.14 ,

.51. .2

. . it
. ’mn.

m

 

 

 

 

1v... Huh 3
).\\t.x
£5va

. 3.. .

If

eight:
« )3.
r n: J' . .
.t‘r’..v it}
3.3 Pl
. x: l!
Kakvtxhwfikwufi
a .

 

a 1 A
Hal... i...
3 .x Syd... .i
fir... ,
5.... xi

, .3
2.53:; ._.. .5 .e

. t
. . 4. . - .fiugwfiksfilnlo .

 

 

P...

affifinnflw ... up: : ...37., H_ .. . . 1,. w. . . . ..

THESIS

-(
l

J

(L3 a 0 9.

This is to certify that the

dissertation entitled

FACE DETECTION AND MODELING FOR RECOGNITION

presented by

Rein-Lien Hsu

has been accepted towards fulfillment
of the requirements for

 

 

Doctoral degree in Computer Science
& Engineering

MW

 

Major professor

Date 3W loll 10° 1/

 

MS U i: an Affirmative Action/Equal Opportunity Institution 042771

 

 

LIBRARY i
Michigan State
University

 

 

 

PLACE IN RETURN Box to remove this checkout from your record.
TO AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

DATE DUE DATE DUE DATE DUE

 

 

 

 

 

 

 

 

 

 

 

 

 

 

6/01 c:/CIRC/DaleDue.p6&p. 15

 

FACE DETECTION AND MODELING FOR RECOGNITION
By

Rein-Lien Hsu

A DISSERTATION

Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of

DOCTOR OF PHILOSOPHY

Department Of Computer Science & Engineering

2002

ABSTRACT
FACE DETECTION AND MODELING FOR RECOGNITION
By

Rein-Lien Hsu

Face recognition has received substantial attention from researchers in biometrics,
computer vision, pattern recognition, and cognitive psychology communities because
of the increased attention being devoted to security, man-machine communication,
content-based image retrieval, and image/video coding. We have proposed two au-
tomated recognition paradigms to advance face recognition technology. Three major
tasks involved in face recognition systems are: (i) face detection, (ii) face modeling,
and (iii) face matching. We have developed a face detection algorithm for color images
in the presence Of various lighting conditions as well as complex backgrounds. Our
detection method first corrects the color bias by a lighting compensation technique
that automatically estimates the parameters of reference white for color correction.
We overcame the difficulty of detecting the low-luma and high-luma skin tones by
applying a nonlinear transformation to the YCbC'r color space. Our method gener-
ates face candidates based on the spatial arrangement. of detected Skin patches. we

constructed eye, mouth, and face boundary maps tO verify each face candidate. Ex-

perimental results demonstrate successful detection of faces with different Sizes, color,
position, scale, orientation, 3D pose, and expression in several photo collections.

3D human face models augment the appearance-based face recognition approaches
to assist face recognition under the illumination and head pose variations. For the two
proposed recognition paradigms, we have designed two methods for modeling human
faces based on (i) a generic 3D face model and an individual’s facial measurements of
Shape and texture captured in the frontal view. and (ii) alignment of a semantic face
graph, derived from a generic 3D face model, onto a frontal face image. Our mod-
eling methods adapt recognition-oriented facial features of a generic model to those
extracted from facial measurements in a global—tO-local fashion. The first modeling
method uses displacement propagation and 2.5D snakes for model alignment. The
resulting 3D face model is visually similar to the true face, and proves to be quite
useful for recognizing non-frontal views based on an appearance-based recognition
algorithm. The second modeling method uses interacting snakes for graph alignment.
A successful interaction of snakes (associated with eyes, mouth, nose, etc.) results in
appropriate component weights based on distinctiveness and visibility of individual
facial components. After alignment, facial components are transformed to a feature
space and weighted for semantic face matching. The semantic face graph facilitates
face matching based on selected components, and effective 3D model updating based
on 2D images. The results of face matching demonstrate that the proposed model
can lead to classification and visualization (e.g., the generation of cartoon faces and

facial caricatures) Of human faces using the derived semantic face graphs.

© Copyright 2002 by Rein-Lien Hsu
All Rights Reserved

To my parents; my lovely wife, Pei-Jing; and my son, Alan

ACKNOWLEDGMENTS

First of all, I would like to thank all the individuals who have helped me during
my Ph.D. study at Michigan State University. I would like to express my deepest
gratitude to my advisor, Dr. Anil K. Jain, for his guidance in academic research and
his support in daily life. He broadened my view in research areas, especially in pattern
recognition and computer vision, and taught me how to focus on research problems.
I will never forget his advice “Just do it,” while being caught in multiple tasks at the
same time. I am grateful to my Ph.D. committee, Dr. Mohamed Abdel—Mottaleb,
Dr. George Stockman, Dr. John J. Weng, and Dr. Sarat C. Dass, for their valuable
ideas, suggestions, and encouragement.

I would also like to thank Dr. Chant-Chin Chen and Dr. Wey-Shiuan Hwang
for their help at the beginning of my study at MSU, and Dr. Shaoyun Chen and
Yonghong Li for their help in the NASA modeling project. I am very grateful to Dr.
Helen Shen and Dr. Mihran Tuceryan for their numerous suggestions and discussions
on model compression. Special thanks are due to Philips Research-USA for Offering
me summer internships in 2000 and 2001; to Dr. Mohamed Abdel—h‘lottaleb for
his guidance and suggestions in my work on face detection; to Dr. Patrick Flynn

vi

for providing the range datasets; to Dr. Wey-Shiuan Hwang for providing his face
recognition software; and to Dennis Bond for his help in creating a graphical user
interface for face editing.

Thanks are also due to Cathy M. Davison, Linda. Moore, Starr Portice, and Beverly
J. Wallace for their assistance in the administrative tasks. Special thanks to all the
Prippies: Lin Hong, Aditya Vailaya, Nicolae Duta, Salil Prabhakar, Dan Gutchess,
Paul Albee, Arun Ross, Anoop N amboodiri, Silviu Minut, Umut Uludag, Xiaoguang
Lu, Martin Law, Miguel Figueroa—Villanue, and Yilu Zhang for their help during my
stay in the PRIP Lab in the Department of Computer Science and Engineering at
MSU. I would also like to thank Michael E. Farmer for giving me an opportunity to
work on human tracking research.

I would like to thank Mark H. McCullen for mentoring me to be a teaching as-
sistant in CSE232; Dr. Jeffrey A. Fessler at the University of Michigan, Ann Arbor,
for his valuable help during my transfer to the Dept. of Computer Science at h’lichi—
gan State University; and Dr. Yung-Nien Sun and Dr. Chin-Hsing Chen in Taiwan
for their encouragement and spiritual support. Special thanks to NASA, Philips
Research-USA, Eaton corporation, and ONR (grant NO. N00014-01-1-0266) for their
financial support during my Ph.D studies.

Finally, but not the least, I would like to thank my parents, my wife, Dr. Pei-jing

Li, and my son, Alan, for all the happiness they have shared with me.

vii

TABLE OF CONTENTS

LIST OF TABLES
LIST OF FIGURES

1 Introduction

1.1 Challenges in Face Recognition ..........
1.2 Semantic Facial Components ............
1.3 Face Recognition Systems .............
1.4 Face Detection and Recognition ..........
1.5 Face Modeling for Recognition ...........
1.5.1 Face Alignment Using 2.5D Snakes .......
1.5.2 ' Model Compression ................
1.5.3 Face Alignment Using Interacting Snakes . . . .
1.6 Face Retrieval ....................
1.7 Outline of Dissertation ...............
1.8 Dissertation Contributions .............

2 Literature Review

2.1 Face Detection ...................
2.2 Face Recognition ..................
2.3 Face Modeling ....................
2.3.1 Generic Face Models ...............
2.3.2 Snakes for Face Alignment .............
2.3.3 3D Model Compression .............
2.4 Face Retrieval ....................
2.5 Summary ......................

3 Face Detection

3.1 Face Detection Algorithm .............
3.2 Lighting Compensation and Skin Tone Detection
3.3 Localization of Facial Features ...........
3.3.1 Eye Map .....................
3.3.2 Mouth Map ....................
3.3.3 Eye and Mouth Candidates ...........
3.3.4 Face Boundary Map ...............
3.3.5 Weight Selection for a Face Candidate .....
3.4 Experimental Results ................
3.5 Summary ......................

nnnnnnnnnnnnn

ooooooooooooo

ooooooooooooo

ooooooooooooo

ooooooooooooo

ooooooooooooo

ooooooooooooo

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

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

ooooooooooooo

ooooooooooooo

ccccccccccccc

nnnnnnnnnnnnn

4 Face Modeling 97

4.1 Modeling Method ............................... 97
4.2 Generic Face Model .............................. 99
4.3 Facial Measurements ............................. 101
4.4 Model Construction .............................. 103
4.5 Summary ................................... 1 10
5 Semantic Face Recognition 111
5.1 Semantic Face Graph as Multiple Snakes .................. 112
5.2 Coarse Alignment of Semantic Face Graph ................. 115
5.3 Fine Alignment of Semantic Face Graph via Interacting Snakes ..... 118
5.3.1 Interacting Snakes and Energy Functional ................ 120
5.3.2 Parametric Active Contours ........................ 127
5.3.3 Geodesic Active Contours ......................... 127
5.4 Semantic Face Matching ........................... 130
5.4.1 Component Weights and Matching Cost ................. 132
5.4.2 Face Matching Algorithm ......................... 133
5.4.3 Face Matching ............................... 134
5.5 Facial Caricatures for Recognition and Visualization ........... 140
5.6 Summary ................................... 143
6 Conclusions and Future Directions 144
6.1 Conclusions .................................. 144
6.2 Future Directions ............................... 147
6.2.1 Face Detection & 'ITacking ......................... 147
6.2.2 Face Modeling ............................... 149
6.2.3 Face matching ............................... 151
APPENDICES 153
A Transformation of Color Space 153
A.1 Linear Transformation ............................ 153
A2 Nonlinear Transformation .......................... 154
A3 Skin Classifier ................................. 156
B Distance between Skin Patches 157
C Image Processing Template Library (IPTL) 160
CI Image and Image Template ......................... 160
G2 Example Code ................................ 163
BIBLIOGRAPHY 165

ix

2.1
2.2
2.3

3.1

3.2

5.1
5.2

LIST OF TABLES

Summary of various face detection approaches ................ 34
Geometric compression efficiency. ...................... 53
Summary of performance Of various face detection approaches ....... 56

Detection results on the HHI image database (Image size 640 x 480) on a

PC with 1.7 GHz CPU. FP: False Positives, DR: Detection Rate. . . 88
Detection results on the Champion database (Image size ~ 150 x 220) on

a PC with 860 MHz CPU. FP: False Positives, DR: Detection Rate. . 88

Error rates on a 50—image database ...................... 136
Dimensions of the semantic graph descriptors for individual facial compo—
nents. ................................... 136

1.1

1.1
1.1
1.2

1.3

1.4

1.5

1.6

1.7

1.8

LIST OF FIGURES

Applications using face recognition technology: (a) and (b) automated
video surveillance (downloaded from Visionics [1] and FaceSnap [2],
respectively); (c) and (d) access control (from Visionics [1] and from
Viisage [3], respectively); (e) management of photo databases (from Vi-
isage [3]); (f) multimedia communication (from Eyematic [4]). Images
in this dissertation are presented in color .................

(Cont’d). ...................................
(Cont’d). ...................................

Comparison of various biometric features: (a) based on zephyr analysis
(downloaded from [5]); (b) based on MRTD compatibility (from [6]). .

Intra—subject variations in pose, illumination, expression, occlusion, acces—
sories (e.g., glasses), color, and brightness. ...............

Face comparison: (a) face verification/ authentication; (b) face identifica-

tion / recognition. Face images are taken from the MSU face database [7].

Head recognition versus face recognition: (a) Clinton and Gore heads with
the same internal facial features, adapted from [8]; (b) two faces of
different subjects with the same internal facial components Show the
important role of hair and face outlines in human face recognition. . .

Caricatures of (a) Vincent Van Gogh; (b) Jim Carrey; (c) Arnold
Schwarzenegger; (d) Einstein; (e) G. W. Bush; and (f) Bill Gates.
Images are downloaded from [9], [10] and [10]. Caricatures reveal the
use of component weights in face identification. ............

Cartoons reveal that humans can easily recognize characters whose facial
components are depicted by simple line strokes and color characteris-
tics: (a) and (b) are frames adapted from the movie Pocahontas; (c)
and (d) are frames extracted from the movie Little Mermaid II. (Disney
Enterprises, Inc.) .............................

Configuration of facial components: (a) face image; (b) face image in (a)
with enlarged eyebrow-to—eye and nose-to—mouth distances; (c) inverted
face of the image in (b). A small change of component configuration
results in a significantly different facial appearance in an upright face
in (b); however, this change may not be perceived in an inverted face
in (c). ...................................

xi

1.9

1.10

1.11

1.12

1.13

1.14

1.15

1.16

1.17

2.1

2.1
2.2

2.3

2.4

2.5

Facial features/ components: (a) five kinds of facial features (i.e., eyebrows,
eyes, nose, ears, and mouth) in a face for reading faces in physiognomy
(downloaded from [11]); (b) a frontal semantic face graph, whose nodes
are facial components that are filled with different Shades. ......

Similarity of frontal faces between (a) twins (downloaded from [12]); and
(b) a father and his son (downloaded from [13]). ............

System diagram of our 3D model-based face recognition system using reg-
istered range and color images. .....................

System diagram of our 3D model-based face recognition system without
the use of range data ............................

Face images taken under unconstrained environments: (a) a crowd of peo-
ple (downloaded from [14]); (b) a photo taken at a swimming pool.

Face images for our detection algorithm: (a) a montage image containing
images adapted from MPEG7 content set [15]; (b) a family photo. . .

Face images not suitable for our detection algorithm: (a) cropped image
(downloaded from [16]); (b) a performer wearing make-up (from [14]);
(c) people wearing face masks (from [14]). ...............

Graphical user interfaces of the FaceGen modeller [17]. A 3D face model
shown (a) with texture mapping; (b) with wireframe overlaid.

A face retrieval interface of the FACEit system [18]: the system gives the
most similar face in a database given a query face image. .......

Outputs of several face detection algorithms; ((a), ((b) Féraud et al. [19];

(c ) Maio et a1. [20],( d) Garciha) et a1. [21];( (f) Schneid)erman et a1.
[22];( g)Rowley et a1. [23];( ),(i) Rowleye et a1. [24];( j)Sung et a1.
[25];( k) Yow et al. [26]; (1) Lew et al. [27] ................
(Cont’d). ...................................

Examples of face images are selected from (a) the FERET database [28];
(b) the MIT database [29]; (c) the XM2VTS database [30] .......
Internal representations of the PCA-based approach and the LDA-based
approach (from Weng and Swets [31]). The average (mean) images are
shown in the first column. Most Expressive Features (MEF) and Most

Discriminating Features (MDF) are shown in (a) and (b), respectively.

Internal representations of the EBGM- based approach (from Wiskott et
a1. [32]): )a graph is overlaid on a face Image; (b) a reconstruction
of the Image from the graph, (c ) a reconstruction of the Image from a
face bunch graph using the best fitting jet at each node. Images are
downloaded from [33]; (d) a bunch graph whose nodes are associated
with a bunch of jets [33]; (e) an alternative interpretation of the concept
of a bunch graph [33]. ..........................
Internal representations of the LFA- based approach (fiom Penev and Atirk
[34]). a)An average face Image is marked with five localized features;
(b) five atopographic kernels associated with the five localized features
are shown in the top row, and the corresponding residual correlations
are shown in the bottom row. ......................

xii

11

13

16

17

18

20

20

22

27

36
37

41

42

43

2.6 A breakdown of face recognition algorithms based on the pose-dependency,
face representation, and features used in matching. ..........
2.7 Face modeling using anthropometric measurements (downloaded

from [35]): (a) anthropometric measurements; (b) a B—spline face model. 48

2.8 Generic face models: (a) Water’s animation model; (b) anthropometric
measurements; (b) six kinds of face models for representing general
facial geometry. ..............................

3.1 Face detection algorithm. The face localization module finds face can-
didates, which are verified by the detection module based on facial
features. ..................................

3.2 Skin detection: (a) a yellow-biased face image; (b) a lighting compensated
image; (c) skin regions of (a) shown in white; (d) Skin regions of (b). .

3.3 The YCbC', color space (blue dots represent the reproducible color on
a monitor) and the skin tone model (red dots represent skin color
samples). (a) The YCbC, space; (b) a 2D projection in the 05-0,
subspace; (c) a 2D projection in the (Cb/Y)-(C,./Y) subspace. . . . .

3.4 The dependency of Skin tone color on luma. The skin tone cluster (red
dots) is Shown in (a) the rgY, (c) the CI E 233/)”, and (e) the H S V color
Spaces; the 2D projection of the cluster is shown in (b) the r — g, (d)
the a: — y , and (f) S — H color subspaces, where blue dots represent the
reproducible color on a monitor. For a better presentation of cluster
shape, we normalize the luma Y in the rgY and the CI E scyY by 255,
and swap the hue and saturation coordinates in the H S V Space. The

49

65

skin tone cluster is less compact at low saturation values in (e) and (f). 66

3.5 2D projections of the 3D skin tone cluster in (a) the Y—C'l> subspace; (b)
the Y—C, subspace. Red dots indicate the Skin cluster. Three blue
dashed curves, one for cluster center and two for boundaries, indicate
the fitted models. .............................

3.6 The nonlinear transformation of the YC’bC, color Space. (a) The trans-
formed YCbC, color space; (b) a 2D projection of (a) in the Cb-Cr
subspace, in which the elliptical Skin model is overlaid on the skin
cluster ....................................

3.7 Nonlinear color transform. Six detection examples, with and without the
transform are shown. For each example, the images shown in the first
column are Skin regions and detections without the transform, while
those in the second column are results with the transform. ......

3.8 Construction of the face mask. (a) Face candidates; (b) one of the face
candidates; (c) grouped skin areas; (d) the face mask ..........

3.9 Construction of eye maps: (a) from chroma; (b) from luma; (c) the com-
bined eye map. ..............................

3.10 An example of a hemispheric structuring element for grayscale morpho—
logical dilation and erosion with a = 1. .................

3.11 Construction of the mouth map ........................

3.12 Computation of face boundary and the eye—mouth triangle .........

xiii

67

67

69

72

74
77

3.13

3.14

3.15

3.16

3.17

3.18

3.19

3.20

3.21

3.22

3.22
3.22
3.23

3.24

4.1

Geometry of an eye-mouth triangle, where vi = —135; unit vectors if} and
u} are perpendicular to the interocular segment and the horizontal axis,
respectively. ................................

Attenuation term, e’3(1“c"32(9r(‘ijik))), plotted as a function of the angle 6,
(in degrees) has a maximal value of 1 at 6’, = 0°, and a value of 0.5 at
6, = 25° ...................................

Face detection examples containing dark skin-tone faces. Each example
contains an input image, grouped skin regions Shown in pseudo color,
and a lighting-compensated image overlaid with detected face and fa—
cial features .................................

Face detection results on closed-eye or open-mouth faces. Each example
contains an original image (top) and a lighting-compensated image
(bottom) overlaid with face detection results ...............

Face detection results in the presence of eye glasses. Each example contains
an original image (top) and a lighting-compensated image (bottom)
overlaid with face detection results. ...................

Face detection results for subjects with facial hair. Each example contains
an original image (top) and a lighting-compensated image (bottom)
overlaid with face detection results. ...................

Face detection results on half-profile faces. Each example contains an orig-
inal image (top) and a lighting-compensated image (bottom) overlaid
with face detection results .........................

Face detection results on a subset of the HHI database: (a) input images;
(b) grouped skin regions; (c) face candidates; ((1) detected faces are
overlaid on the lighting—compensated images ...............

Face detection results on a subset of the Champion database: (a) input
images; (b) grouped skin regions; (c) face candidates; (d) detected faces
are overlaid on the lighting-compensated images .............

Face detection results on a subset of eleven family photos. Each image
contains multiple human faces. The detected faces are overlaid on the
color-compensated images. False negatives are due to extreme lighting
conditions and shadows. Notice the difference between the input and
color-compensated images in terms of color balance. The bias color in
the original images has been compensated in the resultant images. . .

(Cont’d). ...................................

(Cont’d). ...................................

Face detection results on a subset of 24 news photos. The detected faces
are overlaid on the color-compensated images. False negatives are due
to extreme lighting conditions, shadows, and low image quality (i.e.,
high compression rate) ...........................

Graphical user interface (GUI) for face editing. (a) detection mode; (b)
editing mode. ...............................

The system overview of the proposed modeling method based on a 3D
generic face model. ............................

xiv

81

81

83

84

84

85

86

89

90

91
92
93

94

96

98

4.2

4.3

4.4

4.5

4.6

4.7
4.8

4.9

4.10

4.11

4.12

5.1

5.2

5.3

5.4

3D triangular-mesh model and its feature components: (a) the frontal
view; (b) a side view; (c) feature components. .............

Phong—shaded 3D model shown at three viewpoints. Illumination is in
front of the face model ...........................

Facial measurements of a human face: (a) color image; (b) range map; and
the range map with texture mapped for (c) a left view; ((1) a profile
view; (e) a right view. ..........................

Facial features overlaid on the color image, (a) obtained from face detec-
tion; (b) generated for face modeling. ..................

Global alignment of the generic model (in red) to the facial measurements
(in blue): the target mesh is plotted in (a) for a hidden line removal
mode for a side view; (b) for a see-through mode for a profile View.

Displacement propagation ...........................

Local feature alignment and displacement propagation Shown for the
frontal view: (a) the input generic model; the model adapted to (b)
the left eye; (c) the nose; ((1) mouth and chin. .............

Local feature refinement: initial (in blue) and refined (in red) contours
overlaid on the energy maps for (a) the face boundary; (b) the nose;
(c) the left eye; and (d) the mouth. ...................

The adapted model (in red) overlapping the target measurements (in blue),
plotted (a) in 3D; (b) with colored facets at a profile view. ......

Texture Mapping. (a) The texture—mapped input range image. The
texture-mapped adapted mesh model shown for (b) a frontal view;
((1) a left view; (e) a profile view; (f) a right view. ...........

Face matching: the top row shows the 15 training images generated from
the 3D model; the bottom row shows 10 test images of the subject
captured from a CCD camera ......................

Semantic face graph is shown in a frontal view, whose nodes are (a) in-
dicated by text; (b) depicted by polynomial curves; (c) filled with
different shades. The edges of the semantic graph are implicitly stored
in a 3D generic face model and are hidden here. ............

3D generic face model: (a) Waters’ triangular-mesh model shown in the
side view; (b) model in (a) overlaid with facial curves including hair
and ears at a side view; (c) model in (b) shown in the frontal view.

Semantic face graphs for the frontal view are reconstructed using Fourier
descriptors with spatial frequency coefficients increasing from (a) 10%
to (j) 100% at increments of 10%. ....................

Face detection results: (a) and (c) are input face images of size 640 x 480
from the MPEG7 content set; (b) and (d) are detected faces, each of
which is described by an oval and a triangle. ..............

XV

100

100

102

102

103
104

105

107

108

109

110

113

114

115

110

5.5

5.6

5.7

5.8

5.9

5.10

5.11

5.12

5.13

Boundary map and eye component map for coarse alignment: (a) and (b)
are gradient magnitude and orientation, respectively, obtained from
multi-scale Gaussian-blurred edge response; (c) an eye map extracted
from a face image shown in Fig. 5.4(c); (d) a semantic face graph
overlaid on a 3D plot of the eye map; (e) image overlaid with a coarsely
aligned face graph. ............................

Shadow maps: (a) and (c) are luma components Of face images in Figs.
5.4(a) and 5.4(c), overlaid with rectangles within which the average
values of skin intensity is calculated; (b) and (d) are shadow maps
where bright pixels indicate the regions that are darker than average
skin intensity. ...............................

Coarse alignment: (a) input face images of size 640 x 480 from the MPEG7
content set (first three rows), and of size 256 X 384 from the MSU
database (the fourth row); (b) detected faces; (c) locations of eyebrow,
nostril, and mouth lines using shadow maps; (d) face images overlaid
with coarsely aligned face graphs. ....................

Interacting snakes: (a) face region extracted from a face image shown in
Fig. 5.4(a); (b) image in (a) overlaid with a (projected) semantic face
graph; (c) the initial configuration of interacting snakes obtained from
the semantic face graph shown in (b) ...................

Repulsion force: (a) interacting snakes with index numbers marked; (b)
the repulsion force computed for the hair outline; (c) the repulsion
force computed for the face outline. ...................

Gradient vector field: (a) face region of interest extracted from a 640x480
image; (b) thresholded gradient map based on the population of edge
pixels shown as dark pixels; (c) gradient vector field. .........

Component energy (darker pixels have stronger energy): (a) face region of
interest; (b) eye component energy; (c) mouth component energy; (d)
nose boundary energy; (e) nose boundary energy shown as a 3D mesh
surface. ..................................

Fine alignment: (a) snake deformation shown every five iterations; (b)
aligned snakes (currently six snakes~hairstyle, face-border, eyes, and
mouth—are interacting); (c) gradient vector field overlaid with the
aligned snakes. ..............................

Fine alignment with evolution steps: (a) a face image; (b) the face in
(a) overlaid with a coarsely aligned face graph; (c) initial interacting
snakes with different Shades in facial components (cartoon face); (d)
curve evolution shown every five iterations (totally 55 iterations); (e)
an aligned cartoon face. .........................

xvi

119

120

121

122

124

125

126

128

5.14

5.15
5.16
5.17
5.18

5.19

5.20

5.21

5.22

5.23

6.1

Fine alignment using geodesic active contours: (a) a generic cartoon face
constructed from interacting snakes; (b) to (f) for five different sub—
jects. For each subject, the image in the first row is the captured face
image; the second row shows semantic face graphs obtained after coarse
alignment, and overlaid on the color image; the third row shows seman—
tic face graphs with individual components shown in different shades
of gray; the last row shows face graphs with individual components
after fine alignment. ...........................

A semantic face matching algorithm. ....................
Five color images (256 X 384) of a subject. .................
Face images of ten subjects. .........................

Examples of misclassification: (a) input test image; (b) semantic face
, graph of the image in (a); (c) face graph of the misclassified subject;
(d) face graph of the genuine subject obtained from the other images

of the subject in the database (i.e., without the input test image in
(a)). Each row shows one example of misclassification. ........

Cartoon faces reconstructed from Fourier descriptors using all the fre-
quency components: (a) to (j) are ten average cartoon faces for ten
different subjects based on five images for each subject. Individual
components are shown in different shades in (a) to (e). ........

Cartoon faces reconstructed from Fourier descriptors using only 50% of
the frequency components: (a) to (j) are ten average cartoon faces for
ten different subjects based on five images for each subject. Individual
components are shown in different Shades in (a) to (e). ........

Cartoon faces reconstructed from Fourier descriptors using only 30% of
the frequency components: (a) to (j) are ten average cartoon faces for
ten different subjects based on five images for each subject. Individual
components are shown in different shades in (a) to (e). ........

Facial caricatures generated based on a generic 3D face model: (a) a pro-
totype of the semantic face graph, G0, obtained from a generic 3D face
model, with individual components shaded; (b) face images of six dif~
ferent subjects; (c)-(g) caricatures of faces in (b) (semantic face graphs
with individual components shown in different shades) with different
values of exaggeration coefficients, k, ranging from 0.1 to 0.9 ......

Facial caricatures generated based on the average face of 50 faces (5 for
each subject):(a) a prototype of the semantic face graph, G0, ob—
tained from the mean face of the database, with individual components
shaded; (b) face images of Six different subjects; (c)-(g) caricatures of
faces in (b) (semantic face graphs with individual components shown
in different shades) with different values of exaggeration coefficients, 1:,
ranging from 0.1 to 0.9 ...........................

A prototype of a face identification system with the tracking function. . .

xvii

131
134
135
135

137

138

139

139

141

142

148

6.2 An example of motion detection in a video frame: (a) A color video frame;
(b) extracted regions with Significant motion; (c) detected moving skin
patches shown in pseudocolor; (d) extracted face candidates described
by rectangles. ............................... 149
6.3 Face tracking results on a sequence of 25 video frames. These images are
arranged from top to bottom and from left to right. Detected faces are

overlaid on the lighting-compensated images ............... 150
Al Color spaces: (a) RGB; (b) YCbCr. .................... 154
CI Architecture of IPTL class templates ..................... 162

xviii

Chapter 1

Introduction

In recent years face recognition has received substantial attention from researchers
in biometrics, pattern recognition, and computer vision communities (see surveys in
[36], [37], [38]). This common interest among researchers working in diverse fields is
motivated by our remarkable ability to recognize people (although in case of certain
rare brain disability, e. g., prosopagnosia or face blindness [39], this recognition ability
is lost) and the fact that human activity is a primary concern both in everyday
life and in cyberspace. Besides, there are a large number of commercial, security,
and forensic applications requiring the use of face recognition technology. These
applications (see Fig. 1.1) include automated video surveillance (e. g., super bowl face
scans and airport security checkpoints), access control (e.g., to personal computers
and private buildings), mugshot identification (e. g., for issuing driver licenses), design
of human computer interface (HCI) (e.g., classifying the activity of a vehicle driver),
multimedia communication (e.g., generation of synthetic faces), and content-based

image database management [40]. These applications involve locating, tracking, and

recognizing a Single (or multiple) human subject(s) or face(s).

Face recognition is an important biometric identification technology. Facial scan is
an effective biometric attribute / indicator. Different biometric indicators are suited for
different kinds of identification applications due to their variations in intrusiveness,
accuracy, cost, and (sensing) effort [5] (see Fig. 1.2(a)). Among the six biometric
indicators considered in [6], facial features scored the highest compatibility, shown
in Fig. 1.2(b), in a machine readable travel documents (MRTD) system based on a
number of evaluation factors, such as enrollment, renewal, machine requirements, and

public perception [6].

1.1 Challenges in Face Recognition

Humans can easily recognize a known face in various conditions and representations
(see Fig. 1.3). Such a remarkable ability of humans to recognize faces with large
intra—subject variations has inspired vision researchers to develop automated systems
for face recognition based on 2D face images. However, the current state~of-the-art
machine vision systems can recognize faces only in a constrained environment. Note
that there are two types of face comparison scenarios, called (i) face verification (or
authentication) and (ii) face identification (or recognition). As shown in Fig. 1.4, face
verification involves a one-to-one match that compares a query face image against
a template face image whose identity is being claimed, while face identification in-
volves one-to—many matches that compare a query face image against all the template
images in a face database to determine the identity of the query face. The main chal-

2

g .
i
:1
3
:1
1
)—
x
7:1
:x
x
.x
:1
.J
g,
1
1
:1
:1
,x
—x
1
x
:1
:11
.1
.I
1
j
1

”E2 ; vrs_Iomcs;;

 

Dodgy arr ]

'i,

_
j
'i
I
.1
l
I

 

    

According Control

Mace Input Selection

3.1-.1731] 3334 33333

 

(b)
Figure 1.1. Applications using face recognition technology: (a) and (b) automated
video surveillance (downloaded from Visionics [1] and FaceSnap [2], respectively);
(c) and (d) access control (from Visionics [I] and from Viisage [3] respectively); (e)
management of photo databases (from Viisage [3]); (f) multimedia communication
(from Eyematic [4]).

(d)
Figurc1.1. (Cont’d).

[.
ll lam-.1.

.
.
a

 

 

 

 

 

 

Figure 1.1.
Zeohyr" Analysis
@
Ko troke-
Sex: Hand-Scan
Signature-

Faclal-
Scan Scan

Retina- Finger-

can Scan
Iris-Scan VoiceScan
o W W lbs-tic em

 

] uletm-Ammy 0M oEllon
(a)

 

 

 

.5
c
|

Weighted percentage

N
o
l

 

 

 

Face Finger Iland Voice En Signature

(1))

Figure 1.2. Comparison of various biometric features: (a) based on zephyr analysis
(downloaded from [5]); (b) based on MRTD compatibility (from [6]).

.l:

g
El.
5.
a.
a
E'

  

Figure 1.3. Intra—subject variations in pose, illumination, expression, occlusion,
accessories (e.g., glasses), color, and brightness.

 

       

are

One-to-one ,1 One-to:many
Query Template Query Template
(8) (b)
Figure 1.4. Face comparison: (a) face verification/ authentication; (b) face identifi-
cation/recognition. Face images are taken from the MSU face database [7].

 

 

 

 

 

 

lenge in vision-based face recognition is the presence of a high degree of variability in
human face images. There can be potentially very large intra—subject variations (due
to 3D head pose, lighting, facial expression, facial hair, and aging [41]) and rather
small inter-subject variations (due to the similarity of individual appearances). Cur-
rently available vision-based recognition techniques can be mainly categorized into
two groups based on the face representation which they use: (i) appearance~based
which use holistic texture features, and (ii) geometry-based which use geometrical
features of the face. Experimental results Show that appearance-based methods gen-
erally perform a better recognition task than those based on geometry, because it
is difficult to robustly extract geometrical features especially in face images of low

6

resolutions and of poor quality (i.e., to extract features under uncertainty). However,
the appearance-based recognition techniques have their own limitations in recognizing
human faces in images with wide variations in 3D head pose and in illumination [38].
Hence, in order to overcome variations in pose, a large number of face recognition
techniques have been developed to take into account the 3D face shape, extracted
either from a video sequence or range data. As for overcoming the variations in
illumination, several studies have explored features such as edge maps (e.g., eigen-
hills and eigenedges in [42]), intensity derivatives, Gabor-filter responses [43], and
the orientation fields of intensity gradient [44]. However, none of these approaches
by themselves lead to satisfactory recognition results. Hence, the explicit 3D face
model combined with its reflectance model is believed to be the best representation

of human faces for the appearance-based approach [43].

1.2 Semantic Facial Components

Face recognition technology provides useful tools for content-based image and video
retrieval based on a semantic (high-level) concept, i.e., human faces. Is all face pro-
cessing holistic [45]? Some approaches, including feature—based and appearance-based
[46] methods, emphasize that internal facial features (i.e., pure face regions) play the
most important role in face recognition. On the other hand, some appearance-based
methods suggest that in some situations face recognition is better interpreted as
head recognition [8], [31]. An example supporting the above argument was demon-
strated for Clinton and Gore heads [8] (See Fig. 1.5(a)). While the two faces in

7

 

 

 

 

 

 

 

Democrat coalition. ' ,._

(a) (b)
Figure 1.5. Head recognition versus face recognition: (a) Clinton and Gore heads
with the same internal facial features, adapted from [8]; (b) two faces of different
subjects with the same internal facial components show the important role of hair
and face outlines in human face recognition.

 

Fig. 1.5(a) have identical internal features, we can still distinguish Clinton from Gore.
We notice that in this “example” the hair style and the face outline are significantly
different. We reproduce this scenario, across genders, in Fig. 1.5(b). Humans will
usually identify these two persons with different identities. This prompted Liu et
a1. [47] to emphasize that there is no use of face masks (to remove the “non-pure-
face” portion) in their appearance—based method. As a result, we believe that the
separation of external and internal facial features/ components is helpful in assigning

weights on external and internal facial features in the face recognition process.

Modeling facial components at the semantic level (i.e., eyebrows, eyes, nose,
mouth, face outline, ears, and the hair outline) helps to separate external and in-
ternal facial components, and to understand how these individual components con-
tribute to face recognition. Examples of modeling facial components can be found
in the faces represented in caricatures and cartoons. However, the fact that humans

8

can recognize known faces in caricature drawings (e.g., faces shown in Fig. 1.6) and
cartoons (see Fig.1.?) without any difficulty has not been fully explored in research

studies on face recognition [48], [49], [50], [51]. Note that some of the faces shown

       

(a) (9)
Figure 1.6. Caricatures of (a) Vincent Van Gogh; (b) Jim Carrey; (c) Arnold
Schwarzenegger; (d) Einstein; (e) G. W. Bush; and (f) Bill Gates. Images are down-
loaded from [9], [10] and [10]. Caricatures reveal the use of component weights in face
identification.

   

(a) (b) (C) (d)

Figure 1.7. Cartoons reveal that humans can easily recognize characters whose facial
components are depicted by simple line strokes and color characteristics: (a) and (b)
are frames adapted from the movie Pocahontas; (c) and (d) are frames extracted from
the movie Little Mermaid II. (Disney Enterprises, Inc.)

in Fig. 1.6 are represented only by strokes (geometrical features), while some others
have parts of facial features dramatically emphasized with some distortion. Cartoon
faces are depicted by line drawings and color without shading. People can easily
identify faces in caricatures (see, Fig. 1.6) that exaggerate some of the facial compo-
nents/landmarks. Besides, we can also easily identify known faces merely based on

some salient facial components. For example, we can quickly recognize a known face

9

with a distinctive chin no matter whether the face appears in a caricature (e.g., Jim
Carrey shown in Fig. 1.6(b)) or in a real photo [52]. Caricatures reveal that there
are certain facial features which are salient for each individual and that a relatively
easier identification of faces can occur by emphasizing distinctive facial components
(using weights) and their configuration. Besides, the spatial configuration of facial
components has been shown to take a more important role in face recognition than
local texture by using inverted faces [53] in which the (upright) face recognition is

disrupted (see Fig. 1.8). Therefore, we group these salient facial components [48] as

 

(a) (b) (C)

Figure 1.8. Configuration of facial components: (a) face image; (b) face image in (a)
with enlarged eyebrow—to—eye and nose-to—mouth distances; (c) inverted face of the
image in (b). A small change of component configuration results in a significantly
different facial appearance in an upright face in (b); however, this change may not be
perceived in an inverted face in (c).

a graph and derive component weights in our face matching algorithm to improve the
recognition performance.

In addition, humans can recognize faces in the presence of occlusions, i.e., face
recognition can be based on a (selected) subset of facial components. This explains

10

the motivation for studies that attempt to recognize faces from eyes only [54]. The
use of component weights can facilitate face recognition based on selected facial com-
ponents. Furthermore, the shape of facial components (see Fig. 1.9(a)) has been
used in physiognomy (or face reading, an ancient art of deciphering a person’s past
and personality from his/her face). In light of this art, we design a semantic face
graph for face recognition (see in Chapter 5), shown in Fig. 1.9(b), in which ten facial

components are filled with different shades in a frontal view.

 

 

 

 

 

 

 

(a)
Figure 1.9. Facial features/components: (a) five kinds of facial features (i.e., eye-
brows, eyes, nose, ears, and mouth) in a face for reading faces in physiognomy (down-
loaded from [11]); (b) a frontal semantic face graph, whose nodes are facial compo-
nents that are filled with different shades.

For each facial component, the issue of representation also plays an important
role in face recognition. It has been believed that local facial texture and shading are
crucial for recognition [52]. However, some frames of a cartoon video, as shown in Fig.

11

1.7, reveal that line drawings and color characteristics (shades) of facial components
(e.g., dark colors for eyebrows and both bright and dark colors for eyes) provide
sufficient information for humans to recognize the faces of characters in cartoons.
People can even recognize cartoon faces without the use of shading information, which
is rather unstable under different lighting conditions. Consequently, we believe that
curves (or sketches) and shades of facial‘components provide a promising solution
to the representation of facial components for recognition. However, very little work
has been done in face recognition based on facial sketches [55], [56] and (computer-

generated [57]) caricatures [58], [48], [50].

In summary, external and internal facial components, and distinctiveness, config-
uration and local texture of facial components all contribute to the process of face
recognition. Humans can seamlessly blend and independently perform appearance-
based and geometry-based recognition approaches efficiently. Therefore, we believe
that merging [59], [60] the holistic texture features and the geometrical features (es-
pecially at a semantic level) is a promising method to represent faces for recognition.
While we focus on the 3D variations in faces, we should also take the temporal (aging)
factor into consideration while designing face recognition systems [41]. In addition to
large intra—subject variations, another difficulty in recognizing faces lies in the small
inter-subject variations (Shown in Fig. 1.10). Different persons may have very similar
appearances. Identifying people with very similar appearances remains a challenging
task in automatic face recognition.

12

 

(a) (b)
Figure 1.10. Similarity of frontal faces between (a) twins (downloaded from [12]);
and (b) a father and his son (downloaded from [13]).

1.3 Face Recognition Systems

Face recognition applications in fact involve several important steps, such as face
detection for locating human faces, face tracking for following moving subjects, face
modeling for representing human faces, face coding / compression for efficiently archiv-
ing and transmitting faces, and face matching for comparing represented faces and
identifying a query subject. Face detection is usually an important first step. De-
tecting faces can be viewed as a two-class (face vs. non-face) classification problem,
while recognizing faces can be regarded as a multiple—class (multiple subjects) classi-
fication problem within the face class. Face detection involves certain aspects of face
recognition mechanism, while face recognition employs the results of face detection.
We can consider face detection and recognition as the first and the second stages in
a sequential classification system. The crucial issue here is to determine an appro-
priate feature space to represent a human face in such a classification system. We
believe that a seamless combination of face detection, face modeling, and recogni-
tion algorithms has the potential of achieving high performance for face identification

13

applications.

With this principle, we propose two automated recognition paradigms, shown in
Fig. 1.11 and Fig. 1.12, that can combine face detection as well as tracking (not
included in this thesis, but can be realized based on our current work), modeling, and
recognition. The first paradigm requires both video sequences and 2.5D/ 3D facial
measurements as its input in the learning/ enrollment stage. In the recognition/test
stage, however, face images are extracted from video input only. Faces are identified
based on an appearance-based algorithm. The second paradigm requires only video
sequences as its input in both learning and recognition stages. Its face recognition
module makes use of a semantic face matching algorithm to compare faces based on
weighted facial components.

Both paradigms contain three major modules: (i) face detection and feature ex-
traction, (ii) face modeling, and (iii) face recognition. The face detection/location
and feature extraction module is able to locate faces in video sequences. The most
important portion of this module is a feature extraction sub—module that extracts
geometrical features (such as face boundary, eyes, eyebrows, nose, and mouth), and
texture/ color features (estimation of the head pose and illumination is left as a future
research direction). The face modeling module employs these extracted features for
modifying the generic 3D face model in the learning and recognition stages. In this
thesis, we describe the implementation of the face modeling module in both proposed
paradigms for the frontal view only. The extension of the face modeling module to
non-frontal views can be a future research direction. The recognition module makes
use of facial features extracted from an input image and the learned 3D models to

14

verify the face present in an image in the recognition stage. This thesis has developed
a robust face detection module which is used to facilitate applications such as face
tracking for surveillance, and face modeling for identification (as well as verification).
We will briefly discuss the topics of face detection and recognition, face modeling as

well as compression, and face—based image retrieval in the following sections.

1.4 Face Detection and Recognition

Human activity is a major concern in a wide variety of applications such as video
surveillance, human computer interface, face recognition [37], [36], [38], and face
image database management [40]. Detecting faces is a crucial step and usually the
first one in these identification applications. However, due to various head poses,
illumination conditions, occlusion, and distances between teh sensor and the subject
(which may result in a blurred face), detecting human faces is an extremely difficult
task under unconstrained environments (see images in Figs. 1.13 (a) and (b)). Most
face recognition algorithms assume that the problem of face detection has been solved,
that is, the face location is known. Similarly, face tracking algorithms (e.g., [61])
often assume the initial face location is known. Since face detection can be viewed
as a two—class (face vs. non-face) classification problem, some techniques developed
for face recognition (e.g., holistic/template approaches [21], [62], [63], [64], feature-
based approaches [65], and their combination [66]) have been used to detect faces.
However, these detection techniques are computationally very demanding and cannot
handle large variations in faces. In addition to the face location, a face detection

15

Swen: .830 can awash woacumwwm: wEm: Scum? nosEchE can“ womefiamwwca Om .30 mo 8.9%ch 889mm .2; magma

   

 

5.5508:
8mm

 

 

 

 

.. . .3 .4 .. _
a...“ 32%.; y

...:.._3..,,...

=o.=o§xm 258m
a 5:033 8mm

 

 

 

 

 

 

 

 

 

16

dump owns: mo 8: 2: 5055 839mm coSEmoOg 83 8.83-chpo QM So we Seawese Ecuam .NHA 8:th

  
  

 
 
 

8a". m 25. 8a". 85228
=8 8 fine”... swank.“ 8.8.3 8r. one; :5 ass is
r at “UN. 3.. _, .t .
0 .0 a. . ,m
, y ,2.

 

3...??an

    

5:52

 

mmmem.
22:20.88;
i I
macaw
mEEmmq

coaogxm Sega
a 3:088 83

 

 

 

 

 

 

 

 

17

a
‘-

,. *.-_. ‘
,—,,‘.::--».«-,

. a.“ -_
. “3..-kn ~.

 

Figure 1.13. Face images taken under unconstrained environments: (a) a crowd of
people (downloaded from [14]); (b) a photo taken at a swimming pool.

18

algorithm can also provide geometrical facial features for face recognition. Merging
the geometrical features and holistic texture (appearance—based) features is believed
to be a promising method of representing faces for recognition [59], [60]. Therefore,
we believe that a seamless combination of face detection and recognition algorithms
has the potential of providing a high performance face identification algorithm.
Hence, we have proposed a face detection algorithm for color images, which is
able to generate geometrical as well as texture features for recognition. Our approach
is based on modeling skin color and extracting geometrical facial features. The skin
color is detected by using a lighting compensation technique and a nonlinear color
transformation. The geometrical facial features are extracted from eye, mouth, and
face boundary maps. The detected faces, including the extracted facial features, are
organized as a graph for modeling and recognition processes. Our algorithm can detect
faces under different head poses, illuminations, and expressions (see Fig. 1.14(a)), and
family photos (see Fig. 1.14(b)). However, our detection algorithm is not designed
for detecting faces in gray-scale images, cropped face images (see Fig. 1.15(a)) and

faces wearing make—up or mask (see Figs. 1.15(b) and (0)).

1.5 Face Modeling for Recognition

Our face recognition systems are based on 3D face models. 3D models of human faces
have been widely used to facilitate applications such as video compression/ coding,
human face tracking, facial animation, augmented reality, recognition of facial ex-
pression, and face recognition. Figure 1.16 shows two graphical user interfaces of a

19

 

Figure 1.14. Face images for our detection algorithm: (a) a montage image containing
images adapted from MPEG7 content set [15]; (b) a family photo.

 

Figure 1.15. Face images not suitable for our detection algorithm: (a) cropped
image (downloaded from [16]); (b) a performer wearing makeup (from [14]); (0)
people wearing face masks (from [14]).

20

commercial parametric face modeling system [17], FaceGen Modeller, which is based
on face shape statistics. It can efficiently create a character with specified age, gen-
der, race, and caricature morphing. Current trend in face recognition is to employ
3D face model explicitly [67], because such a model provides a potential solution
to identifying faces with variations in illumination, 3D head pose, and facial expres-
sion. These variations, called the intra—subject variations, also include changes due
to aging, facial hair, cosmetics, and facial accessories. These intra—subject variations
constitute the primary challenges in the field of face recognition. As object-centered
representations of human faces, 3D face models not only can augment recognition
systems that utilize viewer-centered face representations (based on 2D face images),
but also can blend together holistic approaches and geometry-based approaches for
recognition. However, the three state-of-the—art face recognition algorithms [68], (1)
the principal component analysis (PCA)-based algorithm; (2) the local feature anal-
ysis (LFA)-based algorithm; and (3) the dynamic-link-architecture—based algorithm,
use only viewer-centered representations of human faces. A 3D model—based match-
ing algorithm is likely to provide a potential solution for advancing face recognition
technology. However, for face recognition, it is more important to capture facial
distinctiveness of recognition-oriented components than to generate a realistic face
model. We briefly introduce our face modeling methods for recognition (i.e., face

alignment) and model compression in the following subsections.

21

 

Generate [Shape] Texture] Tween] Animate] Photo ]
All Races ]Nncan] European] SE Asran] E Indran]

All Races Controls

 

 

 

 

 

Step 1 Oplronal
Generate l Make a random (are Set Mragei Reset to Iverage lace
rep 2
'5' .ms :ep morph_ 'T' . Tenur morph
Use ‘Sync Lock' to synchronize movement OHM 2 sMers
Gender Age Carlcature Asymmetry
S l T S I T S / T S I T
] very male The average T Symmetric
] ] zu [ i
30 l Typlcal ] Tywul
Mlle [ 1
40 [ Cancature |
Female 1 '
5° 3 "ouster ]
‘ 1 Warner:
so ] Q
V 0 1 s Very Tamale 4
lmng p m P" Syn Lock 9 c Lock l" Lock l" Lock
Shading Mode Movement C M 3"" 8W
7 Smoolh r‘ Flat 5 Rolalton r Tnnalalron
Race Morphing
rTenureModt—"a -- ~ 7 - ,
r: 0,. r , ] ] .
7 ~ 7 , T NrRacea All Races NlRacns ’ All Race:
Video Gamma Lighting Option ]
f————- T‘_'——_l
. . . . . . . . . . . [FrontilLrghtmg Ll ‘
1-0 1 0 Miran Europe-n SE Asian East lnalen
[— Wlmame [— Shiny Display Background Colour
Rrght.’ rack and drag on the face to Still: rt Lefi clrclr and drag to .7 . ,, , 1.,
unit rem facet um»

 

 

 

 

 

 

 

    

furl-(um .‘Jtnlullrrr [.IJ [JIMU rMurlt‘l: llrtmult Moduli
Fle Ed

 

Generate] Shape] Tenure TWO" [Animate] Photo ]

Target Face:

 

Load Target

 

 

Symmetry Asymmetry
S / T S / T
r [ Lees similar ] l
] a
T T Comm .1 ]r
] ] in between I ]
Vrmng Optrons [ r g l
Shldrng Mode Movement ; [ 73’9“ [
’; Smooth (‘ Flat I" Rotation r Translation ] , Exaggerate ]
Texture Mode , » 7'7 Sync Loclr F Sync Lock
P On "'
Vrdeo Gamma Lrghtmg Oprron
* F—-—‘ ~_.

 

1 0 2 o 7
V WIMrame F Shmy DlSplZy Background Colour I

Rrg‘ht clrck and drag on the {ace to scale rt Len tlrcka and drag to
more rl There eare (1340 late and 9858‘"?

 

 

 

 

 

(b)
Figure 1.16. Graphical user interfaces of the FaceGen modeller [17]. A 3D face model
shown (a) with texture mapping; (b) with wireframe overlaid.

22

1.5.1 Face Alignment Using 2.5D Snakes

In our first recognition system (shown in Fig. 1.11), we have proposed a face modeling
method which adapts an existing generic face model (a priori knowledge of a human
face) to an individual’s facial measurements (i.e., range and color data). We use
the face model that was created for facial animation by Waters [69] as our generic
face model. Waters’ model includes details of facial features that are crucial for face
recognition. Our modeling process aligns the generic model onto extracted facial
features (regions), such as eyes, mouth, and face boundary, in a global—to—local way,
so that facial components that are crucial for recognition are fitted to the individual’s
facial geometry. Our global alignment is based on the detected locations of facial
components, while the local alignment utilizes two new techniques which we have
developed, displacement propagation and 2.5D active contours, to refine local facial
components and to smoothen the face model. Our goal of face modeling is to generate
a learned 3D model of an individual for verifying the presence of the individual in a
face database or in a video. The identification process involves (i) the modification
of the learned 3D model based on different head poses and illumination conditions
and (ii) the matching between 2D projections of the modified 3D model, whose facial

shape is integrated with facial texture, and sensed 2D facial appearance.

1.5.2 Model Compression

Requirements of easy manipulation, progressive transmission, effective visualization
and economical storage for 3D (face) models have resulted in the need for model

23

compression. The complexity of an object model depends not only on object geom-
etry but also on the choice of its representation. The 3D object models explored in
computer vision and graphics research have gradually evolved from simple polyhedra,
generated in mechanical Computer Aided Design (CAD) systems, to complex free-
form objects, such as human faces captured from laser scanning systems. Although
human faces have a complex shape, modeling them is useful for emerging applica-
tions such as virtual museums and multimedia guidebooks for education [70], [71],
low-bandwidth transmission of human face images for teleconferencing and interactive
TV systems [72], virtual people used in entertainment [73], sale of facial accessories

in e—commerce, remote medical diagnosis, and robotics and automation [74].

The major reason for us to adopt the triangular mesh as our generic human
face model is that it is suitable for describing and simplifying the complexity of facial
geometry. In addition, there are a number of geometry compression methods available
for compressing triangular meshes (e. g., the topological surgery [75] and the multi-
resolution mesh simplification [76]). Beside these helps, we can further obtain a
more compact representation of a 3D face model by carefully selecting vertices of the
triangular mesh for representing facial features that are extracted for face recognition.
Our proposed semantic face graph used in the semantic recognition paradigm (see
Fig. 1.12) is such an example.

24

1.5.3 Face Alignment Using Interacting Snakes

For the semantic recognition system (shown in Fig. 1.12), we define a semantic face
graph. A semantic face graph is derived from a generic 3D face model for identifying
faces at the semantic level. The nodes of a semantic graph represent high—level facial
components (e.g., eyes and mouth), whose boundaries are described by open (or
closed) active contours (or snakes). In our recognition system, face alignment plays
a crucial role in adapting a priori knowledge of facial topology, encoded in semantic
face graph, onto the sensed facial measurements (e.g., face images). The semantic
face graph is first projected onto a 2D image, coarsely aligned to the output of the
face detection module, and then finely adapted to the face images using interacting

snakes.

Snakes are useful models for extracting the shape of deformable objects [77].
Hence, we model the component boundaries of a 2D semantic face graph as a collec-
tion of snakes. We propose an approach for manipulating multiple snakes iteratively,
called interacting snakes, that minimizes the attraction energy functionals on both
contours and enclosed regions of individual snakes and the repulsion energy function-
als among multiple snakes that interact with each other. We evaluate the interacting
snakes through two types of implementations, explicit (parametric active contours)

and implicit (geodesic active contours) curve representations, for face alignment.

Once the semantic face graph has been aligned to face images, we can derive
component weights based on distinctiveness and visibility of individual components.
The aligned face graph can also be easily used to generate cartoon faces and facial

25

caricatures by exaggerating the distinctiveness of facial components. After alignment,
facial components are transformed to a feature space spanned by Fourier descriptors of
facial components for face recognition, called semantic face matching. The matching
algorithm computes the similarity between semantic face graphs of face templates in
a database and a semantic face graph that is adapted to a given face image. The
semantic face graph allows face matching based on selected facial components, and
effective 3D model updating based on 2D face images. The results of our face matching
demonstrate that the proposed face model can lead to classification and visualization
(e. g., the generation of cartoon faces and facial caricatures) of human faces using the

derived semantic face graphs.

1.6 Face Retrieval

Today, people can accumulate a large number of images and video clips (digital con-
tent) because of the growing popularity of digital imaging devices, and because of
the decreasing cost of high-capacity digital storage. This significant increase in the
amount of digital content requires database management tools that allow people to
easily archive and retrieve contents from their digital collections. Since humans and
their activities are typically the subject of interest in consumers’ images and videos,
detecting people and identifying them will help to automate image and video archival
based on a high-level semantic concept, i.e., human faces. For example, we can design
a system that manages digital content of personal photos and amateur videos based
on the concept of human faces, e.g., “retrieve all images containing Carrie’s face.”

26

Using merely low-level features (e. g., skin color or color histograms) for retrieval and
browsing is neither robust nor acceptable to the user. High level semantics have to
be used to make such an image/ video management system useful. Fig. 1.17 shows a

graphical user interface of a facial feature-based retrieval system [18].

 

Figure 1.17. A face retrieval interface of the FACEit system [18]: the system gives
the most similar face in a database given a query face image.

In summary, the ability to group low-level features as a meaningful semantic entity
is a critical issue in the retrieval of visual content. Accurately and efficiently detecting
human faces plays a crucial role in facilitating face identification for managing face
databases. In face recognition algorithms, the high-level concept—a human face—is
implicitly expressed by face representations such as locations of feature points, surface
texture, 2D graphs with feature nodes, 3D head surface, and combinations of them.
The face representation plays an important role in the recognition process because

27

different representations lead to different matching algorithms. We can design a
database management system that utilizes the outputs of our face detection and
modeling modules as indices to search a database based on the semantic concepts,
such as “find all the images containing John’s faces” and “search faces which have

Vincent’s eyes (or face shape).

1.7 Outline of Dissertation

This dissertation is organized as follows. Chapter 2 presents a brief literature review
on face detection and recognition, face modeling (including model compression), and
face retrieval. In Chapter 3, we present our face detection algorithm for color im-
ages. Chapter 4 discusses our range data-based face modeling method for recognition.
Chapter 5 describes the semantic face recognition system, including face alignment
using interacting snakes, a semantic face matching algorithm, and the generation
of cartoon faces and facial caricatures. Chapter 6 presents conclusions and future

directions related to this work.

1.8 Dissertation Contributions

The major contributions of this dissertation are categorized into the topics of face
detection, face modeling, and face recognition. In face detection, we have developed
a new face detection algorithm for multiple non—profile-view faces with complex back—
ground in color images, based on localization of skin-tone color and facial features

28

such as eyes, mouth and face boundary. The main properties of this algorithms are

listed as follows.

0 Lighting compensation: This method corrects the color bias and recovers
the skin-tone color by automatically estimating the reference white pixels in a
color image, under the assumption that an image usually contains “real white”
(i.e., white reference) pixels and the dominant bias color in an image always

appears as “real white”.

0 Non-linear color transformation: In literature, the chrominance compo—
nents of the skin tone have been assumed to be independent of the luminance
component of the skin tone. We found that the chroma of skin tone depends on
the luma. We overcome the difficulty of detecting the low-luma and high-luma
skin tone colors by applying a nonlinear transform to the YCbCr color space.
The transformation is based on the linearly fitted boundaries of our training

skin cluster in YCb and YC, color subspaces.

0 Modeling a skin-tone color classifier as an elliptical region: A simple
classifier which constructs an elliptical decision region in the chroma subspace,
050,, has been designed, under the assumption of the Gaussian distribution of

skin tone color.

0 Construction of facial feature maps for eyes, mouth, and face bound-
ary: With the use of gray-scale morphological operators (dilation and erosion),
we construct these feature maps by integrating the luminance and chrominance
information of facial features. For example, eye regions have high Cb (difference

29

between blue and green colors) and low C. (difference between red and green
colors) values in chrominance components, and have brighter and darker values

in the luminance component.

0 Construction of a diverse database of color images for face detection:
The database includes a MPEG7 content set, mug-shot style web photos, family

photos, and news photos.

 

In face modeling, we have designed two methods for aligning a 3D generic face
model onto facial measurements captured in the frontal view: one uses facial mea-
surements of registered color and range data; the other merely uses color images. In

the first method, we have developed two techniques for face alignment:

0 2.5D snake: A 2.5D snake is designed to locally adapt a contour to each facial
component. The design of snake includes an iterative deformation formula,
placement of initial contours, and the minimization of energy functional. We
reformulated 2D active contours (a dynamic programming approach) for 3D
contours of eye, nose, mouth, and face boundary regions. We have constructed
initial contours based on the outputs of face detection (i.e., locations of the face
and facial components). We form energy maps for individual facial components

based on 2D color image and 2.5D range data, hence the name 2.5D snake.

0 Displacement propagation: This technique is designed to propagate the
displacement of a group of vertices on a 3D face model from contour points on
facial components to other points on non-facial components. The propagation

30

can be applied to a 3D face model whenever a facial component is coarsely

relocated or is finely deformed by the 2.5D snake.
In the second face modeling method, we developed a. technique for face alignment:

0 Interacting snakes: The snake deformation is formulated by a finite differ-
ence approach. The initial snakes for facial components are obtained from the
2D projection of the semantic face graph on a generic 3D face model. We have
designed the interacting snakes technique for manipulating multiple snakes it-
eratively that minimizes the attraction energy functionals on both contours
and enclosed regions of individual snakes and minimizes the repulsion energy

functionals among multiple snakes.

In face recognition, we have proposed two paradigms as shown in Figs. 1.11 and

1.12.

o The first (range data-based) recognition paradigm: This paradigm is de-
signed to automate and augment appearance-based face recognition approaches
based on 3D face models. In this system, we have integrated our face detection
algorithm, face modeling method using the 2.5D snake, and an appearance-
based recognition method using the hierarchical discriminant regression [78].
However, the recognition module can be replaced with other appearance—based
algorithms such as PCA-based and LDA-based methods. The system can learn
a 3D face model for an individual, and generate an arbitrary number of 2D
face images under different head poses and illuminations (can be extended to
different expressions) for training an appearance-based face classifier.

31

o The second (semantic) recognition paradigm: This paradigm is designed
to automate the face recognition process at a semantic level based on the dis-
tinctiveness and visibility of facial components in a given face image captured in
near frontal views. (This paradigm can be extended to face images taken in non—
frontal views). We have decomposed a generic 3D face model into recognition-
oriented facial components and non-facial components, and formed a 3D seman-

tic face graph for representing facial topology and extracting facial components.

 

In this recognition system, we have integrated our face detection algorithm, our
face modeling method using interacting snakes, and our semantic face matching
algorithm. The recognition can be achieved at a semantic level (e.g., comparing
faces based on eyes and the face boundary only) due to the alignment of facial
components. We have also introduced component weights, which play a crucial
role in face matching, to emphasize component’s distinctiveness and visibility.
The system can generate cartoon faces from aligned semantic face graphs and

facial caricatures based on an averaged face graph for face visualization.

32

Chapter 2

Literature Review

We first review the development of face detection and recognition approaches, fol-
lowed by a review of face modeling and model compression methods. Finally, we
will present one major application of face recognition technology, namely, face re-
trieval. We primarily focus on the methods that employ the task—specific cognition
or behaviors specified by humans (i.e., artificial intelligence pursuits), although there
are developmental approaches for facial processing (e.g., autonomous mental devel-

opment [79] and incremental learning [80] methods) that have emerged recently.

2. 1 Face Detection

Various approaches to face detection are discussed in [19], [20], [81],]82], and [83]. The
major approaches are listed chronologically in Table 2.1 for a comparison. For recent
surveys on face detection, see [82] and [83]. These approaches utilize techniques such
as principal component analysis (PCA), neural networks, machine learning, infor-

33

 

Table 2.1

SUMMARY OF VARIOUS FACE DETECTION APPROACHES.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Authors Year Approach Features Head Test Minimal
Used Pose Databases Face
Size
Féraud 2001 Neural net- Motion; Frontal Sussex; 15 x 20
et a1. [19] works Color; to pro— CMU; Web
Texture file images
DeCarlo 2000 Optical Motion; Frontal Videos NA
et a1. [61] flow Edge; to
Deformable profile
face model;
Texture
Maio et a1. 2000 Facial Texture; Frontal Video im— 20 x 27
[20] templates; Directional ages
Hough images
transform
Abdel- 1999 Skin Color Frontal HHI 13 x 13
Mottaleb model; to
et a1. [84] Feature profile
Garcia 1999 Statistical Color; Frontal MPEG 80 X 48
et al. [21] wavelet Wavelet to near videos
analysis coefficients frontal
Wu et a1. 1999 Dizzy color Color Frontal Still color 20 x 24
[85] models; to images
Template profile
matching
Rowley et 1998 Neural net- Texture (Upright) FERET; 20 x 20
a1. [24], works frontal CMU; Web
[23] images
Sung et a1. 1998 Learning Texture Frontal Video 19 x 19
[25] images;
newspaper
scans
Colmenarez 1997 Learning Markov pro- Frontal FERET 11 x 11
et al. [86] cesses
Yow et a1. 1997 Feature; Geometrical Frontal CMU 60 x 60
[26] Belief facial to
networks features profile
Lew et a1. 1996 Markov Most Frontal MIT; 23 X 32
[27] random informative CMU;
field; pixel Leiden
DFFS [64]

 

 

 

 

 

 

 

34

 

 

 

mation theory, geometrical modeling, (deformable) template matching, Hough trans-
form, extraction of geometrical facial features, motion extraction, and color analysis.
Typical detection outputs are shown in Fig. 2.1. In these images, a detected face
is usually overlaid with graphical objects such as a rectangle or an ellipse for a face,
and circles or crosses for eyes. The neural network-based [24], [23] and the view-
based [25] approaches require a large number of face and non-face training examples,
and are designed primarily to locate frontal faces in grayscale images. It is difficult
to enumerate “non-face” examples for inclusion in the training databases. Schnei-
derman and Kanade [22] extend their learning-based approach for the detection of
frontal faces to profile views. A feature-based approach combining geometrical fa-
cial features with belief networks [26] provides face detection for non-frontal views.
Geometrical facial templates and the Hough transform were incorporated to detect
grayscale frontal faces in real time applications [20]. Face detectors based on Markov
random fields [27], [87] and Markov chains [88] make use of the spatial arrangement
of pixel gray values. Model based approaches are widely used in tracking faces and
often assume that the initial location of a face is known. For example, assuming
that several facial features are located in the first frame of a video sequence, a 3D
deformable face model was used to track human faces [61]. Motion and color are very
useful cues for reducing search space in face detection algorithms. Motion information
is usually combined with other information (e.g., face models and skin color) for face
detection and tracking [89]. A method of combining a Hidden Markov Model (HMM)
and motion for tracking was presented in [86]. A combination of motion and color

filters, and a neural network model was proposed in [19].

35

 

 

.E a a 33

3 ”TE .E 8 30> 3v Amfi .E .5 wnsm 3 gwfi Ad pm zofiaom 3 .3 ”HQ .3 am Egom 3 ”fig .E 8 swESquaom CV .on

”SQ am pm. @630 AU gofi .6 no 062 A3 ”a: .3 Ho vsgwm 3V A3 mmfifiiowdm aoBoSoU 8.3 353% m0 3:350 AN 85mg
v

V
V

   

36

.Aquov

.2“ Baa

 

 

37

Categorizing face detection methods based on their representations of faces reveals
that detection algorithms using holistic representations have the advantage of finding
small faces or faces in poor-quality images, while those using geometrical facial fea-
tures provide a good solution for detecting faces in different poses. A combination of
holistic and feature—based methods [59], [60] is a promising approach to face detection
as well as face recognition. Motion [86], [19] and skin-tone color [19], [84], [90], [85],
[21] are useful cues for face detection. However, the color-based approaches face dif-
ficulties in robustly detecting skin colors in the presence of complex background and
variations in lighting conditions. Two color spaces (YCbC', and H S V) have been pro-
posed for detecting the skin color patches to compensate for lighting variations [21].
We propose a face detection algorithm that is able to handle a wide range of color
variations in static images, based on a lighting compensation technique in the RGB
color space and a nonlinear color transformation in the YCbC', color space. Our ap—
proach models skin color using a parametric ellipse in a two-dimensional transformed
color space and extracts facial features by constructing feature maps for the eyes,

mouth and face boundary from color components in the YCbC, space.

2.2 Face Recognition

The human face has been considered as the most informative organ for communication
in our social lives [49]. Automatically recognizing faces by machines can facilitate a
wide variety of forensic and security applications. The representation of human faces
for recognition can vary from a 2D image to a 3D surface. Different representations

38

result in different recognition approaches. Extensive reviews of approaches to face
recognition were published in 1995 [37], 1999 [31], and in 2000 [38]. A workshop
on face processing in 1985 [91] presented studies of face recognition mainly from the
viewpoint of cognitive psychology. Studies of feature-based face recognition, computer
caricatures, and the use of face surfaces in simulation and animation were summarized
in 1992 [49]. In 1997, Uwechue et a1. [92] gave details of face recognition based on high-
order neural networks using 2D face patterns. In 1998, lectures on face recognition
using 2D face patterns were presented from theory to applications [36]. In 1999,
Hallinan et a1. [93] described face recognition using both the statistical models for
2D face patterns and the 3D face surfaces. In 2000, Gong et a1. [94] emphasized
the statistical learning methods in holistic recognition approaches and discussed face
recognition from the viewpoint of dynamic vision.

The above studies show that the face recognition techniques, especially holistic
methods based on the statistical pattern theory, have greatly advanced over the past
ten years. Face recognition systems (e.g., FaceIt [1] and FaceSnap [2]) are being used
in video surveillance and security monitoring applications. However, more reliable
and robust techniques for face recognition as well as detection are required for several
applications. Except for the recognition applications based on static frontal images
that are taken under well-controlled environments (e.g., indexing and searching large
image database of drivers for issuing driving licenses), the main challenge in face
recognition is to be able to deal with the high degree of variability in human face
images. The sources of variations include inter-subject variations (distinctiveness of
individual appearance) and intra—subject variations (in 3D pose, facial expression,

39

facial hair, lighting, and aging). Some variations are not removable, while others
can be compensated for recognition. Persons who have similar face appearances, e. g.
twins, and an individual who could have different appearances due to cosmetics, or
other changes in facial hair and glasses are very difficult to recognize. Variations
due to different poses, illuminations, and facial expressions are relatively easy to
handle. Currently available algorithms for face recognition concentrate on recognizing
faces under those variations which can somehow be compensated for. Because facial
variations due to pose cause a large amount of appearance change, more and more
systems are taking advantage of 3D face geometry for recognition.

The performance of a recognition algorithm depends on the face databases it
is evaluated on. Several face databases, such as MIT [95], Yale [96], Purdue [97],
and Olivetti [98] databases are publically available for researchers. Figure 2.2 shows
some examples of face images from the FERET [28], MIT [29], and XM2VTS [30]
databases. According to Phillips [68], [28], the FERET evaluation of face recognition
algorithms identifies three state-of—the—art techniques: (i) the principal component
analysis (PCA)-based approach [99], [100], [29]; (ii) the elastic bunch graphic match-
ing (EBGM)-based paradigm [32]; and (iii) the local feature analysis (LFA)-based
approach [34], [101]. The internal representations of PCA-based, EBGM-based, and
LFA-based recognition approaches are shown in Figs. 2.3, 2.4, and 2.5, respectively.
To represent and match faces, the PCA-based approach makes use of a set of orthonor-
mal basis images; the EBGM-based approach constructs a face bunch graph, whose
nodes are associated with a set of wavelet coefficients (called jets); the LFA-based
approach uses localized kernels, which are constructed from PCA—based eigenvectors,

40

 

Figure 2.2. Examples of face images are selected from (a) the FERET database [28];
(b) the MIT database [29]; (c) the XM2VTS database [30].

41

for topographic facial features (e.g., eyebrows, cheek, mouth, etc.)

 

Mean MDFl MDF2 MDF3 MDF4 MDF5 MDF6 MDF7

(b)
Figure 2.3. Internal representations of the PCA-based approach and the LDA-based
approach (from Weng and Swets [31]). The average (mean) images are shown in the
first column. Most Expressive Features (MEF) and Most Discriminating Features
(MDF) are shown in (a) and (b), respectively.

The PCA—based algorithm provides a compact but non-local representation of
face images. Based on the appearance of an image at a specific view, the PCA
algorithm works at the pixel level. Hence, the algorithm can be regarded as “picture”
recognition, in other words, it is not explicitly using any facial features. The EBGM-
based algorithm constructs local features (extracted using Gabor wavelets) and global
face shape (represented as a graph), and so this approach is much closer to “face”
recognition. However, the EBGM algorithm is pose-dependent, and it requires initial
graphs for different poses during its training stage. The LFA-based algorithm is
derived from the PCA-based method; it is also called a kernel PCA method. In this
approach, however, the choice of kernel functions for local facial features (e.g., eyes,
mouth, and nose) and the selection of locations of these features still remains an open

42

face bunch graph

 

((1)
Figure 2.4. Internal representations of the EBGM—based approach (from Wiskott et
a1. [32]): (a) a graph is overlaid on a face image; (b) a reconstruction of the image
from the graph; (c) a reconstruction of the image from a face bunch graph using the
best fitting jet at each node. Images are downloaded from [33]; (d) a bunch graph
whose nodes are associated with a bunch of jets [33]; (e) an alternative interpretation
of the concept of a bunch graph [33].

question.

In addition to these three approaches, we categorize face recognition algorithms on
the basis of pose—dependency and matching features (see Fig. 2.6). In pose-dependent
algorithms, 3. face is represented by a set of viewer-centered images. A small number
of 2D images (appearances) of a human face at different poses are stored as a repre—
sentative set of the face, while the 3D face shape is implicitly represented in the set.

43

 

Figure 2.5. Internal representations of the LFA—based approach (from Penev and
Atick [34]). (a) An average face image is marked with five localized features; (b) five
topographic kernels associated with the five localized features are shown in the top
row, and the corresponding residual correlations are shown in the bottom row.

The representative set can be obtained from either digital cameras or extracted from
videos. On the other hand, in pose-invariant approaches, a face is represented by a
3D face model. The 3D face shape of an individual is explicitly represented, while
the 2D images are implicitly encoded in this face model. The 3D face models can be
constructed by using either 3D digitizers or range sensors, or by modifying a generic

face model using a video sequence or still face images of frontal and profile views.

The pose-dependent algorithms can be further divided into three classes:
appearance-based (holistic) [29], [78] feature—based (analytic) [102], [103] and hy-
brid (which combines holistic and analytic methods) [60], [99], [32], [34] approaches.
The appearance—based methods are sensitive to intra—subject variations, especially
to changes in hairstyle, because they are based on global information in an image.
However, the feature—based methods suffer from the difficulty of detecting local fidu-
cial “points”. The hybrid approaches were proposed to accommodate both global and
local face shape information. For example, LFA-based methods, eigen-template meth-

44

Algorithms

............................. Pose-dependency

Pose-Invariant
--. Face
representation

 
    

Pose-dependent
A\

Viewer-centered

Object-centered

   

 

Images Models
------------------------------------------------------------------------
Appearance Hybrid Feature- - Gordon et al., 1995
-based based - Lengqne at al., 1996
("Oust“) {Analytic} - Atick et al., 1996
PCA LDA LFA - Yan et al., 1996
EGBM '- Zhao ct al., 2000
- Zhang at al., 2000
-- Hsu et al., 2001
Figure 2.6. A breakdown of face recognition algorithms based on the pose—

dependency, face representation, and features used in matching.

ods, and shape-and-shape—free [104] methods belong to the hybrid approach which is
derived from the PCA methodology. The EBGM-based methods belong to the hybrid
approach that is based on 2D face graphs and wavelet transforms at each feature node
of the graphs. Although they are in the hybrid approach category, the eigen-template

matching and EBGM-based methods are much closer to feature-based approaches.

In the pose-invariant algorithms, 3D face models are utilized to reduce the varia-
tions in pose and illumination. Gordon et al. [105] proposed an identification system
based on 3D face recognition. The 3D model used by Gordon et al. is represented
by a number of 3D points associated with their corresponding texture features. This
method requires an accurate estimate of the face pose. Lengagne et al. [106] proposed
a 3D face reconstruction scheme using a pair of stereo images for recognition and mod-

45

eling. However, they did not implemented the recognition module. Atick et al. [107]
proposed a reconstruction method of 3D face surfaces based on the Karhonen-Loeve
(KL) transform and the shape-from—shading approach. They discussed the possibility
of using eigenhead surfaces in face recognition applications. Yan et al. [108] proposed
a 3D reconstruction method to improve the performance of face recognition by mak—
ing Atick et al.’s reconstruction method rotation-invariant. Zhao et al. [109] proposed
a method to adapt a 3D model from a generic range map to the shape obtained from
shading for enhancing face recognition performance in different lighting and viewing

conditions.

Based on our brief review, we believe that the current trend is to use 3D face
shape explicitly for recognition. In order to efficiently store an individual’s face,
one approach is to adapt a 3D face model [72] to the individual. There is still a
considerable debate on whether the internal recognition mechanism of a human brain
involves explicit 3D models or not [49], [110]. However, there is enough evidence to
support the fact that humans use information about 3D structure of objects (e.g.,
3D geometry of a face) for recognition. Closing our eyes and imagining a face (or a
chair) can easily verify this hypothesis, since the structure of a face (or a chair) can
appear in our mind without the use of eyes. Moreover, the use of a 3D face model
can separate both geometrical and texture features for facial analysis, and can also
blend both of them for recognition as well as visualization [67]. Our proposed systems

belong to this emerging trend.

46

2.3 Face Modeling

Face modeling plays a crucial role in applications such as human head tracking, facial
animation, video compression/ coding, facial expression recognition, and face recog-
nition. Researchers in computer graphics have been interested in modeling human
faces for facial animation. Applications such as virtual reality and augmented reality
[74] require modeling faces for human simulation and communication. In applications
based on face recognition, modeling human faces can provide an explicit representa-
tion of a face that aligns facial shape and texture features together for face matching

at different poses and in different illumination conditions.

2 .3. 1 Generic Face Models

We fiI‘St review three major approaches to modeling human faces and then point out

an advanced modeling approach that makes use of the a priori knowledge of facial

geometry. DeCarlo et al. [111] use the anthropometric measurements to generate a
general face model (see Fig. 2.7). This approach starts with manually-constructed
B‘spline surfaces and then applies surface fitting and constraint optimization to these
Surfaces. It is computationally intensive due to its optimization mechanism. In the
second approach, facial measurements are directly acquired from 3D digitizers or
Structured light range sensors. 3D models are obtained after a postprocessing, tri-
angul‘c‘trization, on these shape measurements. The third approach, in which models
are reconstructed from photographs, only requires low-cost and passive input devices
(Video cameras). Some computer vision techniques for reconstructing 3D data can

47

 

  

val —.
j.

   

1"; [
Lag _ F,

IN.
F?
,__.

l ear inclination?

 

 

awn.” ‘
-' mentooervlcal angle
(a) (b)
Figure 2.7. Face modeling using anthropometric measurements (downloaded
from [35]): (a) anthropometric measurements; (b) a B-spline face model.

be used for face modeling. For instance, Lengagne et al. [106] and Chen et al. [112]

built face models from a pair of stereo images. Atick et al. [107] and Yan et al.

[108] reconstructed 3D face surfaces based on the Karhonen-Loeve (KL) transform
and the shape-from—shading technique. Zhao et al. [109] made use of a symmet—
ric S118.1)e-from-shading technique to build a 3D face model for recognition. There
are other methods which combine both shape-from-stereo (which extracts low—spatial
freclllency components of 3D shape) and shape-from—shading (extracting high—spatial
frequency components) to reconstruct 3D faces [113], [114], [115]. See [116] for addi-
tional methods to obtain facial surface data. However, currently it is still difficult to
extract. sufficient information about the facial geometry only from 2D images. This
diffiCulty is the reason why Guenter et al. [117] utilize a large number of fiducial
points to capture 3D face geometry for photorealistic animation. Even though we
Can C>btain dense 3D facial measurements from high-cost 3D digitizers, it takes too
much time and it is expensive to scan a large number of human subjects.

48

a
‘3

fl

lll

An advanced modeling approach which incorporates a priori knowledge of fa—
cial geometry has been proposed for efficiently building face models. We call the
model representing the general facial geometry as a generic face model. Waters” face
model [69], shown in Fig. 2.8(a), is a well-known instance of polygonal facial surfaces.

Figure 2.8(b) shows some other generic face models. The one used by Blanz and Vet-

 
 
   

iiflllllflhs
mg "as“.
r WW.
w

 
   
   

Maphnblo
face Mode
: E

‘Ilhlililllllill

. c. , ~\ . 2,,
WV
will;

U

is ;;—:9Asav!figf3- :7
Amberw 999

 

(b)
Figure 2.8. Generic face models: (a) Water’s animation model; (b) anthropometric
meaSlll‘ements; (b) six kinds of face models for representing general facial geometry.

ter is a. statistics—based face model which is represented by the principal components
0f Shape and texture data. Reinders et al. [72] used a fairly coarse wire-frame model,
compared to Waters’ model, to do model adaptation for image coding. Yin et al.
[118] proposed a MPEG4 face modeling method that uses fiducial points extracted
from tWo face images at frontal and profile views. Their feature extraction is simply
based on the results of intensity thresholding and edge detection. Similarly, Lee et
a1. [1 19] have proposed a method that modifies a generic model using either two or-
t'hogonal pictures (frontal and profile views) or range data, for animation. Similarly,

49

for facial animation, Lengagne et al. [120] and Fua [121] use bundle-adjustment and
least—squares fitting to fit a complex animation model to uncalibrated videos. This
algorithm makes use of stereo data, silhouette edges, and 2D feature points. Five
manually-selected features points and initial values of camera positions are essential
for the convergence of this method. Ahlberg [122] adapts a 3D wireframe model
(CANDIDE—3 [123]) to a 2D video image. The two modeling methods proposed in
this thesis follow the modeling approach using a generic face model; both of our meth-
ods make use of a generic face model (Waters’ face model) as a priori knowledge of
facial geometry and employ (i) displacement propagation and 2.5D snakes in the first
method and (ii) interacting snakes and semantic face graphs in the second method

for adapting recognition-orientated features to an individual’s geometry.

2.3.2 Snakes for Face Alignment

AS 8. Computational bridge between the high-level a priori knowledge of object shape
and the low-level image data, snakes (or active contours) are useful models for extract-
ing the Shape of deformable objects. Similar to other template-based approaches such
as Hough transform and active shape models, active contours have been employed
to detect object boundary, track objects, reconstruct 3D objects (stereo snakes and
interrframe snakes), and match/ identify shape. Snakes self converge in an iterative
Way, and deform either with or without topological constraints.

I{esearch on active contours focuses on issues related to representation (e.g., para-
metric curves, splines, Fourier series, and implicit level—set functions), energy func-

50

 

tionals to minimize, implementation methods (e.g., classical finite difference models,
dynamic programming [124], and Fourier spectral methods), convergence rates and
conditions, and their relationship to statistical theory [125] (e.g., the Bayesian esti-
mation). Classical snakes [77], [126] are represented by parametric curves and are de—
formed by finite difference methods based on edge energies. In applications, different
types of edge energies including image gradients, gradient vector flows [127], distance
maps, and balloon force have been proposed. On the other hand, combined with
level-set methods and the curve evolution theory, active contours have emerged as a
powerful tool, called geodesic active contours (GAC) [128], to extract deformable ob-
jects with unknown geometric topology. However, in the GAC approach, the contours
are implicitly represented as level-set functions and are closed curves. In addition to
the edge energy, region energy has been introduced to improve the segmentation re-
sults for homogeneous objects in both the parametric and the GAC approaches (e. g.,
region and edge [129], GAC without edge [130], statistical region snake [131], region
competition [132], and active region model [133]). Recently, multiple active contours
[134], [135] were proposed to extract / partition multiple homogeneous regions that do

not overlap with each other in an image.

In our first alignment method, we have reformulated 2D active contours (a dy-
namic programming approach) in 3D coordinates for energies derived from 2.5D range
and 2D color data. In our second alignment method, we make use of multiple 2D
snakes (a finite difference approach) that interact with each other in order to adapt
facial components.

51

2.3.3 3D Model Compression

Among various representations of 3-D objects, surface models can explicitly represent
shape information and can effectively provide a visualization of these objects. The
polygonal model using triangular meshes is the most prevalent type of surface rep-
resentations for free-form objects such as human faces. The reason is that the mesh
model explicitly describes the connectivity of surfaces, enables mesh simplification,
and is suitable for free-form objects [136]. The polygonization of an object surface ap-
proximates the surface by a large number of triangles (facets), each of which contains
primary information about vertex positions as well as vertex associations (indices),
and auxiliary information regarding facet properties such as color, texture, specu-
larity, reflectivity, orientation, and transparency. Since we use a triangular mesh to
represent a generic face model and an adapted model, model compression is preferred

when efficient transmission, visualization, and storage is required.

In 1995, the concept of geometric compression was first introduced by Deer-
ing [137], who proposed a technique for lossy compression of 3—D geometric data.
Deering’s technique focuses mainly on the compression of vertex positions and facet
properties of 3-D triangle data. Taubin [75] proposed topological surgery which further
contributed connectivity encoding (compression of association information) to geo-
metric compression. Lounsbery et al. [76] performed geometric compression through
multiresolution analysis for particular meshes with subdivision connectivity. Apply-
ing remeshing algorithms to arbitrary meshes, Eck et al. [138] extended Lounsbery’s
work on mesh simplification. Typical compression ratios in this line of development

52

are listed in Table 2.2. All of these compression methods focus on model represen-

Table 2.2
GEOMETRIC COMPRESSION EFFICIENCY.

 

 

Compression [137]

Method Geometric Loss Measure Compressed feature
Compression
Ratio (GCR)

Geometric 6—10 slight losses Positions, normals,

colors

 

 

Topological 20—100 no loss Connectivity;
Surgery [75]
12-30 N / A Positions, facet
properties;
20—100 N / A ASCII-file sizes
Remeshing [138] 54-12 Remeshing & com- Level of detail
pression tolerances (facets)

 

 

 

 

 

 

tation using triangular meshes. However, for more complex 3D shapes, the surface
representation using triangular meshes usually results in a large number of triangu-
lar facets, because each triangular facet is explicitly described. We have developed
a novel compression approach for free-form surfaces using 3D wavelets and lattice
vector quantization [139]. In our approach, surfaces are implicitly represented in-
side a volume in the same way as edges in a 2D image. A further improvement in
our approach can be achieved by making use of integer wavelet transformation [140],

[141].

2.4 Face Retrieval

Face recognition technology provides a useful tool for content-based image and video
retrieval using the concept of human faces. Based on face detection and identification
technology, we can design a system for consumer photo management (or for web
graphic search) that uses human faces for indexing and retrieving image content and

generates annotation (textual descriptions) for the image content automatically.

Traditional text-based retrieval systems for digital libraries can not fulfill a re-
trieval of visual content such as human faces, eye shape, and cars in image or video
databases. Hence, many researchers have been developing multimedia retrieval tech-
niques based on automatically extracting salient features from the visual content (see
[40] for an extensive review). Well known systems for content-based image and video
retrieval are QBIC [142], Photobook [143], CONIVAS [144], FourEyes [145], Virage
[146], ViBE [147], VideoQ [148], Visualseek [149], Netra [150], MARS [151], PicSOM
[152], ImageScape [153], etc. In these systems, retrieval is performed by comparing
a set of low-level features of a query image or video clip with features stored in the
database and then by presenting the user with the content that has the most similar
features. However, users normally query an image or video database based on seman-
tics rather than low-level features. For example, a typical query might be specified
as “retrieve images of fireworks” rather than “retrieve images that have large dark

regions and colorful curves over the dark regions”.

Since the commonly used features are usually a set of unorganized low-level at-
tributes (such as color, texture, geometrical shape, layout, and motion), grouping

54

low-level features can provide meaningful high-level semantics for human consumers.
There has been some work done on automatically classifying images into semantic
categories [154], such as indoors/outdoors and city/landscape images. As for the
semantic concept of faces, the generic facial topology (e.g., our proposed generic se—
mantic face graph) is a useful structure for representing the face in a search engine.
We have designed a graphical user interface for face editing using our face detection
algorithm. Combined with our semantic face matching algorithm, we can build a face

retrieval system.

2.5 Summary

We have briefly described the development of face detection, face recognition, face
modeling and model compression in this chapter. We have summarized the per-
formance of currently available face detection systems in Table 2.3. Note that the
performance of a detection system depends on several factors such as face databases
on which the system is evaluated, system architecture, distance metric, and algorith-
mic parameters. The performance is evaluated based on the detection rate, the false
positive rate (false acceptance rate), and databases. In Table 2.3, we do not include
the false acceptance rate because the false positive rate has not been completely re-
ported in literature. We refer the reader to the FERET evaluation [68], [28] for the

performance of various face recognition systems.

Face detection and face recognition are closely related to each other in the sense
of categorizing faces. Over the past ten years, based on the statistical pattern theory,

55

Table 2.3

SUMMARY OF PERFORMANCE OF VARIOUS FACE DETECTION APPROACHES.

 

 

 

 

 

 

 

 

 

 

Authors Year Head Test Detection Rate
Pose Databases
F éraud et al. 2001 Frontal Sussex; 100% for Sussex;
[19] to CMU testl; 81% ~ 86% for CMU testl;
profile Web images 74.7% ~ 80.1% for Web
images.
Maio et al. 2000 Frontal Static images 89.53% ~ 91.34%
[201
Schneiderman 2000 Frontal CMU; Web 75.24% ~ 92.7%
et al. [22] to images
profile
Garcia et a1. 1999 Frontal MPEG videos 93.27%
[21] to near
frontal
Rowley et al. 1998 (Upright) CMU; 86%[24]; 79.6%[23] for ro-
[24], [23] frontal F ERET; tated faces
Web images
Yow et al. 1997 Frontal CMU 84% ~ 92%
[26] to
profile
Lew et al. [27] 1996 Frontal MIT; CMU; 87% ~ 95%
Leiden

 

 

 

 

 

56

the appearance-based (holistic) approach has greatly advanced the field of face recog-
nition. By categorizing face detection methods based on their representations of the
face, we observe that detection/ recognition algorithms using holistic representations
have the advantage of finding/ identifying small faces or faces in poor-quality images
(i.e. detection/ recognition under uncertainty), while those using geometrical facial
features provide a good solution for detecting / recognizing faces in different poses and
expressions. The internal representation of a human face substantially affects the per-

formance and design of a detection or recognition system. A seamless combination of

 

holistic 2D and geometrical 3D features provides a promising approach to represent
faces for face detection as well as face recognition. Modeling human face in 3D space
has been shown to be useful for face recognition. However, the important aspect
of face modeling is how to efficiently encode the 3D facial geometry and texture as

compact features for face recognition.

57

Chapter 3

Face Detection

We will first describe an overview of our proposed face detection algorithm and then
give details of the algorithm. We will demonstrate the performance and experimental

results on several image databases.

3.1 Face Detection Algorithm

The use of color information can simplify the task of face localization in complex
environments [19], [84], [90], [85]. Therefore, we use skin color detection as the first
step in detecting faces. An overview of our face detection algorithm is depicted in
Fig. 3.1, which contains two major modules: (i) face localization for finding face
candidates; and (ii) facial feature detection for verifying detected face candidates.
The face localization module combines the information extracted from the luminance
and the chrominance components of color images and some heuristics about face
shape (e.g., face sizes ranging from 13 x 13 pixels to about three fourths of the image

58

size) to generate potential face candidates, within the entire image. The algorithm
first estimates and corrects the color bias based on a novel lighting compensation
technique. The corrected red, green, and blue color components are first converted
to the YCbCT color space and then nonlinearly transformed in this color space (see
formulae in Appendix A). The skin-tone pixels are detected using an elliptical skin
model in the transformed space. The parametric ellipse corresponds to contours of
constant Mahalanobis distance under the assumption of the Gaussian distribution of
skin tone color. The detected skin—tone pixels are iteratively segmented using local
color variance into connected components which are then grouped into face candidates
based on both the spatial arrangement of these cemponents (described in Appendix
B) and the similarity of their color [84]. Figure 3.1 shows the input color image, color
compensated image, skin regions, grouped skin regions, and face candidates obtained
from the face localization module. Each grouped skin region is assigned a pseudo
color and each face candidate is represented by a rectangle. Because multiple face
candidates (bounding rectangles) usually overlap, they can be fused based on the
percentage of overlapping areas. However, in spite of this postprocessing there are
still some false positives among face candidates.

It is inevitable that detected skin—tone regions will include some non—face regions
whose color is similar to the skin-tone. The facial feature detection module rejects
face candidate regions that do not contain any facial features such as eyes, mouth,
and face boundary. This module can detect multiple eye and mouth candidates. A
triangle is constructed from two eye candidates and one mouth candidate, and the
best-fitting enclosing ellipse of the triangle is constructed to approximate the face

59

 

 

 

 

 

 

 

   
 
    
  
 

 

 

Regions

 

Segmented
Skin
Regions

 

 
 

 

 

   
 
     
 

 

 

, . ._ . Grouped
[confiscate 2» Skin
_’ ‘ . fireupine :3. Regions

  

 

 

 

1 Face Localization ]

 

Face
Candidates

 

 

 

 

 

s
L Eye/Maui?! fiéecfionj

 

 

[Face Boundary Detection]

Verifying/Weighting l—ji
E es-Mouth Trian les ~77"
Facial

Feature Detection

Face
Descriptor

 

    

 

    

  
 

 

 

 

 

Figure 3.1. Face detection algorithm. The face localization module finds face candi-
dates, which are verified by the detection module based on facial features.

60

boundary. A face score is computed for each set of eyes, mouth and the ellipse.
Figure 3.1 shows a detected face and the enclosing ellipse with its associated eye-
mouth triangle which has the highest score that exceeds a threshold. These detected
facial features are grouped into a structured facial descriptor in the form of a 2D
graph for face description. These descriptors can be the input to subsequent modules
such as face modeling and recognition. We now describe in detail the individual

components of the face detection algorithm.

3.2 Lighting Compensation and Skin Tone

Detection

The appearance of the skin-tone color can change due to different lighting conditions.
We introduce a lighting compensation technique that uses “reference white” to nor-
malize the color appearance. We regard pixels with the top 5% of the luma (nonlinear
gamma-corrected luminance) values as the reference white if the number of these pix—
els is sufficiently large (> 100). The red, green, and blue components of a color image
are adjusted so that these reference-white pixels are scaled to the gray level of 255.
The color components are unaltered if a sufficient number of reference—white pixels
is not detected. This assumption is reasonable not only because an image usually
contains “real white” (i.e., white reference in [155]) pixels in some regions of interest
(such as eye regions), but also because the dominant bias color always appears in
the “real white”. Figure 3.2 demonstrates an example of our lighting compensation

61

 

    
 

. liltilu-n‘kiluw “mm

 

Figure 3.2. Skin detection: (a) a yellow-biased face image; (b) a lighting compensated
image; (c) skin regions of (a) shown in white; (d) skin regions of (1)).
method. Note that the yellow bias color in Fig. 3.2(a) has been removed, as shown in
Fig. 3.2(b). The effect of lighting compensation on detected skin regions can be seen
by comparing Figs. 3.2(a) and 3.2(d). With lighting compensation, our algorithm
detects fewer non-face areas and more skin—tone facial areas. Note that the varia-
tions in skin color among different racial groups, reflection characteristics of human
skin and its surrounding objects (including clothing), and camera characteristics will
all affect the appearance of skin color and hence the performance of an automatic
face detection algorithm. Therefore, if models of the lighting source and cameras are
available, additional lighting correction should be made to remove color bias.
Modeling skin color requires choosing an appropriate color space and identifying
a cluster associated with skin color in this space. It has been observed that the

62

normalized red-green (r-g) space [156] is not the best choice for face detection [157],
[158]. Based on Terrillon et al.’s [157] comparison of nine different color spaces for
face detection, the tint-saturation-luma (TSL) space provides the best results for
two kinds of Gaussian density models (unimodal and mixture of Gaussian densities).
We adopt the YCbCr space since it is perceptually uniform [155], is widely used in
video compression standards (e.g., MPEG and JPEG) [21], and it is similar to the
TSL space in terms of the separation of luminance and chrominance as well as the
compactness of the skin cluster. Many research studies assume that the chrominance
components of the skin-tone color are independent of the luminance component [159],
[160], [158], [90]. However, in practice, the skin-tone color is nonlinearly dependent
on luminance. In order to demonstrate the luma dependency of skin-tone color, we
manually collected training samples of skin patches (853, 571 pixels) from 9 subjects
(137 images) in the Heinrich—Hertz-Institute (HHI) image database [15]. These pixels
form an elongated cluster that shrinks at high and low luma in the YCbCr space,
shown in Fig. 3.3(a). Detecting skin tone based on the cluster of training samples in
the Cb-C, subspace, shown in Fig. 3.3(b), results in many false positives. If we base
the detection on the cluster in the (Cb/ Y)-(Cr / Y) subspace, shown in Fig. 3.3(c), then
many false negatives result. The dependency of skin tone color on luma is also present
in the normalized rgY space in Fig. 3.4(a). the perceptually uniform C I E :er space
in Fig. 3.4(c), and the HSV spaces in Fig. 3.4(e). The 3D cluster shape changes at
different luma values, although it looks compact in the 2D projection subspaces, in
Figs. 3.4(b), 3.4(d) and 3.4(f).

To deal with the skin-tone color dependence on luminance, we nonlinearly trans-

63

form the YCbCr color space to make the skin cluster luma—independent. This is done
by fitting a piecewise linear boundary to the skin cluster (see Fig. 3.5). The details
of the model and the transformation are described in Appendix A. The transformed
space, shown in Fig. 3.6(a), enables a robust detection of dark and light skin tone
colors. Figure 3.6(b) shows the projection of the 3D skin cluster in the transformed
Cb-Cr color subspace, on which the elliptical model of skin color is overlaid. Figure 3.7
shows examples of detection using the nonlinear transformation. More skin-tone pix-
els with low and high luma are detected in this transformed subspace than in the

050, subspace.

64

 

 

 

 

 

 

 

 

 

Figure 3.3. The YCbCr color space (blue dots represent the reproducible color on a
monitor) and the skin tone model (red dots represent skin color samples). (a) The
YCbC, space; (b) a 2D projection in the Cb—C', subspace; (c) a 2D projection in the
(Cb/Y)-(C,/Y) subspace.

65

 

 

*<

9999
e959o

-
l

°PP°P
Anyhm

  

o9 -

.0, . _.
awf,
7' l..f”.~.

Jr"

 

ATTT.‘.0103
on 03305

M
Figure 3.4. The dependency of skin tone color on lunIa. The skin tone cluster (red
dots) is shown in (a) the rgY, (c) the CIE er, and (e) the HSV color spaces; the
2D projection of the cluster is shown in (b) the r — g, (d) the 2: — y , and (f) S —- H
color subspaces, where blue dots represent the reproducible color on a monitor. For
a better presentation of cluster shape. we normalize the luma Y in the rgY and the
CI E .er by 255, and swap the hue and saturation coordinates in the H SV space.
The skin tone cluster is less compact at low saturation values in (e) and (f)

I
.°

 

66

3oq'Cb l — Skin-{OWEREEST 300cr — Skin- tone Samples

, ‘ -[:- FILIEG Biodet -_ ] I I - Fitted Model
250] 250»
zoo] 200
150 150»

   

 

 

 

100 100*
50 50
o E . - .L o . . 1 _, Y -
0 50 100 150 200 250 0 50 100 150 200 250
(a) (b)

Figure 3.5. 2D projections of the 3D skin tone cluster in (a) the Y-Cb subspace; (b)
the Y-C, subspace. Red dots indicate the skin cluster. Three blue dashed curves,
one for cluster center and two for boundaries, indicate the fitted models.

 

 

 

 

Figure 3.6. The nonlinear transformation of the YCbCr color space. (a) The trans-
formed YCbC'r color space; (b) a 2D projection of (a) in the Cb-Cr subspace, in which
the elliptical skin model is overlaid on the skin cluster.

67

35%:er 2: 53» 33mg .93
:EEoe €5.63. 9: 5 $9: amass .anfideb Qt ”:6an 983.08% one 25me as? Ed :EBou umum 23 E 96% memes: 2:
63598 come Em .5997. one. Showcases. 2: 50:33 was its .moEEexo qoSoBoc cam 45898:. 830 Soszcoz Sn 8de

 

 

 

 

 

 

 

 

 

 

 

 

68

3.3 Localization of Facial Features

Among the various facial features, eyes, mouth, and face boundary are the most
prominent features for recognition [103] and for estimation of 3D head pose [161],
[162]. Most approaches for eye [163], [164], [165], [166]. [167], mouth [165], [168],
face boundary [165], and face [20] localization are template based. However, our
approach is able to directly locate eyes, mouth, and face boundary based on their
feature maps derived from the the luma and the chroma of an image, called the eye
map, the mouth map and the face boundary map, respectively. For computing the
eye map and the mouth map, we consider only the area covered by a face mask that
is built by enclosing the grouped skin—tone regions with a pseudo convex hull, which
is constructed by connecting the boundary points of skin-tone regions in horizontal

and vertical directions. Figure 3.8 shows an example of the face mask.

     

(C) (d)
Figure 3.8. Construction of the face mask. (a) Face candidates; (b) one of the face
candidates; (c) grouped skin areas; ((1) the face mask.

69

3.3.1 Eye Map

We first build two separate eye maps, one from the chrominance components and
the other from the luminance component of the color image. These two maps are
then combined into a single eye map. The eye map from the chroma is based on
the observation that high Cb and low C, values are found around the eyes. It is
constructed from information contained in Cb, the inverse (negative) of Cr, and the

ratio Cb/Cr, as described in Eq. (3.1).

EyeMapC = %{ (03) + (a)2 + (Cb/c.) }, (3.1)

where 03, (CA2, and Cb/C'r all are normalized to the range [0,255] and Cr is the
negative of Cr (i.e., 255 — 0,). An example of the eye map from the chroma is shown
in Fig. 3.9(a).

The eyes usually contain both dark and bright pixels in the luma component.
Based on this observation, grayscale morphological operators (e.g., dilation and ero-
sion) [169] can be designed to emphasize brighter and darker pixels in the luma
component around eye regions. These operations have been used to construct feature
vectors for face images at multiple scales for frontal face authentication [66]. We
use grayscale dilation and erosion with a hemispheric structuring element at a single

estimated scale to construct the eye map from the luma, as described in Eq. (3.2).

Yet, y) 69 90a, y)

EyeMapL =
Y(IL',y) e 90(xay) +1

 

, (3.2)

where the grayscale dilation {9 and erosion 9 operations [169] on a function f : f C
R2 ——> R using a structuring function g : g C R2 -—> R are defined as follows.

70

A” .' Hem

 

(C)
Figure 3.9. Construction of eye maps: (a) from chroma; (b) from luma; (c) the
combined eye map.

71

(f 69 go)($, y) = MaX{f(IB - Ca y — 7‘) + 9(Ca7‘)};
(:r—c,y—r) 6f, (c,r)Eg,
(3.3)

(f 9 ga)(:v, y) = Min{f(rr - c, y - r) + 9(6, r)};

(x—c,y—r) Ef, (c,r) 69,
(3.4)

Ial - (ll — (R($,y)/0)2|1/2 — 1); R 5 lol,
—00; R > Io],

R(:c,y) = \/$2+y2, (3.6)

90(1'3 y) = (3.5)

where o is a scale parameter, which will be described later in Eq. (3.11). An example

of a hemispheric structuring element is shown in Fig. 3.10. The construction of the

 

Figure 3.10. An example of a hemispheric structuring element for grayscale morpho-
logical dilation and erosion with o = 1.

eye map from the luma is illustrated in Fig. 3.9(b). Note that before performing the
grayscale dilation and erosion operations, we fill the background of the face mask
with the mean value of the luma in the face mask (skin regions) in order to smooth

the noisy boundary of detected skin areas.

72

The eye map from the chroma is enhanced by histogram equalization, and then

combined with the eye map from the luma by an AND (n‘Iultiplication) operation in

Eq. (3.7).

EyeMap = ( EyelldapC' ) AND ( EyelVIapL ) . (3.7)

The resulting eye map is dilated, masked, and normalized to brighten the eyes and
suppress other facial areas, as can be seen in Fig. 3.9(c). The locations of the eye
candidates are initially estimated from the pyramid decomposition of the eye map,
and then refined using iterative thresholding and binary morphological closing on this

eye map.

3.3.2 Mouth Map

The color of mouth region contains more red component compared to the blue compo—
nent than other facial regions. Hence, the chrominance component 0,, proportional
to (red — Y), is greater than 05, proportional to (blue — Y), near the mouth areas.
We further notice that the mouth has a relatively low response in the Cr/Cb feature,

but it has a high response in CB. We construct the mouth map as follows:

 

Mouthllfap = CE-(CE—n-Cr/Cb)2; (3.8)
.1: 2 Great/)2
(1:,y)€]:g
n = 0.95- , (3.9)
i Z Crf$ay)/Cb($,y)
(2:,y)E.7:g

where both C3 and Cr/Cb are normalized to the range [0,255], and n is the number
of pixels within the face mask, $9. The parameter 7) is estimated as the ratio of the

73

average Cf to the average Cr/Cb. Figure 3.11 shows the major steps in computing
the mouth map of the subject in Fig. 3.9. Note that after the mouth map is dilated,
masked, and normalized, it is dramatically brighter near the mouth areas than at

other facial areas.

Mouth map Dilated &
masked

 

Difference

Figure 3.11. Construction of the mouth map.

3.3.3 Eye and Mouth Candidates

We form an eye-mouth triangle for all possible combinations of two eye candidates
and one mouth candidate within a face candidate. We then verify each eye—mouth
triangle by checking (i) luma variations and average gradient orientations of eye and
mouth blobs; (ii) geometry and orientation constraints of the triangle; and (iii) the
presence of a face boundary around the triangle. A weight is computed for each
verified eye—mouth triangle. The triangle with the highest weight that exceeds a
threshold is selected. We discuss the detection of face boundary in Section 3.3.4, and
the selection of the weight and the threshold in Section 3.3.5.

Note that the eye and mouth maps are computed within the entire areas of the
face candidate, which is bounded by a rectangle. The search for the eyes and the

74

mouth is performed within the face mask. The eye and mouth candidates are located
by using (i) a pyramid decomposition of the eye/ mouth maps and (ii) an iterative
thresholding and binary morphological closing on the enhanced eye and mouth maps.

The number of pyramid levels, L, is computed from the size of the face candidate, as

defined in Eqs. (3.10) and (3.11).

L 2 Max{ [1082(20)la [log2(Min(W, Hl/Fcll } i (310)
o = [von/(2-Fe)], (3.11)

where W and H represent the width and height of the face candidate; Fc X FC is the
minimum expected size of a face candidate; a is a spread factor selected to prevent
the algorithm from removing small eyes and months in the morphological operations;
and Fe is the maximal ratio of an average face size to the average eye size. In our
implementation, Fe is 7 pixels, and Fe is 12 pixels.

The coarse locations of eye and mouth candidates obtained from the pyramid de-
composition are refined by checking the existence of eyes / mouth blobs which are ob-
tained after iteratively thresholding and (morphologically) closing the eye and mouth
maps. The iterative thresholding starts with an initial threshold value, reduces the
threshold step by step, and stops when either the threshold falls below a stopping
value or when the number of feature candidates reaches pre—determined upper bounds,
New, for the eyes and .I mth for the mouth. The threshold values are automatically

computed as follows.

Th 2 E Z Map(:z:,y)+(1— 00- Max .Map(:c,y), (3.12)
n (1‘,y)€fg (:r,y)e]-'g

75

where .M ap(:1:,y) is either the eye or the mouth map; the parameter a is equal to
0.5 for the initial threshold value, and is equal to 0.8 for the stopping threshold.
The use of upper bounds on the number of eye and mouth candidates can prevent
the algorithm from spending too much time in searching for facial features. In our
implementation, the maximum number of eye candidates, New, is 8 and the maximum

number of mouth candidates, Nmth, is 5.

3.3.4 Face Boundary Map

Based on the locations of eyes / mouth candidates, our algorithm first verifies whether
the average orientation of luma gradients around each eye matches the interocular
direction, and then constructs a face boundary map from the luma. Finally, it utilizes
the Hough transform to extract the best-fitting ellipse. The fitted ellipse is used for
computing the eye-mouth triangle weight. Figure 3.12 shows the boundary map that
is constructed from both the magnitude and the orientation components of the luma
gradient within the regions that have positive orientations of the gradient orientations
(i.e., have counterclock-wise gradient orientations). We have modified Canny edge
detection [170] algorithm to compute the gradient of the luma as follows. The gradient
of a luma subimage, S (a: g), which is slightly larger than the face candidate in size

is estimated by

V5013, y) = (0,, Cy) = (130(1) ® SCI?) .71), Da(y) 69 Stay», (3-13)

76

   

  

Orientation

Magnitude

 

Face boundary map

Figure 3.12. Computation of face boundary and the eye-mouth triangle

77

where Da(r) is the derivative of the Gaussian with zero mean and variance 02, and
® is the convolution operator. Unlike the Canny edge detector, our edge detection
requires only a single standard deviation 0 (a spread factor) for the Gaussian that is

estimated from the size of the eye—mouth triangle.

—qu2 “2
Z _— ' Z M . tic, . em 7 -
a (8 ln(wh)) , ws ax(dis dist ) (3 14)

where as is the window size for a Gaussian, which is the maximum value of the
interocular distance (distw) and the distance between the interocular midpoint and
the mouth (distem); wh = 0.1 is the desired value of the Gaussian distribution at the
border of the window. In Fig. 3.12, the magnitudes and orientations of all gradients
have been squared and scaled between 0 and 255. Fig. 3.12 shows that the gradient
orientation provides more information to detect face boundaries than the gradient
magnitude. So, an edge detection algorithm is applied to the gradient orientation
and the resulting edge map is thresholded to obtain a mask for computing the face
boundary. The gradient magnitude and the magnitude of the gradient orientation
are masked, added, and scaled into the interval [0, 1] to construct the face boundary
map. The center of a face, indicated as a white rectangle in the face boundary map

in Fig. 3.12, is estimated from the first-order moment of the face boundary map.

The Hough transform is used to fit an elliptical shape to the face boundary map.
An ellipse in a plane has five parameters: an orientation angle, two coordinates of the

78

 

center, and lengths of major and minor axes. Since we know the locations of eyes and
mouth, the orientation of the ellipse can be estimated from the direction of a vector
that starts from the midpoint between the eyes towards the mouth. The location of
the ellipse center is estimated from the face boundary map. Hence, we need only a
two-dimensional accumulator for estimating the ellipse for bounding the face. The
accumulator is updated by perturbing the estimated center by a few pixels for a more

accurate localization of the ellipse.

3.3.5 Weight Selection for a Face Candidate

For each face in the image, our algorithm can detect several eye-mouth—triangle candi-
dates that are constructed from eye and mouth candidates. Each candidate is assigned
a weight which is computed from the eye and mouth maps, the maximum accumula-
tor count in the Hough transform for ellipse fitting, and face orientation that favors
vertical faces and symmetric facial geometry, as described in Eqs. (3.15)-(3.19). The
eye-mouth triangle with the highest weight (face score) that is above a threshold is
retained. In Eq. (3.15), the triangle weight, tw(i,j, k), for the i-th and the j-th eye
candidates and the k-th mouth candidate is the product of the eye-mouth weight,
emw(i,j, k), the face-orientation weight, ou;(i,j, k), and boundary quality, q(i,j, k).
The eye-mouth weight is the average of the eye-pair weight, ew(i, j), and the mouth
weight, mw(k), as described in Eq. (3.16).

79

two's, k) = 6mw(i,j,kl°mU(i,j./<)40,1316); (3.15)

 

 

1
emw(i,j, k) = §(ew(i,j) + mw(k‘)); (3.16)
. . _ EyeMaptvtz/i)+EyeMap($J-,yjl,
ew(zi]) "7 i
2 - EyeMap(:rm, ym)
i > j; i,j E [1,Neye]; (3.17)
MOUthMapffEk, yr)
k = ' k 1 Nm ; .1
”w“ ) MouthMap<xm,y..‘.)’ E] ’ “’1 (3 8)
2 s _.
mum. k) = flea—“rmrwrtist; comm k» = fl
r=1 r
||u','.|| = 1. (3.19)

Eq. (3.17) describes the eye—pair weight which is the normalized average of the eye
map value around the two eyes, where EyeMap(.r,-, y,) is the eye map value for the
i-th eye candidate (associated with an eye blob and a corresponding pixel in the
lowest level of the image pyramid). Eyehlaphrmym) is the eye map value for the
most significant eye candidate (having the highest response within the eye map). The
mouth weight, m'w(k) in Eq. (3.18), is obtained by normalizing the mouth map value
at the k-th mouth candidate (i.e., a mouth blob), 1W outlilll ap(:1:k, yk), by the mouth
map value at the most significant mouth candidate, 1W outhllf ap(;rm, y,,,). The face-
orientation weight, described in Eq. (3.19), is the product of two attenuation terms,
each of which is an exponential function of a projection ((1056,) of a vector (6,1) along a
particular direction (LIL), where r = 1, 2. As can be seen in Fig. 3.13, one term favors
a symmetric face, and it is a projection (c0561) of the vector iii (from the midpoint of

the two eyes to the mouth) along a vector (217]) that is perpendicular to the interocular

80

segment. The other term favors an upright face, and is a projection of a vector 172
(from the mouth to the midpoint of the two eyes) along the vertical axis (15) of the
image plane. The exponential function, shown in Fig. 3.14, is designed such that the
attenuation has the maximal value of 1 when 91 = 62 2 0° (i.e., when eyes and mouth
form a letter “T” or equivalently the face is upright), and it decreases to below 0.5
at 61 = 62 = 25°. The quality of face boundary, q(i, j, k), can be directly obtained

from the votes received by the best elliptical face boundary in the Hough transform.

 

41%;:th k
11

Figure 3.13. Geometry of an eye-mouth triangle, where v] = —u'§; unit vectors if} and
u} are perpendicular to the interocular segment and the horizontal axis, respectively.

0
Weight

" er "'“
-80 so 40 -20 0’ 20 4060 30100
Figure 3.14. Attenuation term, e‘3(1‘C°32(°r(i’-l*klll, plotted as a function of the angle

9,. (in degrees) has a maximal value of 1 at 6,. 2 0°, and a value of 0.5 at 9, = 25°.

81

The pose-oriented threshold for the face score is empirically determined and used for
removing false positives (0.16 for near-frontal views and 0.13 for half-profile views).
The face pose (frontal vs. profile) is estimated by comparing the distances from each

of the two eyes to the major axis of the fitted ellipse.

3.4 Experimental Results

We have evaluated our algorithm on several face image databases, including family
and news photo collections. Face databases designed for face recognition, includ-
ing the FERET face database [28], usually contain grayscale mugshot-style images,
therefore, in our opinion, are not suitable for evaluating face detection algorithms.
Most of the commonly used databases for face detection, including the Carnegie Mel-
lon University (CMU) database, contain grayscale images only. Therefore, we have
constructed our databases for face detection from MPEG7 videos, the World Wide
Web, and personal photo collections. These color images have been taken taken un-
der varying lighting conditions and with complex backgrounds. Further, these images
have substantial variability in quality and they contain multiple faces with variations

in color, position, scale, orientation, 3D pose, and facial expression.

Our algorithm can detect multiple faces of different sizes with a wide range of
facial variations in an image. Further, the algorithm can detect both dark skin-tones
and bright skin-tones because of the nonlinear transformation of the Cb — Cr color
space. All the algorithmic parameters in our face detector have been empirically de-
termined; same parameter values have been used for all the test images. Figure

82

 

 

(e) (0
Figure 3.15. Face detection examples containing dark skin-tone faces. Each example
contains an input image, grouped skin regions shown in pseudo color, and a lighting-
compensated image overlaid with detected face and facial features.

3.15 demonstrates that our algorithm can successfully detect dark skin faces. Figure
3.16 shows the results for subjects with some facial variations (e.g., closed eyes or
open mouth). Figure 3.17 shows detected faces for subjects who are wearing glasses.
The eye glasses can break up the detected skin tone components of a face into smaller
components, and cause reflections around the eyes. Figure 3.18 shows that the pro—
posed algorithm is not sensitive to the presence of facial hair (moustache and beard).
Figure 3.19 demonstrates that our algorithm can detect non-frontal faces as long as

the eyes and mouth are visible in half—profile views.

A summary of the detection results (including the number of false positives, de—

83

 

 

(C) (6
Figure 3.16. Face detection results on closed-eye or open-mouth faces. Each example

contains an original image (top) and a lighting-compensated image (bottom) overlaid
with face detection results.

 

 

(a) (b) (C) (d) (6)
Figure 3.17. Face detection results in the presence of eye glasses. Each example

contains an original image (top) and a lighting-compensated image (bottom) overlaid
with face detection results.

tection rates, and average CPU time for processing an image) on the HHI MPEG7
image database [15] and the Champion database [171] are presented in Tables 3.1 and
3.2, respectively. Note that the detection rate depends on the database. The HHI
image database contains 206 images, each of size 640 X 480 pixels. Subjects in the
HHI image database belong to several racial groups. Lighting conditions (including
overhead lights and side lights) change from one image to another. Further, these
images contain frontal, near—frontal, half-profile, and profile face views of different
sizes. A detected face is a correct detection if the detected locations of the eyes,

84

 

 

((1)

Figure 3.18. Face detection results for subjects with facial hair. Each example
contains an original image (top) and a lighting-compensated image (bottom) overlaid
with face detection results.

the mouth, and the ellipse bounding a human face are found with a small amount
of tolerance, otherwise it is called a false positive. The detection rate is computed
by the ratio of the number of correct detections in a gallery to that of all human
faces in the gallery. Figure 3.20(a) shows a subset of the HHI images. The detec-
tion results of our algorithm are shown in three stages. In the first stage, we show
the skin-tone regions (Fig. 3.20(b)) using pseudo-color; different colors correspond
to different skin-tone groups. In the second stage, we fuse bounding rectangles that
have significant overlapping areas with neighboring rectangles (Fig. 3.20(c)). Each
bounding rectangle indicates a face candidate. In the third stage, we locally detect
facial features for each face candidate. Figure 3.20(d) shows the final detection re-
sults after these three stages. The detected faces are depicted by yellow-blue ellipses,
and the detected facial features (eyes and mouth) are connected by a triangle. The
detection rates and the number of false positives for different poses are summarized
in Table 3.1. The detection rate after the first two stages is about 97% for all poses.
After the third stage, the detection rate decreases to 89.40% for frontal faces, and to

85

 

 

(f) (g) (h) (i) (I)
Figure 3.19. Face detection results on half-profile faces. Each example contains an
original image (top) and a lightingcompensated image (bottom) overlaid with face
detection results.

90.74% for near—frontal faces, and to 74.67% for half—profile faces. The reason for this
decrease in detection rate is the removal of those faces in which the eyes/ mouth are
not visible. However, we can see that the number of false positives is dramatically
reduced from 9, 406 after the skin grouping stage to just 27 after the feature detection

stage for the whole database containing 206 images.

The Champion database was collected from the Internet, and contains 227 com-
pressed images which are approximately 150 X 220 pixels in size. Because most of the
images in this database are captured in frontal and near-frontal views, we present a
single detection rate for all poses in Table 3.2. The detection rate for the first two
stages is about 99.12%. After the third stage, the detection rate decreases to 91.63%.

86

The number of false positives is also dramatically reduced from 5,582 to 14. We
present face detection results on a subset of the Champion database in Fig. 3.21.
Figure 3.22 shows the detection results on a collection of family photos (total of 55
images). Figure 3.23 shows results on a subset of news photos (total of 327 images)
downloaded from the Yahoo news site [172]. As expected, detecting faces in family
group and news pictures is more challenging, but our algorithm is able to perform
quite well on these images. Detection rate on the collection of 382 family and news
photos (1.79 faces per image) is 80.35%, and the false positive rate (the ratio of the
number of false positives to the number of true faces) is 10.41%. More results are

available at http: //www.cse.msu.edu/ ~hsureinl/facloc/ .

87

 

Table 3.1
DETECTION RESULTS ON THE HHI IMAGE DATABASE (IMAGE SIZE 640 x 480) ON
A PC WITH 1.7 GHz CPU. FP: FALSE POSITIVES, DR: DETECTION RATE.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

DETECTION RESULTS ON THE CHAMPION DATABASE (IMAGE SIZE ~ 150 x 220)
ON A PC WITH 860 MHz CPU. FP: FALSE POSITIVES, DR: DETECTION RATE.

Head Pose Frontal Near- Half- Profile Total
Frontal Profile
NO. of images 66 54 75 11 206
Stage 1: Grouped skin regions
No. of PP I] 3145 2203 3781 277 9406
DR (%) H 95.45 98.15 96.00 100 96.60
Time (sec): average :1: s. d. H 1.56 :1: 0.45
Stage 2: Rectangle merge

No. ofFP I] 468 287 582 39 1376
DR (%) l] 95.45 98.15 96.00 100 96.60
Time (sec): average :1: s. d. I] 0.18 :i: 0.23

Stage 3: Facial feature detection
No. of FF 4 6 14 3 27
DR (%) 89.40 90.74 74.67 18.18 80.58
Time (sec): average :t s. d. 22.97 :1: 17.35

Table 3.2

 

 

 

 

 

 

 

 

 

 

Stage 1 l 2 I 3

No. of images 227

No. of PP 5582 382 14

DR (%) 99.12 99.12 91.63

Time (sec): average 1’: s. d. 0.080 :1: 0.036 0.012 :t 0.020 5.780 :l: 4.980

 

 

88

.- 1“

 

ll

‘11

«Ill

15

:5
Ja‘
'

 

Figure 3.20. Face detection results on a subset of the HHI database: (a) input
images; (b) grouped skin regions; (0) face candidates; (d) detected faces are overlaid
on the lighting—compensated images.

89

 

. a...
15‘!“

 

(a) (b)
Figure 3.21. Face detection results on a subset of the Champion database: (a) input
images; (b) grouped skin regions: (0) face candidates; (d) detected faces are overlaid
on the light.ing—compensated images.

90

 

Figure 3.22. Face detection results on a subset of eleven family photos. Each im-
age contains multiple human faces. The detected faces are overlaid on the color-
compensated images. False negatives are due to extreme lighting conditions and
Shadows. Notice the difference between the input and color—compensated images in
terms of color balance. The bias color in the original images has been compensated
in the resultant images.

91

 

Figure 3.22. (Cont’d).

92

M“
.g—«L

. . mi»...
. v. . . . . M:
. i . .
p H“... kt}... .

.2:

.m
.
_
u
«.

{nu—.22..

a...

:2

22...,

 

Figure 3.22. (Cont’d).

93

 

Figure 3.23. Face detection results on a subset of 24 news photos. The detected faces
are overlaid on the color—compensated images. False negatives are due to extreme
lighting conditions, shadows, and low image quality (i.e., high compression rate).

94

3.5 Summary

We have presented a face detection algorithm for color images using a skin—tone color
model and facial features. Our method first corrects the color bias by a lighting
compensation technique that automatically estimates the reference white pixels. We
overcome the difficulty of detecting the low-luma and high-luma skin tones by ap-
plying a nonlinear transform to the YCbC'r color space. Our method detects skin
regions over the entire image, and then generates face candidates based on the spatial
arrangement of these skin patches. It then constructs eye, mouth, and boundary
maps for detecting the eyes, mouth, and face boundary, respectively. The face can-
didates are further verified by the presence Of these facial features. Detection results
on several photo collections have been demonstrated. Our goal is to design a system
that detects faces and facial features, allows users to edit detected faces (via the user
interface shown in Fig. 3.24), and uses these detected facial features as indices for

identification and for retrieval from image and video databases.

95

.4"

 

   
 

ChoosaDisplay hhflwd
‘W'lfifi"" LE

    

 

.- 7le Images» 7 i]

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

 

(b)
Figure 3.24. Graphical user interface (GUI) for face editing: (a) detection mode; (b)
editing mode.

96

Chapter 4

‘r
. '0' 'L
.5

W:
1,.

Face Modeling

We first introduce an overview Of our modeling method [173], and describe the generic
face model and facial measurements. Then we present an approach for adapting the
generic model to the facial measurements. Finally, an adapted 3D face model Of an
individual is texture-mapped and reproduced at different viewpoints for visualization

and recognition.

4. 1 Modeling Method

For efficiency, we construct a 3D model of a human face from a priori knowledge (a
generic face model) of the geometry Of the human face. The generic face model is a
triangular mesh, whose vertices can precisely specify facial features that are crucial for
recognition, such as eyebrows, eyes, nose, mouth, and face boundary. We call these
features recognition-oriented features. The locations and associated properties Of
these recognition-oriented features are extracted from color texture and range data (or

97

Generic

 

   

 

 

 

 

 

'5 Model 40
‘ Local Adaptation .120 40° x

- Local Ali nment Face
IGloba;AlIgnment1-bll 9 I -) Model

| Local Refinement I

 

 

Facial
Measu rements

 

 

    

Figure 4.1. The system overview of the proposed modeling method based on a 3D
generic face model.

disparity maps) obtained for an individual. The generic face model is modified so that
these recognition—oriented features are fitted to the individual’s facial geometry. The
modeling process aligns and adapts the generic face model to the facial measurements
in a global-to—local fashion. The overview of our face modeling method is given in
Fig. 4.1. The input to the modeling algorithm is the generic face model and the
facial measurements. The modeling method contains two major modules: (i) global
alignment and (ii) local adaptation. The global alignment module changes the size Of
the generic face model, and aligns the scaled generic model according to the 3D head
pose. The local adaptation module refines the facial features of the globally aligned
generic face model iteratively and locally. We do not extract isosurfaces directly from
facial measurements because facial measurements are often noisy (e.g., near the ears

98

and nose in frontal views), and because the extraction is tiIIIe-consuming and usually
generates triangles of the same Size in the mesh. Hence, in our model construction, the
desired recognition-oriented facial features can be Specified and gradually modified
in the 3D generic face model. The modeling algorithm generates an adapted/ learned
3D face model with aligned facial texture. The 2D projections of the texture—mapped

3D model are further used for face verification and recognition.

4.2 Generic Face Model

We choose Waters’ animation model [69], which contains 256 vertices and 441 facets
for one half Of the face, because this model captures most of the facial features that are
needed for face recognition (as well as animation), and because triangular meshes are
suitable for free-form surfaces like faces [136]. Figure 4.2 Shows the frontal and one Side
view Of the model, and facial features such as eyes, nose, mouth, face boundary, and
chin. There are openings at both the eyes and the mouth, which can be manipulated.

The Phong-shaded appearance of this model is shown for three different views in

Fig. 4.3.

99

 

 

L‘AVflllYAVAI»
v- " v“
:e‘x‘V/Ie'

—20246

‘1:—

\
Q1-3‘t'e , /
\firfizazz': .
*gaéb‘

—4

—6

 

 

—8

 

   

Figure 4.2. 3D triangular—mesh model and its feature components: (a) the frontal
view; (b) a side view; (c) feature components.

 

Figure 4.3. Phong—shaded 3D model shown at three viewpoints. Illumination is in
front of the face model.

100

4.3 Facial Measurements

Facial measurements Should include information about face shape and facial texture.
3D shape information can be derived from a stereo pair, a collection of frames in a
video sequence, or Shape from shading. It can also be Obtained directly from range
data. We use the range database Of human faces [174], which was acquired using a
Minolta Vivid 700 digitizer. The digitizer generates a registered 200 x 200 range map
and a 400 x 400 color image for each acquisition. Figure 4.4 shows a color image
and a range map of a frontal view, and the texture-mapped appearance from three
different views. The locations Of face and facial features such as eyes and mouth in
the color texture image can be detected by the face detection algorithm described in
Chapter 3 [175] (see Fig. 4.5(a)). The corners of eyes, mouth, and nose can be easily
Obtained based on the locations of detected eyes and mouth. Figure 4.5(b) Shows the

detected feature points.

101

 

 

    
 

(a)

 

Figure 4.4. Facial measurements of a human face: (a) color image; (b) range map;
and the range map with texture mapped for (c) a left view; (d) a profile view; (e) a
right view.

 

(a) (b)
Figure 4.5. Facial features overlaid on the color image, (a) obtained from face
detection; (b) generated for face modeling.

102

4.4 Model Construction

Our face modeling process consists of global alignment and local adaptation. Global
alignment first brings the generic model and facial measurements into the same co—
ordinate system. Based on the 3D head pose and the face size, the generic model is
then sealed, rotated, and translated to fit the facial measurements. Figure 4.6 shows

the global alignment results in two different modes. Local adaptation consists of local

 

Figure 4.6. Global alignment of the generic model (in red) to the facial measurements
(in blue): the target mesh is plotted in (a) for a hidden line removal mode for a side
view; (b) for a see-through mode for a profile view.

alignment and local feature refinement. Local alignment involves scaling and trans-
lating model features, such as eyes, nose, mouth, chin and face boundary to fit the
extracted facial features. Local feature refinement makes use of two new techniques—
displacement propagation and 2. 5D active contours—to smooth the face model and to
refine local features. The local alignment and the local refinement of each feature

103

(shown in Fig. 4.2(c)) are followed by displacement (of model vertices) propagation,
in order to blend features in the face model.

Displacement propagation inside a triangular mesh mimics the transmission Of
message packets in computer networks. Let N,- be the number Of vertices that are
connected to a vertex V,, J,- be the set of all the indices of vertices that are connected
to the vertex V,, U}, be the sum of weights (each Of which is the Euclidean distance
between two vertices) on all the vertices that are connected to the vertex V,, and d0-
be the Euclidean distance between the vertex V,- and a vertex Vj. Let AV]- be the
displacement of vertex V], and a be the decay factor, which can be determined by
the face size and the size Of the active facial feature in each coordinate. Eq. (4.1)

computes the contribution of vertex V]- to the displacement Of vertex Vi.

.. _ AV 10:21:: 3'1) e-adlj, NZ- > 1, wi : 236.],- dz]
AVj-ea'i, N,=1,]EJ,-.

In other words, AV”- is computed as the product of the displacement, the weight, and

_—_.._—

l; vkllz i” V], 2

 

A 12
f I, dij\2\\ j/x
’ l ' :Av'l1‘
‘ [’di [1 Q :Vertex
V19: @ : New Location
- :Trlangle Edge

 

 

 

. AV]:Q V?

1

Figure 4.7. Displacement propagation.

104

 

 

a feature—dependent decay factor. Figure 4.7 depicts a small portion of a triangular

mesh network around the vertex V,. The mesh network illustrates the displacement,

AV,,-,, contributed by a vertex V], (in blue), and the displacement, A1452, contributed

by a vertex ij (in red). In this case, the vertex V,- has six neighboring vertices, i.e.,

N,- is 6. The total displacement AV,- of V,- can be obtained by summing up all the

displacements contributed by its neighboring vertices as follows.

av; = 2mg,

jEJi

The displacement will decay during propagation and it will continue for few iterations.

The number of iterations is determined by the number of edge connections from the

current feature to the nearest neighboring feature. In future implementations, we

will include the symmetric property of a face and facial topology in computing this

displacement. Figure 4.8 shows the results of local alignment for the frontal view

after three iterations of displacement propagation.

   

    
  

  

  
  
 

   

   
    
 

  
 

' lllVAVA<

    
    

4r].

   
  

    

    

      

          
               

      
                    
  

 
         
     

    

        

    
 

   

       

 

  

  

   

  

       

  

     

    

   

51]"
.11“ j a» .. ,1V:H 77> I:
6 5° Infill); '23.: 50 ‘it‘l‘fifillla'li; 50 ii dll‘
l . 31““ i w w
, . an < a: amt!
y O-fi Y 0.5““: 4- 1 Y urgflflu
2 gift I.» "er/again.
, 60-501.“; soles-+31% ,7;
>_ ‘ .\\, , Itasggggfl-
-2 w : \ ‘f‘31‘“"°»"" .
(7’! L10 . -10 .10 «Wt/1767;.
—4 “\bgmuky‘; “my" LAW}? \ #3555929?"
ktk‘u‘g 15.lk‘7110 15' “G “(fizlgfl 15' ~ “gs-$17
—6 - . ,. .. -
‘_ _ - 4+4 _ , _ , .20: ‘llly L -200 ' *
4 2 Mg 2 4 20 _20 o X “40 80 .20 o x 40 80 -20 0 X 40 80
(a) (b) (c) (90
Figure 4.8. Local feature alignment and displacement propagation shown for the

frontal view: (a) the input generic model; the model adapted to (b) the left eye; (c)
the nose; ((1) month and chin.

105

Local feature refinement follows local alignment to further adapt the aligned face
model to an individual face by using 2.5D active contours (snakes). we modify Amini
et al.’s [124] 2D snakes for our 3D active contours OII boundaries of facial features.
The active contours are useful for detecting irregular shapes by minimizing the (total)
energy of the shape contour. The total energy, Emmi, consists Of the internal energy
Eim (controlling the geometry Of the contour) and external energy Eext (controlling
the desired shape). We reformulate the energy for our 3D snake as follows. Assume
that an active contour includes a set Of N vertices: {01, - -- ,v,_1,v,,v,+1, - -- ,vN}.

The total energy can be computed by Eq. (4.2).

N
Emmi = 2 mm.) + Ema.» . (4.2)

i=1

The internal energy is listed in Eq. (4.3).

Eint(vi) = (ailvi — ’Uz'—1l2 + Hill/’2'“ — 2117 + vi—1l2) /2, (4-3)

where (1,- controls the distance between vertices, and B,- controls the smoothness of

the contours. The norm term I - I in Eq. (4.3) is determined by parameterized 3D co-

ordinates, not merely 2D coordinates. Therefore, we call these contours 2.5D snakes.

The initial contours needed for fitting the snakes are crucial. Fortunately, they can
be obtained from our generic face model. Another important point for fitting snakes

106

’4'

is to find appropriate external energy maps that contain local maximum/ minimum
at the boundaries of facial features. For the face boundary and the nose, the external
energy is computed by the maximum magnitude of vertical and horizontal gradients
from range maps. These two facial features have steeper borders than others. For
features such as eyes and the mouth, the external energy is obtained by a product

of the magnitude of the luminance gradient and the squared luminance. Figure 4.9

 

 

 

 

 

 

 

 

 

 

5- g.— - J: g 30 -_ --'A¥.‘_L~ z 7
5 10 15 20 X 30 35 45 50 1o 15 20 25x30 35 4o 45 50
(c) (d)
Figure 4.9. Local feature refinement: initial (in blue) and refined (in red) contours

overlaid on the energy maps for (a) the face boundary; (b) the nose; (c) the left eye;
and (d) the mouth.

 

107

 

shows the results of local refinement for the face boundary, the nose, the left eye, and
the mouth.

Although our displacement propagation smoothes non-feature skin regions in the
local adaptation, these skin regions can be further updated if a dense range map is
available. However, based on our experiments, we find that the update of non-feature
skin regions does not make a significant difference except in cheek regions because the
displacement propagation already smoothes the skin regions surrounding each facial
feature. Figure 4.10 shows the overlay of the final adapted face model in red and

the target facial measurements in blue. For a comparison with Fig. 4.4, Fig. 4.11

 

Figure 4.10. The adapted model (in red) overlapping the target measurements (in
blue), plotted (a) in 3D, (b) with colored facets at a profile view

shows the texture—mapped face model. The texture-mapped model is visually similar
to the original face. We further use a. face recognition algorithm [78] to demonstrate
the use of 3D model. The training database contains (i) 504 images captured from
28 subjects and (ii) 15 images of one subject generated from our 3D face model,

108

    
  

 

£60 40-20 0x20 40 60 80100
(a) (b)

   

 

O
to
4°
(D
O
V
o
N
°>-
Z 100‘
:23
120 o‘f 120
100' ‘o 100;
80 :1: 40‘
-20‘*
Z 0‘;
3 X 226:, ,,,,,,
3 so “- 02040 50 8°
“‘7 ‘ .20
80 0-80'6040 Y

"31536
(f)
Figure 4.11. Texture Mapping. (a) The texture—mapped input range image. The

texture-mapped adapted mesh model shown for (b) a frontal view; (d) a left view;
(6) a profile View; (f) a right View.

109

which are shown in the top row in Fig. 4.12. All the 10 test images of the subject
shown in the bottom row in Fig. 4.12 were correctly matched to our face model. This
preliminary matching experiment shows that the proposed 3D face model is quite

useful for recognizing faces at non-frontal views based on the facial appearance.

  

Figure 4.12. Face matching: the top row shows the 15 training images generated
from the 3D model; the bottom row shows 10 test images of the subject captured
from a CCD camera.

4.5 Summary

Face representation plays a crucial role in face recognition systems. For face recog—
nition, we represent a human face as a 3D face model that is learned by adapting
a generic 3D face model to input facial measurements in a global—to—local fashion.
Based on the facial measurements, our model construction method first aligns the
generic model globally, and then aligns and refines each facial feature locally using
displacement (of model vertices) propagation and active contours associated with fa-
cial features. The final texture mapped model is visually similar to the original face.
Initial matching experiments based on the 3D face model show encouraging results

for appearance-based recognition.

110

Chapter 5

Semantic Face Recognition

In this chapter, we will describe semantic face matching (see Fig. 1.6 in Chapter 1)
based on color input images and a generic 3D face model. We will give the details of (i)
the face modeling from a single view (i.e., the frontal view), called face alignment, and
(ii) recognition module in the semantic face matching algorithm. Section 5.1 describes
the concept of semantic facial components, the semantic face graph, the generic 3D
face model, and interacting snakes (multiple snakes that interact with each other).
Section 5.2 describes the coarse alignment between the semantic graph and the input
image based on the results of face detection. Section 5.3 presents the process of fine
alignment of the semantic graph using interacting snakes. we explain how to compute
the matching scores for graph alignment, and then Show the resultant facial sketches
and cartoon faces. Section 5.4 describes a semantic face matching method for recog-
nizing faces, the use of component weights based on alignment scores, and the cost
function for face identification. Then we give the algorithm of the proposed semantic
face matching. We illustrate the generated cartoon faces from aligned semantic face

111

graphs. We demonstrate the experiment results on face matching based on a subset
of the MPEG7 content set [15] and Michigan State University (MSU) face database.
Section 5.5 describes the generation of facial caricatures, and discusses the effects of

caricature on face recognition. A summary is given in Section 5.6.

5.1 Semantic Face Graph as Multiple Snakes

A semantic face graph provides a high-level description of the human face. A seman—
tic graph projected onto a frontal view for face recognition is shown in Fig. 5.1. The
nodes of the graph represent semantic facial components (e.g., eyes, mouth, and hair),
each of which is constructed from a subset of vertices of the 3D generic face model and
is enclosed by parametric curves. A semantic graph is represented in a 3D space and
is compared with other such graphs in a 2D projection space. Therefore, the 2D ap-
pearance of the semantic graph looks different at different viewpoints due to the effect
of perspective projection of the facial surface. We adopt Waters’ animation model
[69], [176] as the generic face model because it contains all the internal facial com-
ponents, face outline, and muscle models for mimicking facial expressions. However,
Waters’ model does not include some of the external facial features, such as ears and
hair. The hairstyle and the face outline play a crucial role in face recognition. Hence,
we have created external facial components such as the ear and the hair contours
for the frontal view of Waters’ model. We hierarchically decompose the vertices of
the mesh model into three levels: (i) vertices at the boundaries of facial components,
(ii) vertices constructing facial components, and (iii) vertices belonging to facial skin

112

 

 

 

 

 

Hair Outline

Eyebrow Eyebrow
Eye Eye

Ear Ear
F Nose

a Mouth 3'
C a
e d

B o u n
(a) (b)
Figure 5.1. Semantic face graph is shown in a frontal view, whose nodes are (a)
indicated by text; (b) depicted by polynomial curves; (c) filled with different shades.
The edges of the semantic graph are implicitly stored in a 3D generic face model and
are hidden here.

 

 

 

 

regions. The vertices at the top level are labelled with facial components such as the
face outline, eyebrows, eyes, nose, and mouth (see Fig. 5.2). Let To denote the set of
all semantic facial components, which are nodes of the generic semantic graph, G0.
That is To = {{left eyebrow}, {right eyebrows}, {left eye},..., {hair boundary}}. Let
T be a subset of To, that is T C 2T0. Let M be the number of facial components
in T. For example, T can be specified as {{left eye}, {right eye}}, {mouth}}, where
M is 3. Let the semantic graph projected on a 2D image, represented by the set
T, be G. The coordinates of component boundary of G can be represented by a
pair of sequences 36,01) and y,(n), where n = 0,1,...,N, — 1 and i = 1, . . . , M, for
component i with N,- vertices. The 1D Fourier transform, a,(k), of the complex signal
u,(n) = ;r,~(n) +jy1-(n) (wherej : \/——1) is computed by
N,—1

a,(k) : .7:{u,:('n)} = Z u,(72) . e-firkn/Ni, (5.1)

n=0

    

I
“1’47 ’l/
n 1

1', I
‘ /
\SZ’I‘ /

     

M \. r

.\ v. r.
‘ §griir2
\ \

   
 

 
    

 

(C)

Figure 5.2. 3D generic face model: (a) Waters’ triangular-mesh model shown in the
side view; (b) model in (a) overlaid with facial curves including hair and ears at a
side view; (c) model in (b) shown in the frontal view.

for facial component 2' with a close boundary such as eyes and mouth, and with end-
vertex padding for those having open boundary such as ears and hair components.
The advantage of using semantic graph descriptors for face matching is that these de-
scriptors can seamlessly encode geometric relationships (scaling, rotation, translation,
and shearing) among facial components in a compact format in the spatial frequency
domain, because the vertices of all the facial components are specified in the same
coordinate system with the origin around the nose (see Fig. 5.2). The reconstruction

of semantic face graphs from semantic graph descriptors is obtained by

L,»—1
em) = f-1{a,(k)} : Z mug) . ej2"k"/N’, (5.2)

k=O
where L,- (< N,) is the number of frequency components used for the 2'“ face com—
ponent. Figure 5.3 shows the reconstructed semantic face graphs at different levels

114

of Fourier series truncation. In addition, the coordinates of component boundary

 

(f) (s) (h) (i) (1')
Figure 5.3. Semantic face graphs for the frontal view are reconstructed using Fourier
descriptors with spatial frequency coefficients increasing from (a) 10% to (j) 100% at
increments of 10%.
of G can also be represented by parametric curves, i.e., C(S) = (a:(s),y(s)), where
s 6 [0,1], for explicit curve deformation or for generating level-set functions for im-

plicit curve evolution. Therefore, the component boundaries of a semantic face graph

are associated with a collection of active contours (snakes).

5.2 Coarse Alignment of Semantic Face Graph

Our face recognition system contains four major modules: face detection, pose esti-
mation, face alignment, and face matching. The face detection module finds locations
of face and facial features in a color image using the algorithm in [175]. Figures 5.4(a)
to 5.4(d) Show input color images and the results of face detection. Currently, we as-

115

 

 

sume that the face images have been captured at near frontal views (i.e., all of internal
and external facial components are visible). The face alignment module makes use of
the face detection results to align a semantic face graph onto the input image. The
face alignment can be decomposed into the coarse and the fine alignment modules.

In the coarse alignment, a semantic face graph at an estimated pose is aligned with

 

 

 

 

 

 

 

 

(a) i V ' (c) i (d)

Figure 5.4. Face detection results: (a) and (c) are input face images of size 640 x 480
from the MPEG7 content set; (b) and (d) are detected faces, each of which is described
by an oval and a triangle.
a face image through the global and local geometric transformation (scaling, rotation,
and translation), based on the detected locations of face and facial components. Sec-
tion 5.3 will describe in detail the fine alignment, in which the semantic face graph
is locally deformed to fit the face image.

Coarse alignment involves a rigid 3D transformation of the entire semantic graph.
The parameters used in the transformation (scaling, rotation, and translation) are
estimated from the outputs of the face detection algorithm. Besides the use of face
detection results, we further employ the edges and color characteristics of facial com-
ponents to locally refine the rotation, translation, and scaling parameters for individ—

ual components. This parameter refinement is achieved by maximizing a semantic
facial score (SFS) through a small amount of perturbations of the parameters. The

Semantic face score takes into account the fitness of component boundary and of com-

116

ponent color. The semantic facial score of the set T on a face image I (u, v), S F ST,
is defined by prior weights on facial components and component matching scores as

follows:

 

SFST _.. xiv 01 wt( > - M30)

— -SDMSz’, 5.3
221611)“) p ( ()) ( )

where N is the number of semantic components, wt(z’) and M S (2) are, respectively,
the a priori weight and the matching score of component i, p is a constant used to
penalize the components with high standard deviations of the matching scores, and

SD(:1:) stands for standard deviation of :r.

The matching score for the z“ facial component is computed based on the coher-
ence of the boundary and the coherence of color content (represented by a component

map) by

1 Mi—l 1 A,——1
M50) 74— X Z em, a.)
2 j=0 ’ k=0
ICOS(910(UJ,’U] "' 6(“3': vjlll + f(ujvvJ)

 

Where M,- and A,- are, respectively, the number of pixels along the curve of component

i and those of pixels covered by the component 2', 6,0 and 6, are the normal direction

Of component curve 2' in a semantic graph G and the gradient orientation of the image

I , f is the edge magnitude of the image I , and e(uk, 12;.) is the facial component map
Of the image I at pixel k. The gradients are computed as follows:

117

S

ijij) = ZIVG03(uj7vj)®Y(uj,vj)l (5.5)
3:0
5'

90‘le = ZargWGaJUnW)®Y(Uj,vj)), (5-6)
3:0

where Y is the luma of the color image I , and Gas is the Gaussian function with zero
mean and standard deviation 0,. The largest standard deviation 03 is limited by the
distance between eyes and eyebrows where S = 4, and V and ® are the gradient and
convolution operators. The gradient magnitude, gradient orientation, eye map [175]
and coarse alignment results for the subject in Fig. 5.4(a) are shown in Fig. 5.5.
The eye map is an average of a symmetry map [177] and an eye energy map (will be
explained in Section 5.3.1). Furthermore, we construct a shadow map of a face image
in order to locate eyebrow, nostril, and mouth lines, based on the average value of
luminance intensity on a facial skin region (i.e., rectangles shown in Figs. 5.6(a) and
5.6(c)). These feature lines, shown as dark lines in Figs. 5.7(c), are used to adjust
corresponding facial components of a semantic graph. Fig. 5.7 shows five examples

of coarse alignment.

5.3 Fine Alignment of Semantic Face Graph via
Interacting Snakes

Fine alignment employs active contours to locally refine facial components of a se-
mantic face graph that is drawn from a 3D generic face model. The 2D projection of

118

 

 

 

Figure 5.5. Boundary map and eye component map for coarse alignment: (a) and
(b) are gradient magnitude and orientation, respectively, obtained from multi—scale
Gaussian-blurred edge response; (c) an eye map extracted from a face image shown
in Fig. 5.4(c); (d) a semantic face graph overlaid on a 3D plot of the eye map; (e)
image overlaid with a coarsely aligned face graph.

a semantic face graph produces a collection of component boundaries, each of which
is described by a closed (or open) active contour. The collection of these active con-
tours, called interacting snakes, interact with each other through a repulsion energy
in order to align the general facial topology onto the sensed face images in an iterative
fashion. We have studied two competing implementations of active contours for the
deformation of interacting snakes: (i) explicit (or parametric) and (ii) implicit contour

119

 

 

 

 

 

 

 

 

 

(C) ((1)
Figure 5.6. Shadow maps: (a) and (c) are luma components of face images in Figs.
5.4(a) and 5.4(c), overlaid with rectangles within which the average values of skin
intensity is calculated; (b) and (d) are shadow maps where bright pixels indicate the
regions that are darker than average skin intensity.

representations. The explicit contour representation has the advantage of maintain-
ing the geometric topology. The implicit contour representation requires topological

constraints on the implicit function.

5.3.1 Interacting Snakes and Energy Functional

Active contours have been successfully used to impose high-level geometrical con-
straints on low—level features that are extracted from images. Active contours are
iteratively deformed based on the initial configuration of the contours and the energy
functional that is to be minimized. The initial configuration of interacting snakes is
obtained from the coarsely-aligned semantic face graph, and is shown in Fig. 5.8(c).
Currently, there are eight snakes interacting with each other. These snakes describe
the hair outline, face outline, eyebrows, eyes, nose, and mouth of a face; they are
denoted as V(s) = U£,{v,(s)}, where N (= 8) is the number of snakes, and v,(s) is
the 2'“ snake with the parameter s E [0, 1].

The energies used for minimization include the internal energy of a contour (i.e.,
smoothness and stiffness energies), and the external energy (i.e., the inverse of edge

120

 

 

Figure 5.7. Coarse alignment: (a) input face images of size 640 x 480 from the
MPEG7 content set (first three rows), and of size 256 x 384 from the MSU database
(the fourth row); (b) detected faces; (c) locations of eyebrow, nostril, and mouth lines
using shadow maps; ((1) face images overlaid with coarsely aligned face graphs.

121

 

Figure 5.8. Interacting snakes: (a) face region extracted from a face image shown
in Fig. 5.4(a); (b) image in (a) overlaid with a (projected) semantic face graph; (c)
the initial configuration of interacting snakes obtained from the semantic face graph
shown in (b).

strength) extracted from an image. In addition to minimizing the internal energy of
an individual curve, interacting snakes minimize the attraction energy on both the

contours and enclosed regions of individual snakes, and the repulsion energy among

multiple snakes. The energy functional used by interacting snakes is described in

 

Eq. (5.7).
N 1
Eisnake : Z / pinteTnal(vi(5)) + Erepulsion(vl(3)l
i=1 0 E‘r.
thOT'
+ Eattraction(vi(5)) d3 7 (57)
_V___/
Eobservaticm

where i is the index of the interacting snake. The first two energy terms are based on
the prior knowledge of snake’s shape and snakes’ configuration (i.e., facial topology)
While the third energy term is based on the sensed image (i.e., observed pixel values).

122

 

In the Baysian framework, given an image I , minimizing the energy of interacting
snakes is equivalent to maximizing a posteriori probability p(V|I) of interacting snakes

V(s) with a 0/1 loss function:

p(I|V) °p(V)
19(1) ’

 

p(VII) = (5.8)

where p(I IV) ~ e"E°b"“°“°", p(V) ~ e‘EPm", p(V) is the prior probability of snakes’
structure and p(I IV) is the conditional probability of the image potential of inter-
acting snakes. From calculus of variations, we know that interacting snakes which
minimize the energy function in Eq. (5.7) must satisfy the following Euler-Lagrange

equation:

2 owls)—621%)—VE..p.z..-o..(v.(s))—VE........-o..(v.(s)) =0.

2:1

 

fir
Internal Force Repulsion Force Attraction Force

(5.9)

where a and [3 are coefficients for adjusting the second- and the fourth- order deriva-
tives of a contour, respectively. Repulsion force field is constructed based on the

gradients of distance map among the interacting snakes as follows:

N
—VE,.ep,,13,-0n(v,-(s)) = v EDT( U vj(s)) , (5.10)
i=1aj7éi
where EDT is a signed Euclidean Distance Transform [178]. Figure 5.9 show the
repulsion force fields for the hair outline and the face outline. The use of the repulsion

force can prevent different active contours from converging to the same locations of

123

 

minimum energy. The attraction force field consists of two kinds of fields in Eq.

9

o
(a)
Figure 5.9. Repulsion force: (a) interacting snakes with index numbers marked; (b)

the repulsion force computed for the hair outline; (c) the repulsion force computed
for the face outline.

 

(5.11): one is obtained from edge strength, called gradient vector field (GVF) [127],

and the other from a region pressure field (RPF) [133].

_VEimage(vi(3)) 2 GVF + RPF
= GVF+p~ N(vi(3)) - (1 _ W),

100
(5.11)

where N(v,~(s)) is the normal vector on the 2'“ contour 11,-(5); Ef‘m" is the component
energy of the 2"" component; a, a are the mean and the standard deviation of region
energy over a seed region of the 2"“ component; It is a constant that constrains the
energy variation of a component. The advantage of using GVF for snake deformation
is that its range of influence is larger than that obtained from gradients, and can
attract snakes to a concave shape. A GVF is constructed from an edge map by an
iterative process. However, the construction of GVF is very sensitive to noise in the

124

 

edge map; hence it requires a clean edge map as an input. Therefore, we compute
a GVF by using three edge maps obtained from luma and chroma components of
a color image, and by choosing as the edge pixels the top 19% (= 15%) of edge
pixel population over a face region, as shown in Fig. 5.10(a). Figure 5.10(b) is

the edge map for constructing its GVF that is shown in Fig. 5.10(c). The region

   

(a)

Figure 5.10. Gradient vector field: (a) face region of interest extracted from a 640x480
image; (b) thresholded gradient map based on the population of edge pixels shown
as dark pixels; (c) gradient vector field.

pressure field is available only for a homogeneous region in the image. However,
we can construct component energy maps that reveal the color property of facial
components such as eyes with bright-and-dark pixels and mouth with red lips. Then
a region pressure field can be calculated based on the component energy map and
on the mean and standard deviation of the energy over seed regions (note that we
know the approximate locations of eyes and mouth). Let a color image have color
components in the RGB space denoted as (R, G, B), and those in YCbCr space as
(Y, Cb, Cr). An eye component energy for a color image is computed as follows:

125

Egg?!) : Emsat + Ecsh + Ecdif: (5.12)

 

E...“ = [((R — é)? + (G — 53E)? + (B “1,92
0.5
(was) 1, .5...)
Ea = [[Cr— K/212 — [Cb—M212]. (5.14)
Ea) = [[Crl — [Obi], (5.15)

where Emsat is the modified saturation (that is the distance in the plane between
a point (R, G, B) and (K/3,K/3,K/3)) where R + G' + B = K, E63,, is chroma
shift, Ea“, is chroma difference, K = 256 is the number of grayscales for each color
component, and [11:] indicates a function that normalizes a: into the interval [0, 1]. The

eye component energies for subjects in Fig. 5.11(a) is shown in Fig. 5.11(b). The

     

(b) (C) ((1)
Figure 5.11. Component energy (darker pixels have stronger energy): (a) face re-
gion of interest; (b) eye component energy; (0) mouth component energy; (d) nose
boundary energy; (e) nose boundary energy shown as a 3D mesh surface.

mouth component energy is computed as E233,

= [—[Cl)] — [er]. Figure 5.11(c)

126

 

shows examples of mouth energies. For the nose component, its GVF is usually weak,
and it is difficult to construct an energy map for nose. Hence, for the nose, we utilize
Tsai and Shah’s shape—from-shading (SFS) algorithm [179] to generate a boundary
energy for augmenting the GVF for the nose component. The illumination direction
used in the SFS algorithm is estimated from the average gradient fields of a face
image [180] within a facial region. Figures 5.11(d) and 5.11(e) show examples of nose

boundary energies in a 2D grayscale image and a 3D mesh plot, respectively.

5.3.2 Parametric Active Contours

Once we obtain the attraction force, we can make use of the implicit finite differential
method [77], [127] and the iteratively updated repulsion force to deform the snakes.
The stopping criteria is based on the iterative movement of each snake. Figure 5.12(a)
shows the initial interacting snakes, Fig. 5.12(b) shows snake deformation without the
eyebrow snakes, and Fig. 5.12(c) shows finely aligned snakes. Component matching
scores in Eq. (5.4) are then updated based on the line and region integrals of boundary
and component energies, respectively. We discuss another approach for deforming the

interacting snakes based on geodesic active contours and level-set functions in Section

5.3.3.

5.3.3 Geodesic Active Contours

As implicit contours, geodesic snakes [181], which employ level-sets functions [182]
are designed for extracting complex geometry. We initialize a level-set function using

127

 

 

 

 

Figure 5.12. Fine alignment: (a) snake deformation shown every five iterations; (b)
aligned snakes (currently six snakes—hairstyle, face-border, eyes, and mouth~are
interacting); (c) gradient vector field overlaid with the aligned snakes.

signed Euclidean distances from interacting snakes with positive values inside facial
components such as hair. eyebrows, eyes, nose, mouth, and an additional neck com-
ponent, 9f , where i is an integer, i E [1, 8]; and with negative values over the facial
skin and background regions, 9;, where j is either 1 or 2 . Different shades are filled
in component regions, Of and (2;, to form a cartoon face, as shown in Fig. 5.13(c).
Because facial components have different region characteristics, we modified Chan et
al.’s approach [130] to take multiple regions and edges into account. The evolution
step for the level-sets function,<I>, is described as follows:

128

8(1) . V<I>
—('§t_ =V<I> [pl (dzv(g—) -— n27“ — ag)

|V<I>|
—Z|1—c,|2+Z|I—c,|2 , (5.16)
i J'
90 2 22
=1————, =101+v1 5.17
g ( max(90)) 90 g( l I) ( )
1 dt dt 0
7': / 74 (5.18)

MAXDIST dt = 0

where ,ul is a constant, [1.2 and a are constants in the interval between 0 and 1, I
is the image color component, c,- and Cj are the average color components of facial

component 2' over region Q? and component j over (2;,

respectively, r is the com-
ponent repulsion, dt is the absolute Euclidean distance map of the face graph, and
MAXDIST is the maximum distance in the image. We further preserve facial topology
using topological numbers and the narrow band implementation of level-set functions

[183]. The preliminary results are shown in Fig. 5.13 with evolution details and in

Fig. 5.14 without the evolution details.

The facial distinctiveness of individuals can be seen from the changes among the
generic, the fine fitted, and fine deformed face templates, shown in Figs. 5.13(a),
5.13(d), and 5.13(e). Comparing the two approaches for deforming interacting
snakes, we believe that the first approach, parametric active contours, is better suited
to the deformation of semantic face graphs.

129

 

 

 

 

 

 

(C) (d) (6)
Figure 5.13. Fine alignment with evolution steps: (a) a face image; (b) the face
in (a) overlaid with a coarsely aligned face graph; (c) initial interacting snakes with
different shades in facial components (cartoon face); (d) curve evolution shown every
five iterations (totally 55 iterations); (e) an aligned cartoon face.

5.4 Semantic Face Matching

We have developed a method to automatically derive semantic component weights for
facial components based on coarsely aligned and finely deformed face graphs. These
component weights are used to emphasize salient facial features for recognition (i.e.,
for computing a matching cost for a face comparison using semantic face graphs. The
aligned face graph can also be used for generating facial caricatures.

130

 

a.

@

 

(b) (C) (d) (e) (f)

Figure 5.14. Fine alignment using geodesic active contours: (a) a generic cartoon
face constructed from interacting snakes; (b) to (f) for five different subjects. For
each subject, the image in the first row is the captured face image; the second row
shows semantic face graphs obtained after coarse alignment, and overlaid on the color
image; the third row shows semantic face graphs with individual components shown
in different shades of gray; the last row shows face graphs with individual components
after fine alignment.

131

5.4.1 Component Weights and Matching Cost

After the two phases of face alignment, we can automatically derive a weight (called
semantic component weight) for each facial component 2' for a subject P with Np

training face images by

1 + 8-203(i)/d2(‘i) Np > 1,

P - __
sew (z) — 14.84%“) Np :1, (5.19)
1 N”
. __ . . pk .
d(z)_NP;SFD,(GO,GP,) MS (2), (5.20)
0d(z') =50,c [SFD,-(Go, ka) - MSP“(2')] , (5.21)

where SF D means semantic facial distance, MS is the matching score, SD stands
for standard deviation, Go and ka are the coarsely aligned and finely deformed
semantic face graphs, respectively. The semantic component weights take values
between 1 and 2. The semantic facial distance of facial component 2' between two

graphs is defined as follows

Sim-(Go. GP.) =Dist<SGD.-G°. SGDf'P“)
1 Li G 2 0.5
= 3:]a?°(k)—a.p“(k) , (522)
1 1:20

where SGD stands for semantic graph descriptors. The distinctiveness of a facial
component is evaluated by the semantic facial distance S F D between the generic
semantic face graph and the aligned/ matched semantic graph. The visibility of a
facial component (due to head pose, illumination, and facial shadow) is estimated

132

 

by the reliability of component matching/ alignment (i.e., matching scores for facial
components). Finally, the 2D semantic face graph of subject P can be learned from

Np images under similar pose by
1
Gp = Uf‘l ——§ :SGDGPR . (5.23)

The matching cost between the subject P and the k-th face image of subject Q can

be calculated as

M
C(P, Qk) = Z{scwp(z°) ~ sch“(z') - SFDi(Gp, GQk)}, (5.24)

i=1
where M is the number of facial components. Face matching is accomplished by

minimizing the matching cost.

5.4.2 Face Matching Algorithm

The system diagram of our proposed semantic face recognition method was illustrated
in Fig. 1.6 (in Chapter 1). Figure 5.15 describes the semantic face matching algo-
rithm for identifying faces with no rejection. The inputs of the algorithm are training
images of [if known subjects in the enrollment phase and one query face image of an
unknown subject in the recognition phase. The query input can be easily generalized
to either multiple images of an unknown subject or multiple images of multiple un-
known subjects. Each known subject, P’, can have Nj(2 1) training images. The
output of the algorithm is the identity of the unknown query face image(s) among 111
known subjects (a rejection option can be included by providing a threshold on the

133

 

matching cost in the algorithm). The algorithm uses our face detection method to

locate faces and facial features in all the images, and the coarse and fine alignment

methods to extract semantic facial features for face matching. Finally, it computes a

matching cost for each comparison based on selected facial components, the derived

component weights (distinctiveness), and matching score (visibility).

Figure 5.15. A semantic face matching algorithm.

 

INPUT:

Step 1:
Step 2:
Step 3:

Step 4:

Step 5:

Step 6:
Step 7:

Step 8:

Step 9:

OUTPUT:

- Nj training face images for the subject Pj, j = 1, . . . , M
- one query face image for unknown subject Q

Detect faces for all the images using the method in [175]

—) Generate locations of face and facial features

Form a set of facial components, T, for recognition by assigning prior
component weights

Coarsely align a generic semantic face graph to each image based on T
——> Obtain component matching scores for each graph in Eq. (5.4)
Deform a coarsely-aligned face graph

——> Update component matching scores based on integrals of

component energies

Compute semantic facial descriptors SGD for each graph

using the 1-D Fourier transform in Eq. (5.1).

Compute semantic component weights for each graph in Eqs. (5.19)-(5.21)
Integrate all the face graphs of subject Pj in Eq. (5.23),

resulting in M template face graphs

Compute M matching costs, C(Pj, Qk), between Pi and C2,, in Eq. (5.24),
where k=1,j=1,...,M

Subject PJ with the minimum matching cost has the best matched face
to the unknown subject Qk.

Q=PJ

 

5.4.3 Face Matching

We have constructed a small face database of ten subjects (ten images per subject)

at near frontal views with small amounts of variations in facial expression, face orien—

134

tation, face size, and lighting conditions, during different image capture sessions over
a period of two months. Figure 5.16 shows five images of one subject, while Fig. 5.17

shows one image each of ten subjects.

 

(g) (11)

Figure 5.17. Face images of ten subjects.

We employ 5 images each for 10 subjects for training the semantic face graphs.
With re—substitution and leave-one—out tests. the misclassification rates are shown in
Table 5.1 using different sets of facial components and semantic graph descriptors
with the number of frequency components truncated at three levels. External facial

135

 

 

Table 5.1
ERROR RATES ON A 50-IMAGE DATABASE.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Component Set T2 T3 T4

0 o as as f

l ) if

0
Face Graph .
P (%) RS ] LOO RS WOO RS ] LOO RS ] LOO i ..
100% 0% 6% 0% 6% 12% 24% 16% 30%
50% 0% 6% 0% 6% 12% 24% 16% 30%
30% 0% 6% 0% 12% 16% 24% 18% 34%

 

 

 

 

 

 

P: % of frequency components, T1: All components, T2: External components, T3:
Internal components, T4: Eyes and Eyebrows, RS: Re—substitution, LOO: Leave-one-
out.

 

 

 

 

 

 

 

 

 

Table 5.2
DIMENSIONS OF THE SEMANTIC GRAPH DESCRIPTORS FOR INDIVIDUAL FACIAL
COMPONENTS.

P (%) 100% 50% 30%
Dimension N,- L, L,-
Eyebrow 12 5 3
Eye 13 7 3
Nose 34 13 7
month 14 7 3
Face outline 36 17 11
Ear 11 5 3
Hair 19 9 5

 

 

 

 

 

 

 

 

P: ‘70 of frequency components, N;: the dimension of semantic graph descriptors, L,:
the dimension of truncated descriptors.

136

 

(a) (b) (C) (d)

Figure 5.18. Examples of misclassification: (a) input test image; (b) semantic face
graph of the image in (a); (c) face graph of the misclassified subject; ((1) face graph of
the genuine subject obtained from the other images of the subject in the database (i.e.,
without the input test image in (a)). Each row shows one example of misclassification.

components include face outline, ears, and hairstyle, while internal components are
eyebrows, eyes, nose, and mouth. We can see that the external facial components
play an important role in recognition, and the Fourier descriptors provide compact
features for classification because the dimensionality of our feature space is lower (see
Table 5.2) compared to those used in eigen-subspace methods. Figure 5.18 shows the
three examples of misclassification in a leave—one—out test for the facial component
set T1 using all the frequency components. The false matching may result from the

137

similar configuration of facial components, the biased average facial topology of the
generic face model, and coarse head pose. Figures 5.19, 5.20. and 5.21 show the
reconstructed semantic face graphs, Gp in Eq. (5.23), (compare them with Go in

Fig. 5.1(c)) at three levels of details, respectively. Each coarse alignment and fine

   

(a) (b) (C) (d)
(71" if“ it (7f)
0 O 0 C7
(0 (g) (h) (i)

Figure 5.19. Cartoon faces reconstructed from Fourier descriptors using all the fre-
quency components: (a) to (j) are ten average cartoon faces for ten different subjects
based on five images for each subject. Individual components are shown in different
shades in (a) to (e).

(i)

alignment on an image of size 640 X 480 takes 10 see with C implementation and
460 sec. with MATLAB implementation, repsectively, while each face comparison
takes 0.0029 see with Matlab implementation on a 1.7 GHz CPU. We are conducting
other cross—validation tests for classification, and are in the process of performing
recognition on gallery (containing known subjects) and probe (containing unknown
subject) databases. Although the alignment is off-line currently, there is large room
to enhance the performance of alignment implementation to make it operate in real-

138

 

 

 

(g) (h) (i) (1')

Figure 5.20. Cartoon faces reconstructed from Fourier descriptors using only 50% of
the frequency components: (a) to (j) are ten average cartoon faces for ten different
subjects based on five images for each subject. Individual components are shown in
different shades in (a) to (e).

 

O O
‘8 ‘5’ 28:5 :3 3 8f ° °
O /0 KO /0 O

(f) (g) (h) (i) (1')
Figure 5.21. Cartoon faces reconstructed from Fourier descriptors using only 30% of
the frequency components: (a) to (j) are ten average cartoon faces for ten different
subjects based on five images for each subject. Individual components are shown in
different shades in (a) to (e).

139

time.

5.5 Facial Caricatures for Recognition and Visual-
ization

Facial caricatures are generated based on exaggeration of an individual’s facial dis-
tinctiveness from the average facial topology. Let GE?" represent the face graphs of
caricatures for the subject P, and Go be the face graph of the average facial topology.
Caricatures are generated via the specification of an exaggeration coefficient, k,, in

Eq. (5.25):
G36 = Use-1 (300?? + k,- . (SGDEP — 50030)}. (5.25)

Currently, we use the same value of the coefficient for all the components, i.e., k, = k.
Figure 5.22 shows facial caricatures generated with respect to the average facial topol-
ogy obtained from the 3D generic face model. In Fig. 5.23, facial caricatures are 0p-
timized in the sense that the average facial topology is obtained from the mean facial
topology of training images (total of 50 images for ten subjects). We can see that it is
easier for a human to recognize a known face based on the exaggerated faces. We plan
to quantitatively evaluate the effect of exaggeration of salient facial features on the
performance of a face recognition system. Furthermore, this framework of caricature
generation can be easily employed as an alternative to methods of visualizing high
dimensional data, e.g., Chernoff faces [184].

140

 

(b) (C) (d) (C) (f) (s)
Figure 5.22. Facial caricatures generated based on a generic 3D face model: (a) a
prototype of the semantic face graph, G0, obtained from a generic 3D face model,
with individual components shaded; (b) face images of six different subjects; (c)-(g)
caricatures of faces in (b) (semantic face graphs with individual components shown
in different shades) with different values of exaggeration coefficients, A", ranging from
0.1 to 0.9. 141

I.
5L

 

 

(b) (C) (d) (C) (f) (s)
Figure 5.23. Facial caricatures generated based on the average face of 50 faces (5
for each subject):(a) a prototype of the semantic face graph, G0, obtained from the
mean face of the database, with individual components shaded; (b) face images of
six different subjects; (c)-(g) caricatures of faces in (b) (semantic face graphs with
individual components shown in different shades) with different values of exaggeration

coefficients, k, ranging from 0'1 to 0'9' 142

 

 

 

5.6 Summary

For overcoming variations in pose, illumination, and expression, we propose semantic
face graphs that are extracted from a subset of vertices of a 3D face model, and aligned
to an image for face recognition. We have presented a framework for semantic face
recognition, which is designed to automatically derive weights for facial components
based on their distinctiveness and visibility, and to perform face matching based on
visible facial components. Face alignment is a crucial module for face matching,
and we implement it in a coarse-to—fine fashion. We have shown examples of coarse
alignment, and have investigated two deformation approaches for fine alignment of
semantic face graphs using interacting snakes. Experimental results show that a
successful interaction among multiple snakes associated with facial components makes
the semantic face graph a useful model to represent faces (e.g., cartoon faces and
caricatures) for recognition.

Our automatic scheme for aligning faces uses interacting snakes for various facial
components, including the hair outline, face outline, eyes, nose, and mouth. We
are currently adding snakes for eyebrows to completely automate the whole process
of face alignment. We plan to test the proposed semantic face matching algorithm
on standard face databases. We also plan to implement a pose estimation module
based on the alignment results in order to construct an automated pose-invariant face

recognition system.

143

Chapter 6

Conclusions and Future Directions

We will give conclusions and describe future research directions in the following two

sections, respectively.

6. 1 Conclusions

Face detection as well as recognition are challenging problems and there is still a lot of
work that needs to be done in this area. Over the past ten years, face recognition has
received substantial attention from researchers in biometrics, pattern recognition,
computer vision, and cognitive psychology communities. This common interest in
facial recognition technology among researchers working in diverse fields is motivated
both by our remarkable ability to recognize people and by the increased attention
being devoted to security applications. Applications of face recognition can be found
in security, tracking, multimedia, and entertainment domains. We have proposed two
paradigms to advance face recognition technology. Three major tasks involved in

144

 

 

such vision-based systems are (i) detection of human faces, (ii) construction of face
models/ representations for recognition, and (iii) identification of human faces.

Detection of human faces is the first step in our proposed system. It is also the
initial step in other applications such as video surveillance, design of human computer
interface, face recognition, and face database management. We have proposed a face
detection algorithm for color images in the presence of various lighting conditions
as well as complex backgrounds. Our detection method first corrects the color bias
by a lighting compensation technique that automatically estimates the statistics of
reference white for color correction. We overcame the difficulty of detecting the low-
luma and high-luma skin tones by applying a nonlinear transformation to the YCbCr
color space. Our method detects skin regions over the entire image, and then generates
face candidates based on the spatial arrangement of these skin patches. Next, the
algorithm constructs eye, mouth, and face boundary maps for verifying each face
candidate. Experimental results have demonstrated successful detection of multiple
faces of different size, color, position, scale, orientation, 3D pose, and expression in
several photo collections.

Construction of face models is closely coupled with recognition of human faces,
because the choice of internal representations of human faces greatly affects the design
of the face matching or classification algorithm. 3D face models can help augmenting
the training face databases used by the appearance-based face recognition approaches
to allow for recognition under illumination and head pose variations. For recognition,
We have designed two methods for modeling human faces based on a generic 3D face
model. One requires individual facial measurements of shape and texture (i.e., color

145

images with registered range data) captured in the frontal View; the other takes only
color images as its facial measurements. Both modeling methods adapt facial features
of a generic model to those extracted from an individual’s facial measurements in a
global—to—local fashion. The first method aligns the model globally, uses the 2.5D
active contours to refine feature boundaries, and propagates displacements of model
vertices iteratively to smooth non-feature areas. The resulting face model is visually
similar to the true face. The resulting 3D model has been shown to be quite useful for

recognizing non-frontal views based on an appearance-based recognition algorithm.

The second modeling method aligns semantic facial components, e. g., eyes, mouth,
nose, and the face outline, of the generic semantic face graph onto those in a color face
image. The nodes of a semantic face graph, derived from a generic 3D face model,
represent high-level facial components, and are connected by triangular meshes. The
semantic face graph is first coarsely aligned to the locations of detected face and facial
components, and then finely adapted to the face image using interacting snakes,
each of which describes a semantic component. A successful interaction of these
multiple snakes results in appropriate component weights based on distinctiveness
and visibility of individual components. Aligned facial components are transformed
to a feature space spanned by Fourier descriptors for semantic face matching. The
semantic face graph allows face matching based on selected facial components, and
updating of a 3D face model based on 2D images. The results of face matching
demonstrate the classification and visualization (e.g., the generation of cartoon faces
and facial caricatures) of human faces using the derived semantic face graphs.

146

 

6.2 Future Directions

Based on the two recognition paradigms proposed and implemented in this thesis,
we can extend our work on face detection, modeling and recognition in the following

manner:

6.2.1 Face Detection 85 Tracking

The face detection module can be further improved by

0 Optimizing the implementation for real-time applications;

0 Combining the global (appearance-based) approach and a modified version of
our analytic (feature-based) approach for detecting faces in profile views, in

blurred images, and in images captured at a long distance;

0 Fusing a head (or body) detector in grayscale images and our skin filter in color
images for locating non-skin-tone faces (e.g., faces in gray-scale images or faces

taken under extreme lighting conditions).

In order to make the face detection module useful for face tracking, we need to include

motion detection and prediction submodules as follows.

0 Parametric face descriptors: Face ellipses and eye-mouth triangles are useful
measurements that can be used for the motion prediction of human faces. We
are currently developing a tracking system that combines temporal and shape
(derived from our face detection method) information.

147

 

 

 

 

o A face tracking and recognition prototype: A prototype of a recognition

system with tracking modules is shown in Fig. 6.1. The detection and tracking

 

 

 

 

v O
IMAGES COLOR MOTION
NORMALIZA TION DETECTION -

DETECTION OF *
SKIN TONE

 

DETECTION OF DETECTION OF
FA CE CANDIDATES HUMANt BODIES

 

ETECTION OF FA CIAj I DETECTION OF I

 

 

 

 

 

 

 

 

 

 

COMPONENTS HUMAN GAITS
* MOTION PREDICTION OF
FACES AND BODIES .
I 1‘ Tracked
Faces
[ALIGNMENT OF MATCHING OF
FACIAL MODELS FACIAL MODELS
DATABASE ’ 4 ~

 

 

Figure 6.1. A prototype of a face identification system with the tracking function.

modules include (i) a motion filter if temporal information is available, (ii) a
human body detector and analysis of human gait, and (iii) a motion predictor

of face and the human body.

0 Preliminary tracking results: An example of detection of motion and skin
color is shown in Figure 6.2 (see [185] for more details). The preliminary results
of off-line face tracking based on the detection of interframe difference, skin
color, and facial features are shown in Fig. 6.3, which contains a sequence
of 25 video frames. These images are lighting-compensated and overlaid with
detected faces. The image sequence shows two subjects entering the PRIP Lab

148

 

 

 

 

 

 

 

 

(a) (b)
Figure 6.2. An example of motion detection in a video frame: (a) A color video frame;
(b) extracted regions with significant motion; (c) detected moving skin patches shown
in pseudocolor; (d) extracted face candidates described by rectangles.

in the Engineering building at Michigan State University through two different
doors. Faces of different sizes and poses are successfully tracked under various

lighting conditions.

6.2.2 Face Modeling

We have developed two face modeling methods for range (with registered color) data
and for color data in a frontal view. Once we construct a pose estimator, we can
modify the proposed methods of face alignment for non-frontal views. This extension

of modeling includes the following tasks:

0 Complete head and ear mesh models: We need model polygons for
hair/ head portion and ears in order to generate hair and ear outlines in non—

frontal views.

0 Pose and illumination estimation: We can design a head pose estimator
and an illumination estimator for faces in frontal and non—frontal views based
on the locations of face and facial components, and shadows/shadings on the
face.

149

 

woman: wmpemaomfioomfifiwz 23 no @8785 v.8 $0.3 vouoopom .Emi 8 d8
89a was Eoflon 8 Q8 89a nowadays 2d women: Sonar .88.me 83> mm mo oonoscwm e no mfismoa wfixoafi womb .mb Emmi

.;.r3o:un. _
Ill-n“ . .

 

150

o Non-frontal training views: According to the estimated head pose, we can
rotate the generic face model and generate the boundary curves of the semantic

components for face alignment at the estimated pose.

6.2.3 Face matching

We have designed a semantic face matching algorithm based on the component
weights derived from distinctiveness and visibility of individual facial components.
Currently, the semantic graph descriptors, SGD; in Section 5.4.1, used for compar-
ing the difference between facial components contain only the shape information (i.e.,
component contours). We can improve the performance of the algorithm by including

the following properties:

0 Texture information: Associate a semantic graph descriptor with a set of
texture information (e.g., wavelet coefficients, photometric sketches [55], and
normalized color values) for each facial component. The semantic face matching

algorithm will compare faces based on both the shape and texture information.

o Scalability: Evaluate the matching algorithm on several public domain face

databases.

0 Caricature effects on recognition: Explore other weighting functions on the
distinctiveness of individual facial components based on the visualized facial

caricature and the recognition performance.

0 Facial statistics: Analyze face shape, race, sex, and age, and construct other

151

 

semantic parameters for face recognition, based on a large face database.

152

 

Appendix A

Transformation of Color Space

In this appendix, we will give the detailed formulae of two types of colorspace trans—
formations and an elliptical skins classifier, which are used in our face detection
algorithm. The transformations include a linear transformation between RGB and
YCbCr color spaces and a nonlinear transformation applied to YCbCr for compen-
sating the luma dependency. The skin classifier is described by an elliptical region,

which lies in the nonlinearly transformed YCbCr space.

A.1 Linear Transformation

Our face detection algorithm utilizes a linear transformation to convert the color
components of an input image in the RGB space into those in the YCbCr space for
separating the luma component from chroma components of the input image. The
transformation between these two space is formulated in Eqs. (A1) and (A2) for the
value of color components that range from 0 to 255 (see the details in [155]).

153

 

Y 65 129.057 25.064 R 16

1
Cb = % —37.945 —74.494 112.439 G + 128 (A.1)
Cr 112.439 —94.154 —18.285 B 128
R 1 298.082 0 408583 Y
G : 2—56 298.082 ~100.291 —208.120 Cb (A.2)
B 298.082 516.411 0 Cr

Figures A.1 (a) and (b) illustrate a set of randomly sampled reproducible colors in

the RGB space and its corresponding set in the YCbCr space.

 

(a) (b)
Figure A.1. Color spaces: (a) RGB; (b) YCbCr.

A.2 Nonlinear Transformation

In the YCbCr color space, we can regard the chroma (Cb and C,) as functions of the
luma (Y): Cb(Y) and C,(Y). Let the transformed chroma be C)’,(Y) and C;(Y). The
nonlinear transformation converts the elongated cluster into a cylinder-like shape,

154

based on the skin cluster model obtained from a subset of the HHI database. The
model is specified by the centers (denoted as GAY) and E(Y)) and widths of the
cluster (denoted as W'cb(Y) and W'cr(Y)) (See Fig. 3.5). The following equations

describe how this transformation is computed.

 

(0,-(Y) — 'Ci-(Y)) - ”2:8,, + ‘7-(Kh) if Y < K, or Kh < Y,

 

 

0: (Y) —
C,(Y) if Y E [K1,Kh],
(A.3)
WLc, + <Y“Ym:‘;gfi‘;5::WLC) if Y < K),
Wc,(Y) = (A4)
WHc, + ”mar—YEXEWHC) if K), < Y,

 

_ 108 + (K'"}{’)'_(,‘}?‘1°8) if Y < K),
Cb(Y) = m'" (A5)

108 + (Y‘Qlfilfi.h‘1°8l if K, < Y,

 

154 — “13:35,:31342 if Y < K),

57W) = (A6)

154 + (”50:13:“) if K, < Y,

 

 

where C, in Eqs. (A3) and (A.4) is either Cb 0r Cr, W'cb = 46.97, I'Vch = 23,
)1”ch = 14, We, 2 38.76, l/VLcr = 20, l/VHcr = 10, K1 2 125, and K), = 188. All
values are estimated from training samples of skin patches on a subset of the [HHI
images. Ymm and Yum, in the YCbCr color space are 16 and 235, respectively. Note
that the boundaries of the cluster are described by two curves -C_,(Y) i IVc,(Y) / 2,

155

 

 

and are shown as blue-dashed lines in Fig. 3.5(a) for Cb and in Fig. 3.5(b) for C,.

A.3 Skin Classifier

The elliptical model for the skin tones in the transformed Cg-C; space is described in

Eqs. (A7) and (A8), and is depicted in Fig. 3.6.

 

 

(:1: — ecrr)2 + (y - ecy)2 _

a2 b2 — 1, (A.7)
]- -l [- T F "
r cos 0 sin 0 0,; — ca:
2 , (A.8)
y —sin6 cosbl C,’. — cy

 

 

 

 

 

 

where ex = 109.38, cy = 152.02, 0 = 2.53 (in radian), ecu: = 1.60, ecry = 2.41,
a = 25.39, and b = 14.03. These values are computed from the skin cluster in the

Cg—C; space at 1% of outliers.

156

Appendix B

Distance between Skin Patches

Facial skin areas are usually segmented / split into several clusters of skin patches due
to the presence of facial hair, glasses, and shadows. An important issue here is how
to group / merge these skin regions based on the spatial distance between them. Since
the clusters have irregular shapes, both the Bhattacharrya distance for a generalized
Gaussian distribution and the distance based on the circular approximation of the
cluster areas do not result in a satisfactory merging. Hence, we combine three types

of cluster radii (circular, projection, and elliptical) in order to compute an effective

157

 

 

radius of a cluster ‘i’ w.r.t. another cluster ‘j ’ as follows.

 

R: : mum?) R?) + k ' Rf, (B.1)
Rf = a,|cos(6,j)| , (B2)
1 1/2
.3 2 B.
R; (0052(6ij)/a,2 + Sin2(9Ij)/bz2) a ( 3)
Rf : (Ni/”)1/2 I (13.4)

 

where R,- is the effective radius of the cluster i; R? is its projection radius; Rf is
its elliptical radius; Rf is the circular radius used in [84]; the constant k (equals
0.1) is used to prevent the effective radius from vanishing when two clusters are thin
and parallel; a,- and b,- are the lengths of the major and minor axes of the cluster i,
respectively; 6,-3- is the angle between the major axis of the cluster i and the segment
connecting the centroids of clusters i and j; and N,- is the area of the cluster i. The
major and minor axes of the cluster 2' are estimated by the eigen—decomposition of

the covariance matrix

_ a: 011/
C — , (B5)
5,, a;

 

where 01, (Ty, and 01,, are the second-order central moments of the skin cluster 2'. The

eigenvalues of the covariance matrix C and the lengths of the major and minor axes

158

of the cluster are computed by Eqs. (B.6)-(B.10).

(B.6)

a,:=

 

 

A1 = g. (a: + 0.3 + \/(03, — 03)2 + 40%,!) , (B8)
A2 : g. (of, + a: — (K03 -— 0,3)2 + 403%,) , (B9)

_ 2
a = tan—1(i—0E), (B.10)

where a,- and b,- are the estimated lengths of the major and minor axes of the cluster,
respectively; A1 and A2 are the largest and smallest eigenvalues of the covariance
matrix C, respectively; and a is the orientation of the major axis of the cluster i.
The orientation 0 is used to calculate the angle 922' in Eq. (B.2). Therefore, the
distance between clusters i and j is computed as dij = d — R,- — Rj, where d is the

Euclidean distance between the centroids of these two clusters.

159

 

i.‘

 

Appendix C

Image Processing Template

Library (IPTL)

The face detection algorithm has been implemented using our Image Processing
Template Library (IPTL). This library brings the abstract data class, Image, from
a class level to a container (a class template) level, Image Template. It has the
advantages of easy conversion between images of different pixel types, a high reuse
rate of image processing algorithms, and a better user interface for manipulating data

in the image class level.

C.1 Image and Image Template

An image captures a scene. It is represented by a two—dimensional array of picture
elements (so called pixels) in the digital format. Pixels can be of different data types,
e.g., one bit for binary images, one byte or word for grayscale images, three bytes

160

for true-color images, etc. An image is a concrete object; however, Image can be
regarded as an abstract data class/ type in the field of image processing (IP) and
computer vision (CV). Contemporary IP libraries, including the Intel IPL [186], the
Intel Open CV [187], and the CVIPTool [188], have designed Image classes using
different pixel depths in bits. Our Image Processing Template Library boosts this
abstract Image class from a class level to a class template level. The image template,
ImageT, is designed based on various pixel classes. For example, pixel classes such
as one-bit Boolean, one-byte grayscale, two-byte grayscale, and three-byte true color
are the arguments of the image class template. Hence, the conversion between images
of different pixel classes is performed at the pixel level, not at the class level. Hence,
a large number of algorithms can be reused for images belonging to different pixel
classes.

Figure C.1 shows the architecture of IPTL class templates. The software architec-
ture can be decomposed into five major levels: platform, base, pixel, image/volume,
and movie/space levels. At the platform level, declarations and constants for different
working platforms, e.g., the Microsoft Windows and the Sun Unix, are specified in
the header file ipt|.workingenv.h. At the base level, constants for image processing
are defined in the header file ipt|.base.h, and space-domain classes for manipulation
of different coordinate systems and time—domain classes for evaluation of CPU speed
are defined in ipt|.geom.h and iptl.time.h, rcsrmctively. At the pixel level, the pixel
classes such as GRAYS, GRAYIG, and ColorRGB24 are defined in ipt|.pixe|.h.
At the image level, the image template ImageT is defined based on its argument
of pixel classes. The IPTL is also designed for a. volume template, VolumeT, by

161

 

 

A

considering pixel classes as voxel classes. At the movie level, we can derive templates
for color images (e.g., ImageRGB and ImageYCbCr) for slices, movies, slides,

image display, and image analysis based on the image template. Similarly, based on

 

iptl.workingenv.h
l
I
I I

 

ipt1.base.h iptl.geom.h, iptl.time.h
| ipt1.vector.h
l iptI.matrix.h

iptl.pixel.h
I
I I

 

 

 

iptl.image.h iptl.volume.h

I I

I |

I I I l I

l I I I I

iptl. iptl. iptl. iptl. iptl.

slice h slide.h imagechnl.h space.h scene.h

I I (4D) I

I I I
iptl. iptl. ipt1.volview h

movie.h imagevicw.h I

I I

I l
iptl.imageanal.h iptl.volanal.h

 

Figure C.1. Architecture of IPTL class templates.

the volume template, we can obtain higher level templates for space, scene, volume

display, and volume analysis at the space level.

162

 

C.2 Example Code

Example code of the user interface for the pixel classes and the image template is
given below.

#include " ipt|.imagechnl.h"
int main(){

// Conversion of pixel types

//

GREY8 graypix = 10;
FLOAT32 flpix = 3.5;
ColorRGB colorpix(42,53,64);

colorpix = graypix; // Assign a gray value
graypix = flpix; // Truncate data

colorpix.r = 100; // Change the red component
graypix = colorpix; // Compute luminance

flpix += 20.7; // Arithmetic operations

// Image manipulation
//

// Image Creation

ImageT<GREY8> gray8imageA(HOSTRAM, 128, 128. GREY8(55));
ImageT<GREY8> gray8imageB(HOSTRAM, 128, 128, GREY8(100));
ImageT<GREY16> gray16imageC; // An empty image

/ / Creation of color images
//Data arrangement of the image is RGB... RGB...
ImageT<CoIorRGBZ4> rgbimage(HOSTRAM,32,32,128);

//Data arrangement of the image is RRR...GGG...BBB...
lmageRGB<GREY8> rgbimagechnl(rgbimage1);

// A template function for converting images from one type to another

//

gray8imageA = rgbimage; // Extract Luminance

gray16imageC = gray8imageB; // Enlarge the dynamic range of gray values
gray8imageB -= gray8imageA; // Image subtraction

gray8imageA[5] = 100; // Assess pixels as a 1D vector

163

 

gray8imageA(120,120) = 200; // Assess pixels as a 2D image

} // end of main()

We refer the reader to the IPTL reference manual for the details of template imple-

mentation.

164

Bibliography

 

[1] Visionics Corporation (Facelt), <http://www.faceit.com/>.

[2] F aceSnap, <http: //www.person-spotter.de/htdocs/english/frame_right/
produkte/ facesnap/inhalt.html>.

 

[3] Viisage Technology, <http://www.viisage.com/>.
[4] Eyematic Interfaces, <http://www.eyematic.com/>-
[5] International Biometric Group, <http://www.biometricgroup.com/>.

[6] R. Hietmeyer, “Biometric identification promises fast and secure processing of
airline passengers,” The Int ’1 Civil Aviation Organization Journal, vol. 55, no.
9, pp. 10—11, 2000.

[7] R.-L. Hsu and AK. Jain, “Semantic face matching,” to appear in Proc. IEEE
Int ’l Conf. Multimedia and Expo {ICME}, Aug. 2002.

[8] P. Sinha and T. Poggio, “I think I know that face...,” Nature, vol. 384, pp. 404,
Dec. 1996.

 

[9] Caricature Zone, <http://www.magixl.com/heads/view.html>.
[10] Benjamin Caricature, <http://www.interchile.com/benjamin>.

[11] Destiny Prediction, <http://www.destiny-prediction.com/face/image/ph002.
jpg>.

[12] M arykateandashley. com, <http: //www.marykateandashley.com/ >.

[13] BBC news, <http://news.bbc.co.uk/hi/english/in_depth/americas/2000/us-
elections / profiles / newsid_1012000 / 1012795.stm>.

[14] I ransports. net, <htt p: / / iransportsnet / olympic / photo_gallery / 08.html> .

[15] MPEG7 Content Set from Heinrich Hertz Institute, <http://www.darmstadt.
gmd.de / mobile / hm / projects / MPEG7 / Documents / N2466.html>, Oct. 1998.

[16] Corbis, <http://www.corbis.com/ >.

[17] FaceGen, Singular Inuersions, <http://www.facegen.com>.

165

[18]

[19]

[20]

[21]

[22}

[23]

[24]

[25]

[26]

[27]

[28]

[29]

[30]

[31]

The FACEit System, <http://www.ntu.edu.sg/eee/icis/Staff/yhAng/ faceim—
ages / live_rec . gif> .

R. F éraud, O.J. Bernier, J .-E. Viallet, and M. Collobert, “A fast and accurate
face detection based on neural network,” IEEE Trans. Pattern Analysis and
Machine Intelligence, vol. 23, no. 1, pp. 42—53, Jan. 2001.

D. Maio and D. Maltoni, “Real-time face location on gray-scale static images,”
Pattern Recognition, vol. 33, no. 9, pp. 1525—1539, Sept. 2000.

C. Garcia and G. Tziritas, “Face detection using quantized skin color regions
merging and wavelet packet analysis,” IEEE Transactions Multimedia, vol.
MM-l, no. 3, pp. 264—277, Sept. 1999.

H. Schneiderman and T. Kanade, “A statistical method for 3D object detection
applied to faces and cars,” Proc. IEEE Int ’1 Conf. Computer Vision and Pattern
Recognition, pp. 746—751, June 2000.

HA. Rowley, S. Baluja, and T. Kanade, “Rotation invariant neural network-
based face detection,” Proc. IEEE Int’l Conf. Computer Vision and Pattern
Recognition, pp. 38—44, 1998.

HA. Rowley, S. Baluja, and T. Kanade, “Neural network-based face detection,”

IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 20, no. 1, pp. 23—
38, Jan. 1998.

K.K. Sung and T. Poggio, “Example-based learning for view-based human face
detection,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 20,
no. 1, pp. 39—51, Jan. 1998.

KC. Yow and R. Cipolla, “Feature-based human face detection,” Image and
Vision Computing, vol. 25, no. 9, pp. 713—735, Sept. 1997.

,1

MS. Lew and N. Huijsmans, “Information theory and face detection, Proc.

IEEE Int ’l Conf. Pattern Recognition, pp. 601—605, Aug. 1996.

P.J. Phillips, H. Moon, S.A. Rizvi, and P.J. Rauss, “The FERET evaluation
methodology for face-recognition algorithms,” IEEE Trans. Pattern Analysis
and Machine Intelligence, vol. 22, no. 10, pp. 1090—1104, Oct. 2000.

9

M. Turk and A. Pentland, “Eigenfaces for recognition,’
Neuroscience, vol. 3, no. 1, pp. 71—86, 1991.

Journal of Cognitive

XMZVTS face database, <http://xm2vtsdb.ee.surreyac.uk/home.html>.

J. Weng and D.L. Swets, “Face recognition,” in Biometrics: Personal Identi—
fication in Networked Society, A.K. Jain, R. Bolle, and S. Pankanti, Eds, pp.
67-86, Kluwer Academic, Boston, MA, 1999.

166

 

 

 

[32] L. Wiskott, J .M. Fellous, N. Kruger, and C. von der Malsburg, “F ace recognition
by elastic bunch graph matching,” IEEE Trans. Pattern Analysis and Machine
Intelligence, vol. 19, no. 7, pp. 775—779, 1997.

[33] Reconstruction of Images from Transformed Data 55 BunchCraph, <http
: / / www.cnl.salk.edu / ~wiskott / Projects / { GaborReconstruction.html,
BunchGraph.html}>.

[34] PS. Penev and J .J . Atick, “Local feature analysis: A general statistical theory
for object representation,” Network: Computation in Neural Systems, vol. 7,
no. 3, pp. 477—500, 1996.

[35] Face model generation, <http://www.cs.rutgers.edu/~decarlo/anthface.html>.

[36] H. Wechsler, P. Phillips, V. Bruce, F. Soulie, and T. Huang, Eds, Face Recog-
nition: From Theory to Applications, Springer-Verlag, 1998.

[37] R. Chellappa, C.L. Wilson, and S. Sirohey, “Human and machine recognition
of faces: A survey,” Proc. IEEE, vol. 83, pp. 705-740, May 1995.

[38] W. Zhao, R. Chellappa, A. Rosenfeld, and P.J. Phillips, “Face recog-
nition: A literature survey,” CVL Technical Report, Center for Au-
tomation Research, University of Maryland at College Park, Oct. 2000,
<ftp: //ftp.cfar.umd.edu/TRs/CVL—Reports-2000/TR4167-zhao.ps.gz>.

[39] Faceblind. org, <http://www.faceblind.org/research/index.php>.

[40] A.W.M. Smeulders, M. Worring, S. Santini, A. Gupta, and R. Jain, “Content-
based image retrieval at the end of the early years,” IEEE Trans. Pattern
Analysis and Machine Intelligence, vol. 22, no. 12, pp. 1349—1380, Dec. 2000.

[41] A. Lanitis, C.J. Taylor, and TE Cootes, “Toward automatic simulation of
aging effects on face images,” IEEE Trans. Pattern Analysis and Machine
Intelligence, vol. 24, no. 4, Apr. 2002.

3,

[42] A. Yilmaz and M. Gokmen, “Eigenhill vs. eigenface and eigenedge, Pattern

Recognition, vol. 34, no. 1, pp. 181—184, Jan. 2001.

[43] Y. Adini, Y. Moses, and S. Ullman, “Face recognition: The problem of com-
pensating for changes in illumination direction,” IEEE Trans. Pattern Analysis
and Machine Intelligence, vol. 19, no. 7, pp. 721—732, July 1997.

[44] J .B. Burns, “Recognition via consensus of local moments of brightness and
orientation,” Proc. IEEE Int ’1 Conf. Computer Vision and Pattern Recognition,
pp. 891—898, June 1996.

[45] G.W. Cottrell, M.N. Dailey, C. Padgett, and R. Adolphs, “Is all face processing
holistic? the view from UCSD,” in Computational, Geometric, and Process
Perspectives on Facial Cognition: Contexts and Challenges, M. Wenger and
J. Townsend, Eds, Lawrence Erlbaum Associates, Mahwah, NJ, 2000.

167

 

 

[46]

[47]

[48]

[49]

[50]

[51]

[52]

[531

[55]

[561

[581

Li-Fen Chen, Hong-Yuan Mark Liao, Ja—Chen Lin, and Chin-Chuan Han, “Why
recognition in a statistics-based face recognition system should be based on the

pure face portion: a probabilistic decision-based proof,” Pattern Recognition,
vol. 34, no. 7, pp. 1393-1403, July 2001.

C. Liu and H. Wechsler, “Evolutionary pursuit and its application to face
recognition,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 22,
no. 6, pp. 570—582, June 2000.

G. Rhodes and T. 'Ifemewan, “Understanding face recognition: Caricature

effects, inversion, and the homogeneity problem,” Visual Cognition, vol. 1, pp.
257—311, 1994.

V. Bruce and M. Burton, Eds, Processing Images of Faces, Ablex publishing,
Norwood, NJ, 1992.

MB. Lewis and RA. Johnston, “Understanding caricatures of faces,” Quarterly
Journal of Experimental Psychology A, vol. 51, no. 2, pp. 321-346, May 1998.

RE. Morris and L.H.V. Wickham, “Typicality and face recognition: A critical
re-evaluation of the two factor theory,” Quarterly Journal of Experimental
Psychology A, vol. 54, no. 3, pp. 863—877, Aug. 2001.

B. Bates and J. Cleese, Eds, The Human Face, Dorling kindersley publishing
Inc., New York, NY, 1992.

H. Leder and V. Bruce, “When inverted faces are recognized: The role of
configural information in face recognition,” Quarterly Journal of Experimental
Psychology A, vol. 53, no. 2, pp. 513-536, May 2000.

E. Hjelmas and J. Wroldsen, “Recognizing faces from the eyes only,” Proc. 11th
Scandinavian Conf. Image Analysis, June 1999, <http://citeseer.nj.nec.com/
hjelma899recognizing.html>.

R.G. Uhl Jr. and N. da Vitoria Lobo, “A framework for recognizing a facial
image from a police sketch,” Proc. IEEE Int’l Conf. Computer Vision and
Pattern Recognition, pp. 586—593, June 1996.

H. Chenand Y.Q. Xu, H.Y. Shum, S.C. Zhu, and N.N. Zhen, “Example—based
facial sketch generation with non-parametric sampling,” Proc. IEEE Int ’l Conf.
Computer Vision, Vancouver, Canada, July 2001.

SE. Brennan, “Caricature generator: The dynamic exaggeration of faces by
computer,” Leonardo, vol. 18, no. 3, pp. 170—178, 1985.

R. Mauro and M. Kubovy, “Caricature and face recognition,” Memory 63
Cognition, vol. 20, no. 4, 1992.

168

[59] M. Grudin, “On internal representation in face recognition systems,” Pattern
Recognition, vol. 33, no. 7, pp. 1161—1177, July 2000.

[60] K.M. Lam and H. Yan, “An analytic—to—holistic approach for face recognition
based on a single frontal view,” IEEE Trans. Pattern Analysis and Machine
Intelligence, vol. 20, no. 7, pp. 673—686, July 1998.

[61] D. DeCarlo and D. Metaxas, “Optical flow constraints on deformable models
with applications to face tracking,” Int 7 Journal Computer Vision, vol. 38, no.
2, pp. 99—127, July 2000.

[62] E. Osuna, R. Freund, and F. Girosi, “Training support vector machines: an
application to face detection,” Proc. IEEE Int ’l Conf. Computer Vision and
Pattern Recognition, pp. 130—136, June 1997.

[63] MA. Turk and AP. Pentland, “Face recognition using eigenfaces,” Proc. IEEE
Int ’l Conf. Computer Vision and Pattern Recognition, pp. 586—591, June 1991.

[64] A. Pentland, B. Moghaddam, and T. Starner, “View-based and modular
eigenspaces for face recognition,” Proc. IEEE Int ’l Conf. Computer Vision
and Pattern Recognition, pp. 84—91, June 1994.

[65] R. Brunelli and T. Poggio, “Face recognition: Features vs. templates,” IEEE
Trans. Pattern Analysis and Machine Intelligence, vol. 15, no. 10, pp. 1042—
1052, Oct. 1993.

[66] C. Kotropoulos, A. Tefas, and I. Pitas, “Frontal face authentication using
morphological elastic graph matching,” IEEE Trans. Image Processing, vol. 9,
pp. 555—560, Apr. 2000.

[67] Proc. IEEE Int’l Conf. Automatic Face and Gesture Recognition, 2000,
<http: //www-prima.inrialpes.fr/FG2000/ >.

[68] P. Phillips, H. Wechsler, J. Huang, and P. Rauss, “The FERET database
and evaluation procedure for face-recognition algorithms,” Image and Vision
Computing, vol. 16, no. 5, pp. 295-306, Mar. 1998.

[69] F1. Parke and K. Waters, “Appendix 1: Three-dimensional muscle model fa-
cial animation,” in Computer Facial Animation, pp. 337—338, A.K. Peters,
1996, < http: / / www.crl.researchdigitalcom / publications / books / waters / wa-
ters_book. html>.

[70] G. Miller, E. Hoffert, S.E. Chen, E. Patterson, D. Blackketter, S. Rubin, S. A.
Applin, D. Yim, and J. Hanan, “The virtual museum: Interactive 3D naviga-
tion of a multimedia database,” The Journal of Visualization and Computer
Animation, vol. 3, no. 3, pp. 183—197, July 1992.

169

[71] H. Fuchs, G. Bishop, K. Arthur, L. McMillan, R. Bajcsy, S. Lee, H. Farid, and
T. Kanade, “Virtual space teleconferencing using a sea of cameras,” Technical

Report, Department of Computer Science, University of North Carolina-Chapel
Hill, Number TR94-033, May 1994.

[72] M.J.T. Reinders, P.J.L. van Beek, B. Sankur, and J.C.A. van der
Lubbe, “Facial feature localization and adaptation of a generic face
model for model-based coding,” Signal Processing: Image Communica-
tion, vol. 7, no. 1, pp. 57—74, Mar. 1995, <http://www-it.et.tudelft.nl/
itbibliography / reports / 1995 / journal / imagecom95 .reinders.ps. gz> .

[73] A. Hilton, D. Beresford, T. Gentils, R. Smith, and W. Sun, “Virtual peo-
ple: Capturing human models to populate virtual worlds,” Proc. IEEE Conf.
Computer Animation, pp. 174—185, May 1999.

[74] RT. Azuma, “A survey of augmented reality,” Presence: Teleoperators and
Virtual Environments, vol. 6, no. 4, pp. 355—385, Aug. 1997.

[75] G. Taubin and J. Rossignac, “Geometric compression through topological
surgery,” ACM Trans. Graphics, vol. 17, pp. 84—115, Apr. 1998.

[76] M. Lounsbery, T.D. DeRose, and J. Warren, “Multiresolution analysis for
surfaces of arbitrary topological type,” ACM Trans. Graphics, vol. 16, pp.
34—73, Jan. 1997.

l77l M- Kass, A- Witkin, and D. Terzopoulos, “Snakes: active contour models,”
Int ’1 Journal Computer Vision, vol. 1, no. 4, pp. 321—331, 1988.

[78] W.-S. Hwang and J. Weng, “Hierarchical discriminant regression,” IEEE Trans.
Pattern Analysis and Machine Intelligence, vol. 22, pp. 1277—1293, Nov. 2000.

[79] J. Weng, J. McClelland, A. Pentland, O. Sporns, I. Stockman, M. Sur, and
E. Thelen, “Autonomous mental development by robots and animals,” Science,
vol. 291, no. 5504, pp. 599—600, Jan. 2000.

[80] J. Weng, C.H. Evans, and W.S. Hwang, “An incremental learning method
for face recognition under continuous video stream,” Proc. IEEE Int ’l Conf.
Automatic Face and Gesture Recognition, pp. 251—256, Mar. 2000.

[81] M. Pantic and L.J.M. Rothkrantz, “Automatic analysis of facial expressions:
The state of the art,” IEEE Trans. Pattern Analysis and Machine Intelligence,
vol. 22, no. 12, pp. 1424—1445, Dec. 1996.

[82] M.-H. Yang, N. Ahuja, and D. Kriegman, “Detecting faces in images: A survey,”
IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 24, no. 1, pp. 34—
58, Jan. 2001.

[83] E. Hjelm and BK. Low, “Face detection: A survey,” Computer Vision and
Image Understanding, vol. 83, pp. 236—274, Sept. 2001.

170

 

[84]

[85]

[86]

[87]

[88]

[89]

[90]

[91]

[92]

[93]

[94]

[951
[96]
[97]
[98]
[99]

M. Abdel-Mottaleb and A. Elgammal, “Face detection in complex environments
from color images,” Proc. IEEE Int ’1 Conf. Image Processing, pp. 622—626,
1999.

H. Wu, Q. Chen, and M. Yachida, “Face detection from color images using a
fuzzy pattern matching method,” IEEE Trans. Pattern Analysis and Machine
Intelligence, vol. 21, pp. 557—563, June 1999.

A. Colmenarez, B. Frey, and T. Huang, “Detection and tracking of faces and
facial features,” Proc. IEEE Int’l Conf. Image Processing, pp. 657—661, Oct.
1999.

SC. Bass and A.K. Jain, “Markov face models,” Proc. IEEE Int ’1 Conf.
Computer Vision, pp. 680—687, July 2001.

A.J. Colmenarez and TS. Huang, “Face detection with information based max-
imum discrimination,” Proc. IEEE Int ’l Conf. Computer Vision and Pattern
Recognition, pp. 782—787, June 1997.

V. Bakic and G. Stockman, “Menu selection by facial aspect,” Proc. Vision
Interface, Canada, pp. 203—209, May 1999.

K. Sobottka and I. Pitas, “A novel method for automatic face segmentation,
facial feature extraction and tracking,” Signal Processing: Image Communica-
tion, vol. 12, no. 3, pp. 263—281, June 1998.

H.D. Ellis, M.A. Jeeves, F. Newcombe, and A. Young, Eds, Aspects of Face
Processing, Martinus N ijhoff Publishers, Dordrecht, Netherlands, 1985.

O.A. Uwechue and AS. Pandya, Eds, Human Face Recognition Using Third-
order Synthetic Neural Networks, Kluwer Academic Publishers, Norwell, MA,
1997.

PL. Hallinan, G.G. Gordon, A.L. Yuille, P. Giblin, and D. Mumford, Two- and
Three- Dimensional Patterns of the Face, A.K. Peters, Natick, MA, 1999.

S. Gong, S.J. McKenna, and A. Psarrou, Dynamic Vision: From Images to
Face Recognition, Imperial College Press, London, 1999.

MIT face database, <ftp://whitechapel.media.mit.edu/pub/images/>.

Yale face database, <http://cvc.yale.edu/projects/yalefaces/yalefaces.html>.
AR face database, <http://rv11.ecn.purdue.edu/~aleix/aleix-face.DB.html>.
Olivetti face database, <http://www.cam-orl.co.uk/facedatabase.html>.

B. Moghaddam and A. Pentland, “Face recognition using view-based and mod-
ular eigenspaces,” Automatic Systems for the Identification and Inspection of
Humans, Proc. SPIE, vol. 2257, pp. 12—21, July 1994.

171

[100]

[101]

[102]

[103]

[104]

[105]

[106]

[107]

B. Moghaddam and A. Pentland, “Probabilistic visual learning for object rep—
resentation,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 19,
no. 7, pp. 696—710, July 1997.

3,

PS. Penev and L. Sirovich, “The global dimensionality of face space, Proc.
IEEE Int ’l Conf. Automatic Face and Gesture Recognition, pp. 264—270, Mar.
2000.

I.J. Cox, J. Ghosn, and RN. Yianilos, “Feature-based face recognition us-
ing mixture-distance,” Proc. IEEE Int ’1 Conf. Computer Vision and Pattern
Recognition, pp. 209—216, June 1996.

MS. Kamel, H.C. Shen, A.K.C. Wong, and R. I. Campeanu, “System for the
recognition of human faces,” IBM Systems Journal, vol. 32, no. 2, pp. 307—320,
1993.

I. Craw, N. Costen, T. Kato, G. Robertson, and S. Akamatsu, “Automatic face
recognition: Combining configuration and texture,” Proc. IEEE Int ’l Conf.
Automatic Face and Gesture Recognition, pp. 53—58, Sept. 1995.

CC. Gordon and M. E. Lewis, “Face recognition using video clips and mug
shots,” Proc. Oflice of National Drug Control Policy (0NDCP) Int’l Techni-
cal Symposium, Oct. 1995, <http://http://www.vincent-net.com/gaile/papers
/ONDCPPaper/ONDCPPaper.html>.

R. Lengagne, J.-P. Tarel, and O. Monga, “From 2D images to 3D face geometry,”
Proc. IEEE Int ’l Conf. Automatic Face and Gesture Recognition, pp. 301—306,
Oct. 1996.

J .J . Atick, P.A. Griffin, and AN. Redlich, “Statistical approach to shape from
shading: Reconstruction of 3D face surfaces from single 2D images,” Neural
Computation, vol. 8, no. 6, pp. 1321—1340, Aug. 1996.

[108] Y. Yan and J. Zhang, “Rotation-invariant 3D recognition for face recognition,”

[109]

[110]

[111]

[112]

Proc. IEEE Int ’1 Conf. Image Processing, vol. 1, pp. 156—160, Oct. 1998.

W.Y. Zhao and R. Chellappa, “3D model enhanced face recognition,” Proc.
IEEE Int’l Conf. Image Processing, Sept. 2000, <http://www.cfar.umd.edu/
~wyzhao/icip003Dmodel.ps>.

P. Sinha and T. Poggio, “Role of learning in three—dimensional form percep—
tion,” Nature, vol. 384, pp. 460—463, Dec. 1996.

D. DeCarlo, D. Metaxas, and M. Stone, “An anthropometric face model using
variational techniques,” Proc. SIGGRAPH Conf., pp. 67—74, July 1998.

Q. Chen and G. Medioni, “Building human face models from two images,”
IEEE 2nd Workshop on Multimedia Signal Processing, pp. 117—122, Dec. 1998.

172

 

 

[113] B. Kim and P. Burger, “Depth and shape from shading using the photometric
stereo method,” Computer Vision, Graphics, and Image Processing: Image
Understanding, vol. 54, no. 3, pp. 416—427, 1991.

[114] DR. Hougen and N. Ahuja, “Estimation of the light source distribution and
its use in integrated shape recovery from stereo and shading,” Proc. IEEE Int ’1
Conf. Computer Vision, pp. 148—155, May 1993.

[115] J. Cryer, P.-S. Tsai, and M. Shah, “Integration of shape from shading and
stereo,” Pattern Recognition, vol. 28, no. 7, pp. 1033—1043, July 1995.

[116] F.I. Parke and K. Waters, “Modeling faces,” in Computer Facial Animation,
pp. 55—104, A.K. Peters, Wellesley, MA, 1996.

[117] B. Guenter, C. Grimm, D. Wood, H. Malvar, and F. Pighin, “Making faces,”
Proc. SIGGRAPH Conf., pp. 55—66, July 1998.

[118] L. Yin and A. Basu, “MPEG4 face modeling using fiducial points,” Proc.
IEEE Int ’l Conf. Image Processing, pp. 109—112, Oct. 1997.

[119] W. Lee and N. Magnenat-Thalmann, “Fast head modeling for animation,”
Image and Vision Computing, vol. 18, no. 4, pp. 355—364, Mar. 2000.

[120] R. Lengagne, P. Fua, and O. Monga, “3D stereo reconstruction of human faces
driven by differential constraints,” Image and Vision Computing, vol. 18, no.
4, pp. 337—343, Mar. 2000.

[121] P. Fua, “Using model-driven bundle-adjustment to model heads from raw video
sequences,” Proc. IEEE Int ’1 Conf. Computer Vision, pp. 46—53, Sept. 1999.

[122] J. Ahlberg, “An experiment on 3D face model adaptation using the active
appearance algorithm,” Technical Report, Dept. of EE, Linko'ping University,
Sweden, Jan. 2001.

[123] J. Ahlberg, “Candide-3 - an updated parameterized face,” Report No. LiTH-
IS Y-R-2326, Dept. of EE, Linkoping University, Sweden, Jan. 2001.

[124] A.A. Amini, T.E. Weymouth, and RC. Jain, “Using dynamic programming
for solving variational problems in vision,” IEEE Trans. Pattern Analysis and
Machine Intelligence, vol. 12, no. 9, pp. 855—867, Sept. 1990.

[125] B. Olstad and AH. Torp, “Encoding of a priori information in active contour
models,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 18, no.
9, pp. 863—872, Sept. 1996.

[126] L.H. Staib and J .8. Duncan, “Boundary finding with parametrically deformable
models,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 14, no.
11, pp. 1061—1075, Nov. 1992.

173

[127] C.Y. Xu and J.L. Prince, “Snakes, shapes, and gradient vector flow,” IEEE
Trans. Image Processing, vol. 7, no. 3, pp. 359—369, Mar. 1998.

[128] R. Goldenberg, R. Kimmel, E. Rivlin, and M. Rudzsky, “Fast geodesic active
contours,” IEEE Trans. Image Processing, vol. 10, no. 10, pp. 1467—1475, Oct.
2001.

[129] X.M. Pardo, M.J. Carreira, A. Mosquera, and D. Cabello, “A snake for CT
image segmentation integrating region and edge information,” Image and Vision
Computing, vol. 19, no. 7, pp. 461—475, May 2001.

[130] TE Chan and L.A. Vese, “Active contours without edges,” IEEE Trans. Image
Processing, vol. 10, no. 2, pp. 266—277, Feb. 2001.

[131] C. Chesnaud, P. Refregier, and V. Boulet, “Statistical region snake-based seg-
mentation adapted to different physical noise models,” IEEE Trans. Pattern
Analysis and Machine Intelligence, vol. 21, no. 11, pp. 1145—1157, Nov. 1999.

[132] SC. Zhu and A. Yuille, “Region competition - unifying snakes, region growing,
and Bayes/MDL for multiband image segmentation,” IEEE Trans. Pattern
Analysis and Machine Intelligence, vol. 18, no. 9, pp. 884—900, Sept. 1996.

[133] J. Ivins and J. Porrill, “Statistical snakes: active region models,” Proc. Fifth
British Machine Vision Conf. (BM VC), vol. 2, pp. 377-386, Dec. 1994.

[134] T. Abe and Y. Matsuzawa, “Multiple active contour models with application
to region extraction,” Proc. 15th Int ’1 Conf. Pattern Recognition, vol. 1, pp.
626—630, Sept. 2000.

[135] V. Chalana, D.T. Linker, D.R. Haynor, and Y.M. Kim, “A multiple active
contour model for cardiac boundary detection on echocardiographic sequences,”
IEEE Trans. Medical Imaging, vol. 15, no. 3, pp. 290—298, 1996.

[136] B. Fleming, 3D Modeling 6% Surfacing, Morgan Kaufmann, San Francisco,
California, 1999.

[137] M. Deering, “Geometry compression,” Proc. SIGGRAPH Conf., pp. 13—20,
Aug. 1995.

[138] M. Eck, T. DeRose, T. Duchamp, H. Hoppe, M. Lounsbery, and W. Stuetzle,
“Multiresolution analysis of arbitrary meshes,” Proc. SIGGRAPH Conf., pp.
173—182, Aug. 1995.

[139] R.-L. Hsu, A.K. Jain, and M. Tuceryan, “lVlultiresolution model compression
using 3-D wavelets,” Proc. Asian Conf. Computer Vision, pp. 74—79, Jan. 2000.

[140] A.R. Calderbank, I. Daubechies, W. Sweldens, and B.-L. Yeo, “Lossless image
compression using integer to integer wavelet transforms,” Proc. IEEE Int ’l
Conf. Image Processing, vol. 1, pp. 596—599, Oct. 1997.

174

[141] MD. Adams and F. Kossentini, “Reversible integer-tainteger wavelet trans-
forms for image compression: Performance evaluation and analysis,” Proc.
IEEE Int ’l Conf. Image Processing, vol. 9, no. 6, pp. 1010—1024, June 2000.

[142] IBM Query By Image Content (QBIC), <http://wwwqbic.almaden.ibm.com/>.
[143] Photobook, <http: / / www-white.media.mit.edu / vismod / demos / photobook / >.

[144] M. Abdel-Mottaleb, N. Dimitrova, R. Desai, and J. Martino, “Conivas:
Content-based image and video access system,” Proc. Fourth ACM Multimedia
Conf., pp. 427—428, Nov. 1996.

[145] F ourEyes, <http: / / www-white.media.mit.edu / vismod / demos / photobook /
foureyes/ >.

[146] Virage, <http://www.virage.com/>.

[147] J.-Y. Chen, C. Taskiran, E.J. Delp, and CA. Bouman, “Vibe: A new
paradigm for video database browsing and search,” Proc. IEEE Workshop
Content-Based Access of Image and Video Libraries, pp. 96—100, June 1998,
<http: //stargate.ecn.purdue.edu/~ips/ViBE/ >.

[148] S.-F. Chang, W. Chen, H. Meng, H. Sundaram, and D. Zhong,
“An automated content-based video search system using visual cues,”
Proc. ACM Multimedia, vol. 18, no. 1, pp. 313—324, Nov. 1997,
<http: //www.ctr.columbia.edu/VideoQ/ >.

[149] Visualseek, <http: //www.ctr.columbia.edu/visualseek/ >.

[150] W.Y. Ma and BS. Manjunath, “Netra: A toolbox for navigating large image
databases,” Proc. IEEE Int ’l Conf. Image Processing, vol. 1, pp. 568—571, Oct.
1997.

[151] S. Mehrotra, Y. Rui, M. Ortega, and TS. Huang, “Supporting content-based
queries over images in MARS,” Proc. IEEE Int’l Conf. Multimedia Computing
and Systems, pp. 632—633, June 1997.

[152] J. Laaksonen, M. Koskela, S. Laakso, and E. Oja, “Picsom - content-based
image retrieval with self-organizing maps,” Pattern Recognition Letters, vol.
21, 110. 13-14, pp. 1199—1207, Dec. 2000, <http://www.cishut.fi/picsom/>.

[153] MS. Lew, “N ext-generation web searches for visual content,” IEEE Computer,
pp. 46—53, Nov. 2000, <http://skynet.liacs.nl>.

[154] A. Vailaya, M. Figueiredo, A.K. Jain, and H.-J. Zhang, “Image classification
for content-based indexing,” IEEE Trans. Image Processing, vol. 10, no. 1, pp.
117—130, Jan. 2001.

[155] GA. Poynton, A Technical Introduction to Digital Video, John Wiley & Sons,
New York, 1996.

 

[156]

[157]

[158]

[159}

[160]

[161]

[162]

[163]

[164]

[165]

[166]

[167]

[168]

L.M. Bergasa, M. Mazo, A. Gardel, M.A. Sotelo, and L. Boquet, “Unsupervised
and adaptive gaussian skin-color model,” Image and Vision Computing, vol.
18, pp. 987—1003, Sept. 2000.

J.C. Terrillon, M.N. Shirazi, H. Fukamachi, and S. Akamatsu, “Comparative
performance of different skin chrominance models and chrominance spaces for
the automatic detection of human faces in color images,” Proc. IEEE Int’l
Conf. Automatic Face and Gesture Recognition, pp. 54—61, Mar. 2000.

E. Saber and A.M. Tekalp, “Frontal-view face detection and facial feature
extraction using color, shape and symmetry based cost functions,” Pattern
Recognition Letters, vol. 19, no. 8, pp. 669—680, June 1998.

M.J. Jones and J .M. Rehg, “Statistical color models with application to skin
detection,” Proc. IEEE Int’l Conf. Computer Vision and Pattern Recognition,
vol. 1, pp. 274—280, June 1999.

B. Menser and M. Brunig, “Locating human faces in color images with complex

background,” Intelligent Signal Processing and Communications Systems, pp.
533—536, Dec. 1999.

T. Horprasert, Y. Yacoob, and LS. Davis, “Computing 3-D head orientation
from a monocular image,” Proc. IEEE Int ’1 Conf. Automatic Face and Gesture
Recognition, pp. 242—247, Oct. 1996.

A. Nikolaidis and I. Pitas, “Facial feature extraction and determination of
pose,” Pattern Recognition, vol. 33, pp. 1783—1791, 2000.

W. Huang, Q. Sun, C.-P. Lam, and J .-K. Wu, “A robust approach to face and
eyes detection from images with cluttered background,” Proc. IEEE Int ’1 Conf.
Pattern Recognition, vol. 1, pp. 110—114, Aug. 1998.

K.M. Lam and H. Yan, “Locating and extracting the eye in human face images,”
Pattern Recognition, vol. 29, no. 5, pp. 771—779, 1996.

A. Lanitis, C.J. Taylor, and TR Cootes, “Automatic interpretation and cod-
ing of face images using flexible models,” IEEE Trans. Pattern Analysis and
Machine Intelligence, vol. 19, no. 7, pp. 743—756, July 1997.

S. Sirohey and A. Rosenfeld, “Eye detection,” Technical Report CS-TR-3971,
Univ. of Maryland, Dec. 1998.

F. Smeraldi, O. Carmona, and J. Bigiin, “Saccadic search with Gabor fea-
tures applied to eye detection and real-time head tracking,” Image and Vision
Computing, vol. 18, no. 4, pp. 323—329, 2000.

D. Stork and M Henneke, “Speechreading: An overview of image process-
ing, feature extraction, sensory integration and pattern recognition techniques,”
Proc. IEEE Int ’1 Conf. Automatic Face and Gesture Recognition, pp. xvi—xxvi,
1996.

176

[169] PT. Jackway and M. Deriche, “Scale-space properties of the multiscale mor-
phological dilation-erosion,” IEEE Trans. Pattern Analysis and Machine Intel-
ligence, vol. 18, pp. 38—51, Jan. 1996.

[170] J. Canny, “A computational approach to edge detection,” IEEE Trans. Pattern
Analysis and Machine Intelligence, vol. 8, pp. 679—698, Nov. 1986.

[171] The Champion Database, <http://www.libfind.unl.edu/alumni/events/ break—
fast_for_championshtm>, Mar. 2001.

[172] Yahoo news photos, <http://dailynewsyahoo.com/h/g/ts/>, Dec. 2001.

[173] R.-L. Hsu and A.K. Jain, “Face modeling for recognition,” Proc. IEEE Int’l
Conf. Image Processing, vol. 2, pp. 693—696, Oct. 2001.

[174] SAMPL range databases, <http://sampl.eng.Ohio-state.edu/wsampl/data/
3DDB / RID / minolta/ faceimages.0300 / > .

[175] R.-L. Hsu, M. Abdel-Mottaleb, and A.K. Jain, “Face detection in color images,”
IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 24, no. 5, pp. 696-
706, May 2002.

[176] D. Terzopoulos and K. Waters, “Analysis and synthesis of facial image sequences
using physical and anatomical models,” IEEE Trans. Pattern Analysis and
Machine Intelligence, vol. 15, no. 6, pp. 569—579, 1993.

[177] D. Reisfeld, H. Wolfson, , and Y. Yeshurun, “Contex free attentional operators:

The generalized symmetry transform,” Int ’l Journal Computer Vision, vol. 14,
pp. 119—130, 1995.

[178] H. Breu, J. Gil, D. Kirkpatrick, and M. Werman, “Linear time euclidean dis-
tance transform algorithms,” IEEE Trans. Pattern Analysis and Machine In-
telligence, vol. 17, no. 5, pp. 529—533, May 1995.

[179] R. Zhang, P.-S. Tsai, J. Cryer, and M. Shah, “Shape from shading: a survey,”
IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 21, no. 8, pp.
690—706, Aug. 1999.

[180] E. Trucco and A. Verri, Introductory Techniques for 3—D Computer Vision,
Prentice Hall, 1999.

[181] V. Caselles, R. Kimmel, , and G. Sapiro, “Geodesic active contours,” Int’l
Journal Computer Vision, vol. 22, no. 1, pp. 61—79, 1997.

[182] D. Adalsteinsson and J .A. Sethian, “A fast level set method for propagating
interfaces,” Journal Computational Physics, vol. 118, pp. 269—277, 1995.

[183] X. Han, C. Xu, and J .L. Prince, “A topology preserving deformable model using
level sets,” Proc. IEEE Int’l Conf. Computer Vision and Pattern Recognition,
vol. II, pp. 765-770, Dec. 2001.

177

[184] H. Chernoff, “The use of faces to represent points in k-dimensional space graph-
ically,” Journal of the Americal Statistics Association, vol. 68, pp. 361—368,
1973.

[185] R.-L. Hsu and A.K. Jain, “Detection and tracking of multiple faces in video,”
Technical Report MS U- CSE-02-1 3, vol. Dept. Computer Science & Engineering,
pp. Michigan State University, May 2002.

[186] Intel Image Processing Library, <http: / / developer.intel.com / software/ products
/perflib/ipl/index.htm>.

[187] Intel Open Source Computer Vision Library, <http://developer.intel.com/ soft-
ware/opensource/cvfl/opencv_download.htm>.

[188] SE. Umbaugh, Computer Vision and Image Processing: A Practical Approach
Using CVIPTools, Prentice Hall, Upper Saddle River, NJ, 1997.

178

IIIIIIIIIIIIIIIIIIIIIIIIIIIIII

llll]llllllll lllllllllll][][]lll

1293 02328 7