.
Accessed
November 20, 2012.
[13] H. T. F. Rhodes, Alphonse Bertillon, Father of Scientiﬁc Detection. New York: AbelardSchuman, 1956.

171

[14] National
Law
Enforcement
Museum
Insider,
“Bertillon
system
of
criminal
identiﬁcation.”
http://www.nleomf.org/museum/
news/newsletters/online-insider/november-2011/
bertillon-system-criminal-identification.html, November 2011.
[15] H. Faulds, “On the skin-furrows of the hand,” Nature, vol. 22, p. 605, October 1880.
[16] W. Herschel, “Skin furrows of the hand,” Nature, vol. 23, p. 76, November 1880.
[17] F. Galton, “Personal identiﬁcation and description,” Nature, vol. 38, pp. 201–202, 1888.
[18] H. Cummins and C. Midlo, Finger Prints, Palms and Soles: An Introduction to Dermatoglyphics. Dover Publications, Inc., 1961.
[19] Biography,
“Juan
Vucetich
(18581925).”
https://www.nlm.nih.gov/
visibleproofs/galleries/biographies/vucetich.html,
February
2006.
[20] C. Wilson et al., “Fingerprint vendor technology evaluation 2003: Summary of results and
analysis report.” http://biometrics.nist.gov/cs_links/fpvte/report/
ir_7123_summary.pdf, June 2004. NISTIR 7123.
[21] P. Grother, G. W. Quinn, J. R. Matey, M. Ngan, W. Salamon, G. Fiumara, and C. Watson,
“IREX III: Performance of iris identiﬁcation algorithms.” http://www.nist.gov/
itl/iad/ig/irexiii.cfm, April 2012. NISTIR 7836.
[22] S. Z. Li and A. K. Jain, eds., Encyclopedia of Biometrics, vol. 1. Springer Science + Business
Media, 2009.
[23] E. R. Henry, Classiﬁcation and Uses of Finger Prints. London: Routledge, 1900.
[24] SWGFAST, “ANSI/NIST ITL 1-2000 American National Standard for Information Systems - data format for the interchange of ﬁngerprint, facial, & scar mark & tattoo (SMT) information.” http://fingerprint.nist.gov/standard/cdeffs/
Docs/SWGFAST_Memo.pdf, November 2005.
[25] A. M. Bazen, G. T. B. Verwaaijen, S. H. Gerez, L. P. J. Veelenturf, and B. J. van der Zwaag,
“A correlation-based ﬁngerprint veriﬁcation system,” in 11th Annual Workshop on Circuits,
Systems and Signal Processing ProRISC, pp. 205–213, November-December 2000.
[26] T. Hatano, T. Adachi, S. Shigematsu, H. Morimura, S. Onishi, Y. Okazaki, and H. Kyuragi,
“A ﬁngerprint veriﬁcation algorithm using the differential matching rate,” in International
Conference on Pattern Recognition ICPR, pp. 799–802, 2002.
[27] M. Tico and P. Kuosmanen, “Fingerprint matching using and orientation-based minutia descriptor,” IEEE Trans. on Pattern Analysis and Machine Intelligence, vol. 25, pp. 1009–
1014, August 2003.
[28] A. K. Jain and J. Feng, “Latent ﬁngerprint matching,” IEEE Trans. on Pattern Analysis and
Machine Intelligence, vol. 33, pp. 88–100, January 2011.
172

[29] N. K. Ratha, K. Karu, S. Chen, and A. K. Jain, “A real-time matching system for large
ﬁngerprint databases,” IEEE Trans. on Pattern Analysis and Machine Intelligence, vol. 18,
pp. 799–813, August 1996.
[30] X. Shuai, C. Zhang, and P. Hao, “Fingerprint indexing based on composite set of reduced
SIFT features,” in International Conference on Pattern Recognition (ICPR), pp. 1–4, December 2008.
[31] R. S. Germain, A. Califano, and S. Colville, “Fingerprint matching using transformation
parameter clustering,” IEEE Computational Science and Engineering, vol. 4, no. 4, pp. 42–
49, 1997.
[32] B. Bhanu and X. Tan, “Fingerprint indexing based on novel features of minutiae triplets,”
IEEE Trans. on Pattern Analysis and Machine Intelligence, vol. 25, no. 5, pp. 616–622,
2003.
[33] X. Jiang, M. Liu, and A. C. Kot, “Fingerprint retrieval for identiﬁcation,” IEEE Trans. on
Information Forensics and Security, vol. 1, pp. 532–542, December 2006.
[34] Y. Wang, J. Hu, and D. Phillips, “A ﬁngerprint orientation model based on 2D fourier expansion (FOMFE) and its application to singular-point detection and ﬁngerprint indexing,”
IEEE Trans. on Pattern Analysis and Machine Intelligence, vol. 29, pp. 573–585, April
2007.
[35] C. Champod, C. Lennard, P. Margot, and M. Stoilovic, Fingerprints and Other Ridge Skin
Impressions. CRC Press LLC, 2004.
[36] R. Cappelli, M. Ferrara, and D. Maltoni, “FVC-onGoing FIDX benchmark.” https://
biolab.csr.unibo.it/fvcongoing/UI/Form/ICB2013FIDX.aspx, 2013.
Accessed May 9, 2013.
[37] “FVC 2002 ﬁngerprint veriﬁcation competition.” http://bias.csr.unibo.it/
fvc2002/, 2002.
[38] R. Cappelli, M. Ferrara, and D. Maltoni, “Fingerprint indexing based on minutia cylindercode,” IEEE Trans. on Pattern Analysis and Machine Intelligence, vol. 33, pp. 1051–1057,
May 2011.
[39] B. Klare, “On a taxonomy of facial features,” in Biometrics: Theory, Applications and Systems (BTAS), pp. 1–8, September 2010.
[40] M. Turk and A. Pentland, “Eigenfaces for recognition,” Journal of Cognitive Neuroscience,
vol. 3, no. 1, pp. 71–86, 1991.
[41] P. N. Belhumeur, J. P. Hespanha, and D. Kriegman, “Eigenfaces vs. ﬁsherfaces: Recognition using class speciﬁc linear projection,” IEEE Trans. on Pattern Analysis and Machine
Intelligence, vol. 19, no. 7, pp. 711–720, 1997.
[42] NIST, “Face recognition grand challenge.” http://www.nist.gov/itl/iad/ig/
frgc.cfm.
173

[43] P. J. Grother, G. W. Quinn, and P. J. Phillips, “Multiple-biometric evaluation (MBE): Report
on the evaluation of 2D still-image face recognition algorithms.” http://www.nist.
gov/customcf/get_pdf.cfm?pub_id=905968, 2010. NISTIR 7709.
[44] J. G. Daugman, “High conﬁdence visual recognition of persons by a test of statistical independence,” IEEE Trans. on Pattern Analysis and Machine Intelligence, vol. 15, pp. 1148–
1161, November 1993.
[45] “Unique identiﬁcation authority of india.” https://portal.uidai.gov.in/
uidwebportal/dashboard.do#.
[46] “Biometrics provides economy and security.” http://www.scribd.com/doc/
313998/Jornal-Impressao-Digital, 8 2007.
[47] S. de Sainte Croix, “Brazil´ most secure voting ever.” http://riotimesonline.
s
com/brazil-news/front-page/brazils-most-secure-voting-ever/
#, March 2010.
[48] P. Kallender, “Japanese banks choose vein-recognition security system.” http:
//www.computerworld.com/s/article/95545/Japanese_banks_
choose_vein_recognition_security_system, August 2004.
[49] “Monozygotic twins.” Dorland’s Medical Dictionary for Health Consumers, 2007. http:
//medical-dictionary.thefreedictionary.com/monozygotic+twins.
[50] M. Indovina, V. Dvornychenko, E. Tabassi, G. Quinn, P. Grother, S. Meagher, and M. Garris,
“An evaluation of automated latent ﬁngerprint identiﬁcation technology (Phase II),” tech.
rep., NIST Interagency/Internal Report (NISTIR) - 7577, April 2009.
[51] K. D. Kochanek, S. E. Kirmeyer, J. A. Martin, D. M. Strobino, and B. Guyer, “Annual
summary of vital statistics: 2009,” Pediatrics, vol. 129, pp. 338–348, February 2012.
[52] J. A. Martin, H.-C. Kung, T. J. Mathews, D. L. Hoyert, D. M. Strobino, B. Guyer, and S. R.
Sutton, “Annual summary of vital statistics: 2006,” Pediatrics, vol. 121, no. 4, pp. 788–801,
2008.
[53] J. G. Hall, “Twins and twinning,” American Journal of Medical Genetics, vol. 61, pp. 202–
204, 1996.
[54] J. J. Nora, F. C. Fraser, J. Bear, C. R. Greenberg, D. Patterson, and D. Warburton, Medical
Genetics: Principles and Practice, ch. Twins and Their Use in Genetics, pp. 395–402. Lea
& Febiger, 4 ed., 1993.
[55] “Identical twins guilty in bank fraud scheme,” United Press International, September 24,
1986. Section: Domestic News.
[56] “Twin convicted in third rape trial,” March 22nd 2006. http://origin.foxnews.
com/story/0,2933,188801,00.html.

174

[57] Digital Persona, “Unimed case study.” http://www.digitalpersona.com/
uploadedFiles/Collateral/Case_Studies/CS_Unimed20080905.pdf,
2008.
[58] A. K. Jain, S. Prabhakar, and S. Pankanti, “On the similarity of identical twin ﬁngerprints,”
Pattern Recognition, vol. 35, pp. 2653–2663, 2002.
[59] A. W.-K. Kong, D. Zhang, and G. Lu, “A study of identical twins’ palmprints for personal
veriﬁcation,” Pattern Recognition, vol. 39, pp. 2149–2156, 2006.
[60] K. Kodate, R. Inaba, E. Watanabe, and T. Kamiya, “Facial recognition by a compact parallel
optical correlator,” Measurement Science and Technology, vol. 13, pp. 1756–1766, 2002.
[61] A. Ariyaeeinia, C. Morrison, A. Malegaonkar, and S. Black, “A test of the effectiveness of
speaker veriﬁcation for differentiating between identical twins,” Science and Justice, vol. 48,
pp. 182–186, 2008.
[62] J. Daugman and C. Downing, “Epigenetic randomness, complexity and singularity of human
iris patterns,” in Proceedings of the Royal Society of London, vol. 268, pp. 1737–1740, 2001.
[63] Y. Han, C. Ryu, J. Moon, H. Kim, and H. Choi, “A study on evaluating the uniqueness of
ﬁngerprints using statistical analysis,” in International Conference on Information Security
and Cryptology, pp. 467–477, 2005.
[64] H. A. Patil and T. K. Basu, “The teager energy based features for identiﬁcation of identical
twins in multi-lingual environment,” in International Conference on Neural Information
Processing, pp. 333–337, 2004.
[65] A. M. Bronstein, M. M. Bronstein, and R. Kimmel, “Three-dimensional face recognition,”
International Journal of Computer Vision, vol. 64, no. 1, pp. 5–30, 2005.
[66] S. N. Srihari, H. Srinivasan, and G. Fang, “Discriminability of ﬁngerprints of twins,” Journal
of Forensic Identiﬁcation, vol. 58, no. 1, pp. 109–127, 2008.
[67] Z. Sun, A. A. Paulino, J. Feng, Z. Chai, T. Tan, and A. K. Jain, “A study of multibiometric
traits of identical twins,” in Biometric Technology for Human Identifcation VII, vol. SPIE
7667, 2010.
[68] K. Hollingsworth, K. W. Bowyer, S. Lagree, S. P. Fenker, and P. J. Flynn, “Genetically
identical irises have texture similarity that is not detected by iris biometrics,” Computer
Vision and Image Understanding, vol. 115, pp. 1493–1502, 11 2011.
[69] M. T. Pruitt, J. M. Grant, J. R. Paone, and P. J. Flynn, “Facial recognition of identical twins,”
in International Joint Conference on Biometrics (IJCB), pp. 1–8, 10 2011.
[70] S. Biswas, K. W. Bowyer, and P. J. Flynn, “A study of face recognition of identical twins by
humans,” in International Workshop on Information Forensics and Security (WIFS), pp. 1–6,
December 2011.

175

[71] B. Klare, A. A. Paulino, and A. K. Jain, “Analysis of facial features in identical twins,” in
International Joint Conference on Biometrics (IJCB), pp. 1–8, October 2011.
[72] V. Vijayan, K. W. Bowyer, P. J. Flynn, D. Huang, L. Chen, M. Hansen, O. Ocegueda, S. K.
Shah, and I. A. Kakadiaris, “Twins 3D face recognition challenge,” in International Joint
Conference on Biometrics (IJCB), pp. 1–7, 10 2011.
[73] N. Srinivas, G. Aggarwal, P. Flynn, and R. W. V. Bruegge, “Analysis of facial marks to
distinguish between identical twins,” IEEE Trans. on Information Forensics and Security,
vol. 7, pp. 1536–1550, October 2012.
[74] X. Tao, X. Chen, X. Yang, and J. Tian, “Fingerprint recognition with identical twin ﬁngerprints,” PLoS ONE, vol. 7, p. e35704, 4 2012.
[75] J. M. Lytle, “Biometric face scanner tells identical twins apart,” May 9th
2008. http://www.techradar.com/news/world-of-tech/future-tech/
face-scanner-tells-identical-\\twins-apart-363150.
[76] U. Park and A. K. Jain, “Face matching and retrieval using soft biometrics,” IEEE Trans. on
Information Forensics and Security, vol. 5, pp. 406–415, 9 2010.
[77] Z. Sun and T. Tan, “Ordinal measures for iris recognition,” IEEE Trans. on Pattern Analysis
and Machine Intelligence, vol. 31, no. 12, pp. 2211–2226, 2009.
[78] “Symwave: BioPrint USB Portable and Peripheral Solutions.” http://www.symwave.
com/prod\_biometric.shtml.
[79] Neurotechnology Inc, “Neurotechnology:
VeriFinger SDK.”
neurotechnology.com/verifinger.html.

http://www.

[80] A. A. Aﬁﬁ and S. P. Azen, Statistical Analysis: A computer oriented approach. New York:
Academic Press, 1972.
[81] “Cognitec:
FaceVACS - SDK.”
FaceVACS-SDK.19.0.html.

http://www.cognitec-systems.de/

[82] “SDK-Beijing IrisKing Ltd. Co..” www.irisking.com/en/sdk.html.
[83] Neurotechnology Inc, “Neurotechnology:
VeriEye
neurotechnology.com/verieye.html.

SDK.”

http://www.

[84] A. K. Jain, P. Flynn, and A. A. Ross, Handbook of Biometrics, ch. 23. Springer, 2008.
[85] D. R. Ashbaugh, Quantitative-Qualitative Friction Ridge Analysis. CRC Press, 1999.
[86] FBI, “Next generation identiﬁcation.” http://www.fbi.gov/about-us/cjis/
fingerprints\_biometrics/ngi.
[87] A. A. Paulino, A. K. Jain, and J. Feng, “Latent ﬁngerprint matching: Fusion of manually
marked and derived minutiae,” in 23rd SIBGRAPI - Conference on Graphics, Patterns and
Images, pp. 63–70, August 2010.
176

[88] A. A. Paulino, J. Feng, and A. K. Jain, “Latent ﬁngerprint matching using descriptor-based
Hough transform,” IEEE Trans. on Information Forensics and Security, vol. 8, pp. 31–45,
January 2013.
[89] J. Feng and A. K. Jain, “Fingerprint reconstruction: from minutiae to phase,” IEEE Trans.
on Pattern Analysis and Machine Intelligence, vol. 33, pp. 209–223, February 2011.
[90] NIST Special Database 14, “Mated ﬁngerprint card pairs 2.” http://www.nist.gov/
srd/nistsd14.cfm.
[91] R. Cappelli, M. Ferrara, and D. Maltoni, “Minutia cylinder-code: a new representation
and matching technique for ﬁngerprint recognition,” IEEE Trans. on Pattern Analysis and
Machine Intelligence, vol. 32, pp. 2128–2141, December 2010.
[92] J. Qi, S. Yang, and Y. Wang, “Fingerprint matching combining the global orientation ﬁeld
with minutia,” Pattern Recognition Letters, vol. 26, pp. 2424–2430, 2005.
[93] J. Gu, J. Zhou, and C. Yang, “Fingerprint recognition by combining global structure and
local cues,” IEEE Trans. on Image Processing, vol. 15, pp. 1952–1964, July 2006.
[94] X. Jiang and W.-Y. Yau, “Fingerprint minutiae matching based on the local and global structures,” in International Conference on Pattern Recognition (ICPR), pp. 1038–1041, 2000.
[95] X. Chen, J. Tian, and X. Yang, “A new algorithm for distorted ﬁngerprints matching
based on normalized fuzzy similarity measure,” IEEE Trans. on Image Processing, vol. 15,
pp. 767–776, March 2006.
[96] W. Xu, X. Chen, and J. Feng, “A robust ﬁngerprint matching approach: Growing and fusing of local structures,” in International Conference on Biometrics (ICB), (Seoul, Korea),
pp. 134–143, August 2007.
[97] J. Feng, “Combining minutiae descriptors for ﬁngerprint matching,” Pattern Recognition,
vol. 41, no. 1, pp. 342–352, 2008.
[98] J. Feng, S. Yoon, and A. K. Jain, “Latent ﬁngerprint matching: Fusion of rolled and plain
ﬁngerprints,” in International Conference on Biometrics (ICB), June 2009.
[99] A. K. Jain, J. Feng, A. Nagar, and K. Nandakumar, “On matching latent ﬁngerprints,” in
CVPR Workshops on Biometrics, pp. 1–8, June 2008.
[100] S. Yoon, J. Feng, and A. K. Jain, “On latent ﬁngerprint enhancement,” in SPIE, vol. 7667766707, April 2010.
[101] S. Yoon, J. Feng, and A. K. Jain, “Latent ﬁngerprint enhancement via robust orientation
ﬁeld estimation,” in International Joint Conference on Biometrics (IJCB), October 2011.
[102] J. Feng, J. Zhou, and A. K. Jain, “Orientation ﬁeld estimation for latent ﬁngerprint enhancement,” IEEE Trans. on Pattern Analysis and Machine Intelligence, vol. 35, pp. 925–940,
April 2013.

177

[103] A. Sankaran, T. I. Dhamecha, M. Vatsa, and R. Singh, “On matching latent to latent ﬁngerprints,” in International Joint Conference on Biometrics (IJCB), pp. 1–6, October 2011.
[104] M. Vatsa, R. Singh, A. Noore, and K. Morris, “Simultaneous latent ﬁngerprint recognition,”
Applied Soft Computing, vol. 11, pp. 4260–4266, October 2011.
[105] NIST, “Evaluation of latent ﬁngerprint technologies.” http://www.nist.gov/itl/
iad/ig/latent.cfm.
[106] NIST, “Summary of the results of Phase I ELFT testing.” http://biometrics.
nist.gov/cs\_links/latent/elft07/phase1\_aggregate.pdf, September 2007.
[107] M. Indovina, R. A. Hicklin, and G. I. Kiebuzinski, “NIST evaluation of latent ﬁngerprint
technologies: Extended feature sets [Evaluation #1],” tech. rep., NIST Interagency/Internal
Report (NISTIR) - 7775, April 2011.
[108] B. T. Ulery, R. A. Hicklin, J. Buscaglia, and M. A. Roberts, “Accuracy and reliability of
forensic latent ﬁngerprint decisions,” Proc. of the National Academy of Sciences of the USA,
vol. 108, no. 19, pp. 7733–7738, 2011.
[109] B. T. Ulery, R. A. Hicklin, J. Buscaglia, and M. A. Roberts, “Repeatability and reproducibility of decisions by latent ﬁngerprint examiners,” PLoS ONE, vol. 7, p. e32800, March 2012.
[110] I. E. Dror, D. Charlton, and A. E. Peron, “Contextual information renders experts vulnerable
to making erroneous identiﬁcations,” Forensic Science International, vol. 156, no. 1, pp. 74–
78, 2006.
[111] B. G. Sherlock and D. M. Monro, “A model for interpreting ﬁngerprint topology,” Pattern
Recognition, vol. 26, no. 7, pp. 1047–1055, 1993.
[112] J. Zhou and J. Gu, “Modeling orientation ﬁelds of ﬁngerprints with rational complex functions,” Pattern Recognition, vol. 37, pp. 389–391, 2004.
[113] S. Huckemann, T. Hotz, and A. Munk, “Global models for the orientation ﬁeld of ﬁngerprints: an approach based on quadratic differentials,” IEEE Trans. on Pattern Analysis and
Machine Intelligence, vol. 30, pp. 1507–1519, September 2008.
[114] J. Feng and J. Zhou, “A performance evaluation of ﬁngerprint minutia descriptors,” in International Conference on Hand-Based Biometrics, pp. 1–6, 2011.
[115] L. Hong, Y. Wan, and A. K. Jain, “Fingerprint enhancement: Algorithm and performance
evaluation,” IEEE Trans. on Pattern Analysis and Machine Intelligence, vol. 20, pp. 777–
789, August 1998.
[116] T. K. Ho, J. J. Hull, and S. N. Sihari, “Decision combination in multiple classiﬁer systems,”
IEEE Trans. on Pattern Analysis and Machine Intelligence, vol. 16, pp. 66–75, January
1994.
[117] A. K. Jain, P. Flynn, and A. A. Ross, eds., Handbook of Biometrics. Springer, 2008.
178

[118] R. Cappelli, M. Ferrara, D. Maltoni, and M. Tistarelli, “MCC: a baseline algorithm for
ﬁngerprint veriﬁcation in FVC-onGoing,” in 11th Int. Conf. Control, Automation, Robotics
and Vision, pp. 19–23, December 2010.
[119] Biometric
System
Laboratory
team,
https://biolab.csr.unibo.it/FVCOnGoing/UI/Form/Home.aspx.

“FVC-onGoing.”

[120] A. A. Paulino, J. Feng, and A. K. Jain, “Latent ﬁngerprint matching using descriptor-based
Hough transform,” in International Joint Conference on Biometrics (IJCB), pp. 1–7, October
2011.
[121] V. N. Dvornychenko, “Evaluation of fusion methods for latent ﬁngerprint matchers,” in
International Conference on Biometrics (ICB), March-April 2012.
[122] A. A. Ross, K. Nandakumar, and A. K. Jain, Handbook of Multibiometrics. Springer, 2006.
[123] The Federal Bureau of Investigation (FBI), “Next generation identiﬁcation.” http://
www.fbi.gov/about-us/cjis/fingerprints_biometrics/ngi.
[124] A. Lumini, D. Maio, and D. Maltoni, “Continuous versus exclusive classiﬁcation for ﬁngerprint retrieval,” Pattern Recognition Letters, vol. 18, pp. 1027–1034, 1997.
[125] X. Tan, B. Bhanu, and Y. Lin, “Fingerprint identiﬁcation: Classiﬁcation vs. indexing,” in
IEEE Conference on Advanced Video and Signal Based Surveillance, AVSS ’03, (Washington, DC, USA), pp. 151–156, IEEE Computer Society, 2003.
[126] A. K. Jain, Y. Chen, and M. Demirkus, “Pores and ridges: High-resolution ﬁngerprint
matching using level 3 features,” IEEE Trans. on Pattern Analysis and Machine Intelligence,
vol. 29, no. 1, pp. 15–27, 2007.
[127] A. Gionis, P. Indyk, and R. Motwani, “Similarity search in high dimensions via hashing,”
in Proceedings of the 25th International Conference on Very Large Data Bases, VLDB ’99,
pp. 518–529, 1999.
[128] S.-O. Lee, Y.-G. Kim, and G.-T. Park, “A feature map consisting of orientation and interridge spacing for ﬁngerprint retrieval,” in Audio- and Video-Based Biometric Person Authentication (T. Kanade, A. Jain, and N. K. Ratha, eds.), vol. 3546 of Lecture Notes in Computer
Science, pp. 184–190, Springer Berlin Heidelberg, 2005.
[129] M. Liu, X. Jiang, and A. C. Kot, “Fingerprint retrieval by complex ﬁlter responses,” in
International Conference on Pattern Recognition ICPR, pp. 1042–1045, 2006.
[130] J. Li, W.-Y. Yau, and H. Wang, “Fingerprint indexing based on symmetrical measurement,”
in International Conference on Pattern Recognition ICPR, pp. 1038–1041, 2006.
[131] A. Gyaourova and A. Ross, “A novel coding scheme for indexing ﬁngerprint patterns,” in
Joint IAPR International Workshop on Structural, Syntactic, and Statistical Pattern Recognition, (Berlin, Heidelberg), pp. 755–764, Springer-Verlag, 2008.

179

[132] R. Cappelli, “Fast and accurate ﬁngerprint indexing based on ridge orientation and frequency,” IEEE Trans. on Systems, Man, and Cybernetics, Part B: Cybernetics, vol. 41,
pp. 1511–1521, December 2011.
[133] R. Cappelli and M. Ferrara, “A ﬁngerprint retrieval system based on level 1 and level 2
features,” Expert Systems with Applications, vol. 39, pp. 10465–10478, 2012.
[134] M. Liu and P.-T. Yap, “Invariant representation of orientation ﬁelds for ﬁngerprint indexing,”
Pattern Recognition, vol. 45, pp. 2532–2542, 2012.
[135] X. Liang, A. Bishnu, and T. Asano, “A robust ﬁngerprint indexing scheme using minutia neighborhood structure and low-order delaunay triangles,” IEEE Trans. on Information
Forensics and Security, vol. 2, pp. 721–733, December 2007.
[136] O. Iloanusi, A. Gyaourova, and A. Ross, “Indexing ﬁngerprints using minutiae quadruplets,”
in IEEE CVPR Workshop on Biometrics, pp. 127–133, June 2011.
[137] A. A. Paulino, E. Liu, K. Cao, and A. K. Jain, “Latent ﬁngerprint indexing: Fusion of level
1 and level 2 features,” in Biometrics: Theory, Applications and Systems (BTAS), (Washington, D.C.), September 29 - October 2 2013.
[138] K. Nilsson and J. Bigun, “Localization of corresponding points in ﬁngerprints by complex
ﬁltering,” Pattern Recognition Letters, vol. 24, no. 13, pp. 2135–2144, 2003.
[139] K. Nandakumar, Y. Chen, S. C. Dass, and A. K. Jain, “Likelihood ratio-based biometric
score fusion,” IEEE Trans. on Pattern Analysis and Machine Intelligence, vol. 30, pp. 342–
347, February 2008.

180