algorithms to process and measure biometric

ALGORITHMS TO PROCESS AND MEASURE
BIOMETRIC INFORMATION CONTENT IN
LOW QUALITY FACE AND IRIS IMAGES
RICHARD YOUMARAN
Thesis submitted to the
Faculty of Graduate Studies and Research
In partial fulfillment of the requirements
For the degree of Doctor of Philosophy in
Electrical and Computer Engineering
School of Information Technology
Faculty of Graduate and Postdoctoral Studies
University of Ottawa
©Richard Youmaran, Ottawa, Canada, 2011
Abstract
Biometric systems allow identification of human persons based on physiological or
behavioral characteristics, such as voice, handprint, iris or facial characteristics. The use
of face and iris recognition as a way to authenticate user’s identities has been a topic of
research for years. Present iris recognition systems require that subjects stand close
(<2m) to the imaging camera and look for a period of about three seconds until the data
are captured. This cooperative behavior is required in order to capture quality images for
accurate recognition. This will eventually restrict the amount of practical applications
where iris recognition can be applied, especially in an uncontrolled environment where
subjects are not expected to cooperate such as criminals and terrorists, for example. For
this reason, this thesis develops a collection of methods to deal with low quality face and
iris images and that can be applied for face and iris recognition in a non-cooperative
environment. This thesis makes the following main contributions:
I.
For eye and face tracking in low quality images, a new robust method is
developed. The proposed system consists of three parts: face localization, eye detection
and eye tracking. This is accomplished using traditional image-based passive techniques
such as shape information of the eye and active based methods which exploit the spectral
properties of the pupil under IR illumination. The developed method is also tested on
underexposed images where the subject shows large head movements.
II.
For iris recognition, a new technique is developed for accurate iris
segmentation in low quality images where a major portion of the iris is occluded. Most
existing methods perform generally quite well but tend to overestimate the occluded
2
regions, and thus lose iris information that could be used for identification. This
information loss is potentially important in the covert surveillance applications we
consider in this thesis. Once the iris region is properly segmented using the developed
method, the biometric feature information is calculated for the iris region using the
relative entropy technique. Iris biometric feature information is calculated using two
different feature decomposition algorithms based on Principal Component Analysis
(PCA) and Independent Component Analysis (ICA).
III.
For face recognition, a new approach is developed to measure biometric
feature information and the changes in biometric sample quality resulting from image
degradations. A definition of biometric feature information is introduced and an
algorithm to measure it proposed, based on a set of population and individual biometric
features, as measured by a biometric algorithm under test. Examples of its application
were shown for two different face recognition algorithms based on PCA (Eigenface) and
Fisher Linear Discriminant (FLD) feature decompositions.
3
Acknowledgments
I would like to dedicate this thesis to my son, Alexandre and to my wife, Nadine who
never stopped being a source of support and encouragement.
I would also like to take this opportunity and thank my supervisor, Dr Andy Adler, for his
support, guidance and understanding. His comments and suggestions for further
development as well as his assistance during writing this thesis are invaluable to me.
I would also like to thank my friends, Liepeng Xie and Yednek Asfaw for their constant
support and help. We have had many useful discussions that played a major role in my
learning process.
Finally, I want to express my deepest appreciation for my parents, my sister and my
brother for being a constant source of loving support throughout the years.
4
Contents
Abstract ............................................................................................................................... 2
Acknowledgments............................................................................................................... 4
1
Chapter 1................................................................................................................... 14
Introduction....................................................................................................................... 14
1.1
Thesis Objectives .............................................................................................. 14
1.2
Thesis contributions .......................................................................................... 16
1.3
Thesis outline .................................................................................................... 19
1.4
Image databases ................................................................................................ 21
2
Chapter 2................................................................................................................... 23
Biometrics review ............................................................................................................. 23
2.1
Biometrics technology ...................................................................................... 23
2.2
Multimodal biometric systems.......................................................................... 28
2.3
Properties of Biometrics ................................................................................... 29
2.4
Classification of Biometric System .................................................................. 31
2.5
Biometric Sample Quality Measures ................................................................ 34
2.6
Face recognition................................................................................................ 35
2.7
Face tracking algorithms................................................................................... 39
2.7.1
Knowledge-Based Classifiers ................................................................... 40
2.7.2
Learning-Based Classifiers ....................................................................... 45
2.7.3
Motion estimation ..................................................................................... 47
2.8
Iris Recognition................................................................................................. 49
2.8.1
Iris structure .............................................................................................. 52
2.8.2
Iris texture pattern and colors ................................................................... 52
2.8.3
Imaging Systems....................................................................................... 55
2.8.4
Iris Localization and Segmentation .......................................................... 57
2.8.5
Size-invariant Unwrapping and Representation ....................................... 64
2.8.6
Feature Extraction..................................................................................... 67
2.8.7
Matching Algorithms and Distance Measure ........................................... 73
2.8.8
Evaluation Metrics .................................................................................... 75
2.8.9
Image Database and Open Source Software............................................. 80
2.9
Non-Cooperative Iris Recognition.................................................................... 81
2.10 Summary ........................................................................................................... 84
3
Chapter 3................................................................................................................... 86
Using infrared illumination to improve eye & face tracking in low quality video images
........................................................................................................................................... 86
3.1
Introduction....................................................................................................... 86
3.2
Algorithm design .............................................................................................. 88
3.3
Face detection ................................................................................................... 89
3.3.1
Experimental setup.................................................................................... 89
3.3.2
Non-linear image enhancement and denoising ......................................... 90
5
3.3.3
Histogram stretch ...................................................................................... 91
3.3.4
Non-linear coarse edge enhancement ....................................................... 92
3.3.5
Morphological image erosion operation and Edge detection ................... 92
3.3.6
Initial face contour extraction ................................................................... 93
3.4
Initial eye detection........................................................................................... 94
3.5
Eye and Face tracking....................................................................................... 97
3.5.1
Template correlation ................................................................................. 97
3.6
Eigen-eyes......................................................................................................... 98
3.7
Face detection using the previous face template............................................. 100
3.8
Pupil candidate regions computation and eye tracking................................... 101
3.9
Kalman filtering .............................................................................................. 105
3.10 Eye contour extractor...................................................................................... 106
3.11 Experimental Results ...................................................................................... 107
3.12 Discussion ....................................................................................................... 111
3.13 Summary ......................................................................................................... 112
4
Chapter 4................................................................................................................. 113
Improved Identification of Iris and Eyelash Features..................................................... 113
4.1
Introduction..................................................................................................... 113
4.2
Enhanced Segmentation.................................................................................. 115
4.2.1
Pupil-Iris region localization and boundary extraction........................... 115
4.2.2
Eyelash detection .................................................................................... 119
4.2.3
Enhanced Iris Recognition...................................................................... 123
4.2.4
Iris Unwrapping ...................................................................................... 126
4.2.5
Pattern Matching..................................................................................... 131
4.3
Results............................................................................................................. 132
4.4
Discussion ....................................................................................................... 143
4.5
Summary ......................................................................................................... 145
5
Chapter 5................................................................................................................. 146
Measuring Information Content in Biometric Features .................................................. 146
5.1
Introduction..................................................................................................... 146
5.2
Theoretical framework.................................................................................... 150
5.2.1
Requirements for biometric feature information .................................... 150
5.2.2
Distribution modeling ............................................................................. 152
5.2.3
Regularization Methods for degenerate features .................................... 156
5.2.4
Regularization Methods for insufficient data ......................................... 157
5.2.5
Average information of a biometric system............................................ 158
5.2.6
Information loss due to degradation ....................................................... 159
5.3
Face recognition.............................................................................................. 160
5.3.1
Biometric information calculations......................................................... 163
5.3.2
Degraded features ................................................................................... 169
5.4
Biometric Iris Features Information................................................................ 172
5.5
Discussion ....................................................................................................... 177
5.6
Summary ......................................................................................................... 184
6
6
Chapter 6................................................................................................................. 186
6.1
Discussion ....................................................................................................... 186
6.2
Future Work .................................................................................................... 192
7
List of Figures
Figure 1.1: Images taken from (a) Underexposed, (b) RPI ISL IR , (c) CASIA and (d)
Aberdeen databases, respectively. .................................................................................... 22
Figure 2.1: Evaluation Simplified block diagram representation of a biometric system.. 25
Figure 2.2: Evaluation of the matching accuracy of a biometric system. Histograms of the
genuine and impostor matching scores are represented as well as the two types of errors
that can arise in a biometric system given a matching score threshold (T). The areas A
and B represent false accept rate (FAR) and false reject rate (FRR), respectively. ......... 27
Figure 2.3: Receiver operating characteristic curve (ROC) showing the relation between
false acceptance and false rejection in a biometric system............................................... 28
Figure 2.4: Integral image used to calculate the sum of pixels in a rectangle. For example,
the value of the integral image at location 1 is the sum of the pixels in rectangle A. The
value at location 2 is A+B, at location 3 is A+C, and at 4 is A+B+C+D. The sum within
D is computed as (4+1)-(2+3)........................................................................................... 47
Figure 2.5: Different stages in an iris recognition system ................................................ 51
Figure 2.6: Iris image........................................................................................................ 55
Figure 2.7: The Daugman iris imaging system................................................................. 56
Figure 2.8: The Wildes iris imaging system ..................................................................... 57
Figure 2.9: Iris image showing severe specular reflections taken from Montgomery
(2007)................................................................................................................................ 62
Figure 2.10: Unwrapping of the Iris using Daugman's Rubber Sheet Model................... 65
Figure 2.11: Uniform feature points sampling with the Daugman's rubber sheet model. 67
Figure 2.12: Feature points sampling with displaced pupil and iris centers..................... 67
Figure 2.13: Phase quantization (taken from Masek (2003)) ........................................... 70
Figure 2.14: FAR and FRR seen from the overlap of the intra-class and inter-class
distributions. Also, as an example, the separation Hamming distance in this figure is 0.35
(Masek (2003)).................................................................................................................. 76
Figure 2.15: Possible non-cooperative face detection and iris recognition system. ......... 84
8
Figure 3.1: Multistage Algorithm block diagram showing three stages. At each stage, the
input image (F or A) is filtered using a Gaussian [5 × 5] low-pass filter. An image
containing only high-frequencies H(x,y) is obtained by subtracting the smoothed output
from the input. The edge amplification parameter si is selected at each stage, i, based on
the level of high frequency noise in Hi(x,y). c is a scalar controlling the contrast level in
the enhanced image........................................................................................................... 91
Figure 3.2: (a) Low quality underexposed IR image showing shows that most pixels have
low intensity values due to poor illumination, (b) its corresponding histogram, (c) Noise
reduction and contrast enhancement using the log-ratio approach. The histogram stretch
operation (d) presents better edge delineation around the face region. Image a) shows that
the pupil classification problem is very challenging since subject’s eyes show very weak
reflections.......................................................................................................................... 91
Figure 3.3: (a) Coarsely enhanced image using non-linear enhancement algorithm, (b)
binary image using tH as threshold.................................................................................... 92
Figure 3.4: (a) Inner face contour extracted using a Sobel operator and connected
component analysis, (b) plot of pixel intensity summation in the vertical direction. ....... 93
Figure 3.5: Initial Eye detection block diagram................................................................ 95
Figure 3.6: Image differencing process: (a) The original bright pupil image, (b) dark pupil
image, (c) image obtained after morphological opening with a disk structure of size 2, (d)
Subtraction result of images: [a-b], (e) Subtraction result of: [a-c], (f) Thresholded image
(d) using a very small threshold to account for most reflections in the image, (g)
Thresholded image (e) using the same small threshold, (h) Image obtained using the
logical AND operator which keeps the bright regions which appear in both thresholded
images. .............................................................................................................................. 97
Figure 3.7: Block diagram showing face detection using prior face template. This
algorithm is activated when the processed image does not contain valid eye regions
possibly due to out-of-plane head rotation or occlusions. .............................................. 101
Figure 3.8: Block diagram showing eye and face tracking algorithm. The algorithm is
initialized with the previously computed eyes and face location at time t-1. Subsequent
frames are then processed using a Kalman-based and adaptive thresholding techniques to
successfully track eyes in low quality images. ............................................................... 104
Figure 3.9: Eye detection results obtained using low quality images taken in very low
illumination conditions. .................................................................................................. 109
Figure 3.10: Eye detection results obtained using images from the ISL IR EYE database.
Frames (b,d) show eye detection under challenging conditions where the subject has his
eyes closed. ..................................................................................................................... 110
9
Figure 4.1: Iris segmentation algorithm based on local image enhancement ................. 119
Figure 4.2: Four different masks for detecting horizontal, vertical and diagonal edges,
respectively. .................................................................................................................... 121
Figure 4.3: Eyelash detection algorithm and ideal iris region segmentation. ................. 122
Figure 4.4: Iris image taken from the CASIA database.................................................. 125
Figure 4.5: Iris segmentation using the Masek’s algorithm. As seen in the image, the iris
region includes some eyelash occlusion. In addition, some of the valid iris pixels are
discarded as noise. .......................................................................................................... 125
Figure 4.6: Iris segmentation using the developed enhanced eyelash detection algorithm.
......................................................................................................................................... 126
Figure 4.7: Example of the iris region unwrapping using the rubber sheet model described
in (Xie (2007)). ............................................................................................................... 128
Figure 4.8: Iris feature sampling points using the Masek’s iris segmentation scheme... 129
Figure 4.9: Iris feature sampling points using the enhanced iris segmentation scheme. 129
Figure 4.10: Iris region unwrapping using the Masek’s technique................................. 130
Figure 4.11: Iris region unwrapping using the enhanced technique. .............................. 130
Figure 4.12: Example of a binary mask applied on the unwrapped iris image in the
Masek’s segmentation technique. The binary 0s (black pixels) indicate noise and are used
to discard the underlying pixel values in the iris template. The pixel value corresponding
to 1s (white) are used for the Hamming distance calculation......................................... 130
Figure 4.13: Example of a binary mask applied on the unwrapped iris image in the
enhanced segmentation technique. The mask shows accurate eyelash and noise detection.
......................................................................................................................................... 130
Figure 4.14: Image enhancement result: (a) Original image of the eye, (b) Non-linear
image enhancement, (c) Binarized image....................................................................... 135
Figure 4.15: Accurate iris boundary extraction and enhancement: (a) Approximated
location of the iris outer boundary using the Hough transform, (b) Edge map and accurate
iris boundary calculation, (c) Accurate pupil-iris boundary extraction, (d) Exact Pupil-iris
region segmentation, (e) Pupil-iris local region enhancement, (f)Non-iris eye image, (g)
Non-iris local image enhancement.................................................................................. 137
10
Figure 4.16: Eyelash detection and iris segmentation examples: (a, b, c) Original eye
images, (d, e, f) Computed candidate eyelash points using our algorithm, (g, h, i)
Accurate segmentation of the iris regions without eyelash occlusions........................... 138
Figure 4.17: Normalized iris images in the polar space. (a) shows a normalized iris region
without eyelash detection, (b) shows the result of iris normalization using the Masek’s
algorithm and (c) represents normalization using the enhanced segmentation algorithm.
The y-axis represents the radial resolution and the x-axis corresponds to the
circumferential (circular) resolution. .............................................................................. 139
Figure 4.18: (a) Plot of the intra-class and inter-class distribution using the Masek’s
segmentation algorithm, (b) Genuine-impostor distributions using the enhanced
segmentation algorithm................................................................................................... 140
Figure 4.19: The decidability ( ) measure showing a higher decidability measure for the
enhanced algorithm......................................................................................................... 141
Figure 4.20: DET curve showing results difference between the Masek's segmentation
and the enhanced segmentation. The enhanced segmentation method shows better results
at FAR>0.05 where a lower FRR is obtained compared to Masek’s algorithm. On the
other hand, Masek’s method seems to be superior for FAR<0.05. ................................ 142
Figure 4.21: Cumulative Match Curve comparison. A 95.25% rank-1 identification rate is
obtained using our proposed method while 93.67% is calculated using Masek’s
technique. ........................................................................................................................ 143
Figure 5.1: An example of PCA (Eigenface) face features. From left to right, PCA
features number 3, 15, 35, 55 are shown. The PCA features are othonormal and fit the
data in a least squares sense. ........................................................................................... 162
Figure 5.2: An example of FLD face features. From left to right, FLD features number 7,
10, 30, 50 are shown. FLD attempts to maximize class separation while minimizing the
within class scatter. ......................................................................................................... 162
Figure 5.3: Biometric information (bits) as a function of number of features for (A) PCA
(Eigenface)...................................................................................................................... 166
Figure 5.4: The regularized intra-person covariance matrix Sp showing dominant
components along its diagonal. Since Σp represents similar information to Σq it is
reasonable to expect the matrices have similar eigenvectors, resulting in strong diagonal
components in Σp............................................................................................................ 166
Figure 5.5: Biometric information (in bits) (y-axis) vs. the mask size (L) (x-axis) for each
person. Each subfigure represents a different value of Np (images of the same person):
diverges as becomes
(A) 8, (B) 12, (C) 16 and (D) 18. The curves show that
singular (L ≥ Np). The relative entropy increases with the size of the mask. ................ 167
11
Figure 5.6: Average
vs L (x-axis) for Np = 18. Each line represents the average of
information calculated for a population of 16 individuals with 18 images each using PCA
(middle), FLD (bottom) and a fusion of PCA and FLD features (top)........................... 168
Figure 5.7: DET curve showing the FRR vs FAR for PCA, Fisher and fusion of PCA, and
Fisher features, respectively............................................................................................ 168
Figure 5.8: Degraded image obtained by applying a Gaussian blur to (b) a section of the
and to (c) the entire image
. ....................................................... 170
original image
Figure 5.9: ∆BI as a function of an increasing blur level for images taken from (a) and
(b) ............................................................................................................................... 171
Figure 5.10: Biometric Eigen iris feature information computed for 327 iris features. The
y-axis represents the biometric information for each feature (in bits) and the x-axis is the
feature number. The top graph is calculated using the Masek’s algorithm while the
bottom graph is generated using the enhanced technique. The standard deviation is also
plotted at the bottom of each graph................................................................................. 176
Figure 5.11: Biometric ICA iris feature information computed for 327 iris features where
the features are extracted from the iris region at a constant angle/varying radius. The yaxis represents the biometric information for each feature (in bits) and the x-axis is the
feature number. The top graph is calculated using the Masek’s algorithm while the
bottom graph is generated using the enhanced technique. The standard deviation is also
plotted at the bottom of each graph................................................................................. 176
Figure 5.12: Biometric ICA iris feature information computed for 327 iris features where
the features are extracted from the iris region at a varying angle/constant radius. The yaxis represents the biometric information for each feature (in bits) and the x-axis is the
feature number. The top graph is calculated using the Masek’s algorithm while the
bottom graph is generated using the enhanced technique. The standard deviation is also
plotted at the bottom of each graph................................................................................. 177
as a function of the mean feature variance (arbitrary units)
Figure 5.13: Average
(x-axis) for 16 different persons. The mean feature variance is computed by summing all
the diagonal components of matrix for each person. The correlation coefficient is
, which is significant at
.............................................................................. 184
12
Nomenclature
BI
Biometric Information
CMC
Cumulative Match Curve
DCAC
Discrete Circular Active Contour
DET
Detection Error Trade-off
EER
Equal Error Rate
FAR
False Accept Rate
FDA
Fisher Discriminant Analysis
FMR
False Match Rate
FNMR
False Non Match Rate
FRR
False Reject Rate
HD
Hamming Distance
ICA
Independent Component Analysis
IR
Infrared
LDA
Linear Discriminant Analysis
LoG
Laplacian of Gaussian
NC
Normalized Correlation
NIR
Near Infrared
PCA
Principal Component Analysis
ROC
Receiver Operating Characteristic
SVD
Singular Value Decomposition
SBI
System Biometric Information
WED
Weighted Euclidean Distance
13
1 Chapter 1
Introduction
1.1 Thesis Objectives
Biometrics is an emerging field of information technology that is crucial for human
identification and verification. Biometric technologies measure and recognize human
physical and behavioral characteristics for authentication purposes. Some of the most
common physical characteristics include fingerprints, irises, and facial patterns. The use
of face and iris recognition as a way to authenticate user’s identities has been a major
topic of research (Jain et al. (2004)). While many image processing algorithms exist for
iris and face recognition, their performance is not completely reliable, especially in
situations with variable lighting, and when dealing with low resolution images (Ma et al.
(2004), Tajbakhsh et al. (2008)). Face and iris recognition algorithms are widely used; in
most cases, images are taken from a cooperative individual under a controlled
environment in order to provide satisfactory results (Ma et al. (2004)). Present iris
recognition systems require that subjects stand close (< 2m) to the imaging camera and
look for a period of about three seconds until the data are captured (International
Biometric Group (2005)). This cooperative behavior is required in order to capture good
quality images for accurate recognition. This will eventually restrict the amount of
practical applications where iris recognition can be applied, especially in an uncontrolled
environment where subjects are not expected to cooperate. Hence, these techniques have
14
limited capability of identifying non-cooperative subjects for applications such as
surveillance, where the observed individuals are non-cooperating and/or non-habituated
to the technology.
For this reason, this thesis aims on developing a new system that is composed of a
collection of methods that deal with low quality face and iris images and that can be
applied for face and iris recognition in a non-cooperative environment. Hence, the
developed techniques can be used for automatically detecting and recognizing human
subjects via their face and iris traits using images taken at a long distance without any
subject cooperation.
Images taken from a non-cooperating individual at-a-distance in a very dynamic
environment tend to include more distortions and noise (i.e. low quality underexposed
images, obstructions by eyelids or eyelashes and reflections). For the terms of our work
and of this thesis, all these factors are considered as noise and the processed images are
described as low quality.
The thesis objectives consist of the following:
i.
Using infrared illumination to improve eye and face tracking in low quality and
underexposed video images where subject is moving.
15
ii.
Improved identification of iris and eyelash features in low quality images where a
major portion of the iris is occluded by eyelash noise, eyelid or specular
reflections.
iii.
Measuring information content in biometric iris and face features.
iv.
Measuring biometric sample quality in terms of biometric information for face
features.
1.2 Thesis contributions
The major thesis contributions include:
i.
For eye and face tracking in low quality images, a new system is developed using
the “Bright-pupil” effect — specular reflection from the retina of human eyes under coaxial infrared illumination (Haro et al. (2000), Morimoto et al. (1998)). By using IR
illumination, it is possible to get information from which the eye positions in the image
can be calculated. Our algorithm consists of three parts: face localization, eye detection
and eye tracking. This is accomplished using traditional image-based passive techniques
such as shape information of the eye and active based methods which exploit the spectral
properties of the pupil under IR illumination. We also address the problem of processing
eye image containing weak reflections due to occlusion or eye closure using adaptive
thresholding techniques to extract the eye contour. The developed method is also tested
on underexposed images where the subject shows large head movements.
16
This work resulted in the following publications:
•
Conference paper in the 24th Queen's Biennial Symposium on Communications,
QBSC 2008, Kingston, Canada, June 24 - June 26 (Youmaran et al. (2008)).
•
Conference paper in Canadian Conference of Computer and Electrical
Engineering (CCECE), Ottawa, Canada, May 7-10 (Youmaran et al. (2006))
ii.
A new technique is developed for accurate iris segmentation using low quality iris
images. A major portion of the iris is occluded with eyelash, eyelids and/or specular
reflections. Most existing methods tend to overestimate the occluded regions, and thus
lose iris information that could be used for identification. For this reason, the new method
addresses most of these issues by using a collection of image processing techniques such
as: non-linear image enhancement, edge detection, morphological operators, Hough
transform, intensity gradient based algorithm and a block mean and variance method
using region’s local statistics.
This work resulted in the following publication:
•
Conference paper in the 24th Queen's Biennial Symposium on Communications,
QBSC 2008, Kingston, Canada, June 24 - June 26 (Youmaran et al. (2008)).
iii.
For face recognition, an approach is developed to measure biometric feature
information and the changes in biometric sample quality resulting from image
degradations. A definition of biometric feature information is introduced and an
algorithm to measure it proposed, based on a set of population and individual biometric
17
features, as measured by a biometric algorithm under test. Examples of its application
were shown for two different face recognition algorithms based on PCA (Eigenface) and
FLD feature decompositions. Subsequently, we introduced a measure of information loss
as a function of image degradation.
This work resulted in the following publications:
•
Journal paper in Pattern Analysis & Applications, 12:261-270 (Adler et al.
(2009)).
•
Book chapter in Biometrics: Theory, Methods, and Applications N.V. Boulgouris,
K.N. Plataniotis, and E.Micheli-Tzanakou (Eds), In press: Wiley/IEEE2008
(Youmaran et al. (2008)).
•
In Biometrics Consortium Conference 2006, Sep. 19-21, Baltimore, MD, USA
(Youmaran et al. (2006)).
•
Conference paper in Canadian Conference of Computer and Electrical
Engineering (CCECE), Ottawa, Canada, May 7-10 (Youmaran et al. (2006)).
iv.
For iris recognition, a new algorithm is developed to calculate biometric
information for a set of iris features using the relative entropy measure developed in this
thesis (section 5.2). The developed algorithm is divided in the following steps:
i. Distribution modeling of iris biometric features, ii. Relative entropy calculations, iii.
ICA iris feature extraction and biometric information calculation. The iris regions used in
the entropy calculation are obtained using the enhanced iris segmentation techniques
developed in section (4.2).
18
1.3 Thesis outline
Chapter 2 presents a review of Biometrics. It includes a brief description of current
biometric technology related to face detection, face recognition and iris recognition. It
also describes some of the face detection and recognition algorithms that are used in
similar applications.
Chapter 3 presents a new system for face detection and tracking that is designed for low
quality video images using a Kalman filter type tracker. The technique extracts and tracks
face and eye positions from surveillance type images with infrared strobe taken under
poor illumination. In the case where many reflections occur, the algorithm will find all
possible eye locations and presents the best solution using multi-stage classification
techniques. The algorithm is tested using 1800 images taken from two different IR image
databases.
Chapter 4 presents a new technique for eyelash noise detection, accurate iris boundary
extraction and ideal iris segmentation. The proposed techniques locate the iris region
using logarithmic image enhancement and the Hough transform techniques, locate the iris
boundary, extract the exact iris contour, detect eyelashes based on the local image
statistics and block intensity and propose an improved iris model for accurate iris
recognition. The developed method overcomes the limitations encountered in other iris
segmentation and eyelash detection techniques such that it detects accurately separable
and multiple eyelashes, extracts the exact iris contour and is illumination invariant. The
19
performances of the enhanced method are studied and compared to the existing Masek’s
technique implemented based on Daugman’s work (Masek (2003), Daugman (2003)).
Chapter 5 presents a new approach to measure biometric feature information and the
changes in biometric sample quality resulting from image degradations. A definition of
biometric feature information is introduced and an algorithm to measure it proposed,
based on a set of population and individual biometric features, as measured by a
biometric algorithm under test. Examples of its application were shown for two different
face recognition algorithms based on PCA (Eigenface) and FLD feature decompositions.
Subsequently, a new measure of information loss as a function of image degradation is
introduced.
This chapter presents a new approach to measure biometric feature information using the
segmented iris region from chapter 4. A definition of iris biometric feature information is
introduced and an algorithm to measure it proposed, based on a set of biometric iris
features. Biometric information (in bits) is calculated for iris using two different feature
decomposition spaces. First, biometric information is computed on PCA features and
then using ICA feature decomposition scheme. Since iris features tend to be nonGaussian, biometric information using ICA features seems to be a more accurate measure
since ICA maximizes non-Gaussianity.
Finally, chapter 6 concludes this thesis by summarizing the work proposed in it and
discussing possible future applications where the developed methods can be used.
20
1.4 Image databases
The following image databases are used to test the methods developed in this thesis
(Figure 1.1).
Databases
Underexposed face
images
(Asfaw et al. (2002))
Aberdeen face images
(Belhumeur et al. (1997))
Images
600 images
288 images
• 18 images per subject
• 16 subjects
• Resolution:
RPI ISL IR eye images
(Wang et al. (2005))
1200 images
• 9 subjects
• Resolution:
Description
•
•
•
•
•
IR illumination
Very low intensity images
Poor edge delineation
Subject moving
Differing pose (angle)
•
•
•
•
Low quality
Non-uniform illumination
Various facial expressions
Out-of-focus
•
•
•
IR illumination
Challenging for tracking
Eye closure and
occlusions
Rapid head movement
In/out – of plane rotation
Various Facial
expressions
Non-uniform illumination
•
•
•
•
CASIA iris images
(Chinese Academy of
Sciences Institute of
Automation. (2004))
689 images
•
•
• 108 different classes
• 6 or 7 images per class
• Resolution:
•
•
•
Noisy iris images
Eyelids and eyelash
occlusions
Out-of-focus
Specular reflections
Lighting reflections
21
(a)
(c)
(b)
(d)
Figure 1.1: Images taken from (a) Underexposed, (b) RPI ISL IR, (c) CASIA and (d) Aberdeen
databases, respectively.
22
2 Chapter 2
Biometrics review
2.1 Biometrics technology
Biometrics is an emerging field of information technology which aims to automatically
identify individuals using their unique biological traits. By measuring the physiological
and behavioral characteristics using the individual’s biological samples, it has been
shown that information characteristics of each individual can be extracted in order to
verify the identity of that individual in a population. The main advantages of biometrics
over other standard security systems are that biometric traits cannot be forgotten or lost.
They are difficult to copy, share and distribute and they require the person to be present
at the time of authentication. In most applications, a biometric system is a real-time
identification system which requires the measurement of unique information (i.e.
features) characterizing the individual being enrolled or tested and later comparing these
samples against a database containing several candidates. Figure 2.1 shows a simplified
block diagram representation of a biometric system. The first block (the sensor) acquires
the necessary data to be processed and represents the interface between the real-world
and the biometric system. Afterwards, the pre-processing block is used to remove
artifacts and noise from the data using advanced image processing techniques that
enhance the acquired information at the input of the system. Once the data are cleaned,
the feature extractor creates feature vectors designed to be unique to each individual.
23
These features are used for the identification and matching process. Using these features,
a template vector or image is then created for further processing. At the enrollment stage,
the templates are stored in the database while during the matching (i.e. identification)
phase, the extracted feature vectors are passed to a matcher that compares them to
existing templates in the database. A similarity distance is then calculated using an
algorithm which then sends a decision to the output of the biometric system allowing or
restricting the individual for further operations. Depending on the context and the
application, a biometric system can be either a verification/authentication (Am I whom I
claim I am?) or an identification system (Who am I?) (Jain et al. (2004)). Verification
implies making sure that the person, whose biometric information is already known in the
system, is perhaps the one that he is claiming to be. This is a 1:1 match verification
process which involves extracting new biometric features and then comparing them to the
ones in the database in order to confirm or deny a person’s claimed identity. On the other
hand, identification involves establishing a person’s identity. This involves extracting
biometric information from the person and comparing them to the database. It is a 1: N
match verification operation. It is a more computationally expensive process since most
databases tend to contain a large number of templates.
24
Enrollment
Stored
templates
Stored feature set
Pre-processing
Feature
extractor
Template
generator
Identification
Matcher
Feature set
Sensor
Accept/reject
Biometric system
Figure 2.1: Evaluation Simplified block diagram representation of a biometric system
In order to analyze the performance of a biometric system, the probability distribution of
genuine and impostor matching scores is examined. A genuine matching score is
obtained by comparing two feature sets of the same individual while an impostor
matching score is the result of comparing feature sets corresponding to two different
individuals. If the matching score is higher than a certain threshold, it is then assumed
that the two feature sets correspond to the same individual; otherwise, they are assumed
to come from different subjects. Thus, there are two types of error that a verification
biometric system can make (Jain et al. (2004)): (i) False rejection (type I error) which
occurs when a genuine matching score does not exceed the threshold which means that a
legitimate user is rejected and (ii) False acceptance (type II error) which occurs when an
impostor matching score exceeds the threshold, meaning that an illegitimate user is
accepted as someone else. False rejection tends to occur when the threshold in the
Biometric System is selected to be more severe such as in facilities where high security is
25
required. For this reason, depending on the context, the biometric system must be
calibrated according to the desired level of security. One example is credit card
applications where it is important to keep the false reject rate low since it will not be
convenient for a legitimate user to fail the authentication process when the system isn’t
properly calibrated.
In a biometric system, the probability that a genuine person is rejected is called false
rejection rate (FRR), while false acceptance rate (FAR) is the probability that an impostor
is accepted as legitimate (Figure 2.2). A receiver operating characteristic (ROC) curve
plots the false reject rate (FRR) against the false accept rate (FAR) where FRR represents
the percentage of genuine scores that do not exceed the threshold while FAR corresponds
to the percentage of impostor scores that exceed the threshold chosen depending on the
nature of the application (Figure 2.3). The point where FRR and FAR are equal is called
the equal error rate (EER). The EER of a system can be used to give a threshold
independent performance measure. The lower the EER is, the better is the system's
performance, as the total error rate which is calculated as the sum of the FAR and the
FRR at the point of the EER decreases. It is also important to mention that generally, in
watch list applications, it is preferable that the biometric system produce a low FRR
while in a high security context, the objective is to obtain a low FAR.
Besides these types of errors, in some cases, some individuals cannot provide good
biometric data (i.e. poor quality fingerprint ridges) since they do not have the biometric
feature from which there can be produced repeatable templates. The expected proportion
of the population for whom the biometric system is unable to obtain good templates is
called the failure to enroll rate (FTE). Similarly, a system may also be unable to capture
26
or locate an image of sufficient quality (Mansfield et al. (2002)). For example, this could
be because of worn, cut or unrecognizable prints as well as the quality of the captured
image is inadequate. In this case, the expected proportion of cases that failed to provide
good features is called the failure to acquire rate (FTAR).
Genuine
Distribution
Impostor
Distribution
Figure 2.2: Evaluation of the matching accuracy of a biometric system. Histograms of the genuine
and impostor matching scores are represented as well as the two types of errors that can arise in a
biometric system given a matching score threshold (T). The areas A and B represent false accept rate
(FAR) and false reject rate (FRR), respectively.
27
Commercial
applications
Figure 2.3: Receiver operating characteristic curve (ROC) showing the relation between false
acceptance and false rejection in a biometric system.
2.2 Multimodal biometric systems
Most biometric systems deployed in real-world applications are unimodal since they rely
on a single source of information for authentication (e.g., single fingerprint, iris or face).
These systems suffer from different problems such as the presence of noisy data resulting
for example from defective sensors, unfavorable ambient conditions, poor illumination,
incorrect facial pose and many others. Some of these limitations can be overcome by
including multiple sources of information for establishing identity. Those systems are
called multimodal biometric systems and are expected to be more reliable due to the
presence of multiple and independent biometric features. These systems can meet the
severe performance requirements imposed by various real world applications by
achieving higher accuracy and improved system’s performance (Kuncheva et al. (2000)).
28
Also, since a single biometric feature can sometimes lead to type I and type II errors as
well as higher failure to acquire (FTA) and failure to enroll (FTE) rates, a multimodal
biometric system tends to be more reliable for the same application. For instance,
fingerprints can be copied (Sandström (2004)) or altered by cuts and bruises (Jain et al.
(2004)), a face recognition algorithm can result in too many false acceptances (Kosmerlj
(2004)) and many other drawbacks exist in other biometric authentication schemes (Jain
et al. (2004)). For this reason, a multimodal biometric system (Jain et al. (2004))
combines the use of more than one biometric feature to solve these issues. For example,
the system can solve the problem of distinguishing between people with similar faces (i.e.
identical twins) by using fingerprints as an additional biometric feature, while at the same
time issues caused in a fingerprint biometric system by people having worn fingerprints
and people missing fingers are handled by using face recognition. Furthermore, a
multimodal system also provides anti-spoofing measures (Jain et al. (2004), Schuckers
(2002)) by making it more challenging for an impostor to fool the system.
2.3 Properties of Biometrics
In
order
for
a
Biometric
system
to
achieve
good
performance
at
the
enrollment/authentication and at the matching levels, the following properties of
biometric characteristics must be met:
•
Invariance: The biometric characteristics should remain constant over a long
period of time. This eliminates the need for updating the biometric feature
29
templates stored in the database, reduces the complexity of the system and
improves the recognition rate. For example, facial characteristics may change
over time due to aging while iris features remain constant throughout a person’s
lifetime (Jain et al. (2004)).
•
Measurability and Timeliness: The process for extracting biometric samples
should be simple and fast. This is very important for continuous authentication in
applications where real-time identification is necessary. Extracted features in a
biometric system must be measureable in order to automatically compare them to
an expected norm. For example, in airports, biometric samples must be taken at a
distance and computation must be done rapidly as subjects are walking by the
gate.
•
Singularity/Uniqueness: Biometric characteristics should have sufficient unique
properties for different individuals in order to distinguish one person from
another. This is true for all biometrics except that some of them, present more
unique and accurate features compared to others (i.e. iris contains more
information compared to hand geometry).
•
Reducibility: In a biometric system, the extracted feature templates should be
reduced in size for handling and storage purpose as long as information is
impossible to duplicate. This becomes a crucial property especially when
30
information is transmitted across secure channels when the controller of the
results is located in a remote area.
•
Reliability: The biometric algorithm should ensure high reliability and integrity.
It becomes very inconvenient and costly when a biometric system does not
provide consistent results.
•
Privacy: In a biometric system, the privacy of the individual cannot be violated in
any way. Information about the identity of a person cannot leak out of the system
and should remain confidential, otherwise, people will hesitate to use it.
All the properties mentioned above are important to all biometric characteristics (e.g.
face, iris, fingerprint, hand geometry, etc…) and must be met in a biometric system in
order to be able to provide an accurate way of authentication.
2.4 Classification of Biometric System
Biometric systems can be classified according to six perspectives (Dessimoz et al.
(2005)) as follows:
•
Overt / covert: An application is defined as overt if and only if the user is aware
about the acquisition of his biometric data otherwise, the application is said to be
covert which is in fact one of the most concerning public perception of a
31
biometric system since it is related to a privacy issue. In overt applications, there
are fewer concerns with data acquisition and sample quality since images are
taken in a controlled environment where subject is asked to cooperate during the
data acquisition process. On the other hand, in a covert environment such as at an
airport checkpoint where face images of passengers are captured and compared to
a watch list without their knowledge, the quality of the captured images can be
problematic since they are taken in an uncontrolled environment without any user
cooperation.
•
Attended / non-attended: The biometric recognition process is performed
attended if the user is observed and guided by supervisors during the process. On
the other hand, if the process is unsupervised, the process is considered to be nonattended. In attended applications, biometric samples tend to be of a better quality
compared to the ones acquired in a non-attended system where subject
cooperation is non-existent.
•
Standard / non-standard environment: A standard environment involves that
most conditions in the biometric system are controlled and the recognition takes
place indoors within a constrained environment, otherwise, the use is called in
non-standard environment. For example, customs and airport security systems are
considered standard since the entire biometric recognition process is completed in
a controlled environment.
32
•
Habituated / non-habituated: If the majority of the users interact with the
biometric system on a frequent/daily basis then the recognition is said to be
performed in the habituated mode. On the other hand, when the system’s usage
frequency is low, the recognition is performed in the non-habituated mode. This is
relevant to the degree of cooperation and training demanded from the users.
•
Public / private: The application is considered to be public if the users are not
employees or work within the association that owns the biometric recognition
system. If the users are the employees, the application is then called private. An
example of a private application is internal bank security where employees are
asked to voluntarily present their biometric traits for authentication.
•
Open / closed: The biometric application is considered to be closed if the system
uses completely proprietary formats. On the contrary, when the system is able to
exchange data with others, the system is then called open and privacy issues
should be addressed properly.
Based on the above description of a biometric system, we classify the systems considered
by this thesis as covert, non-standard, non-habituated, non-attended, public and open
biometric recognition systems.
33
2.5 Biometric Sample Quality Measures
Biometric systems are designed to identify a person based on physiological or behavioral
characteristics (Wayman (2001)). Some of the most popular biometric systems presently
used include automatic fingerprint, iris, and face recognition systems. Currently, these
systems are seeing an increasing level of interest in a wide variety of applications such as
in national identification applications, criminal searches, airport security and for access
control. Lately, a significant level of interest is seen in the development of standards for
measurement of biometric quality. According to (ISO (2007)), biometric sample quality
may be considered from the point of view of character (inherent features), fidelity
(accuracy of features), or utility (predicted biometrics performance). In general, a good
biometric quality measure should demonstrate that images evaluated as higher quality
must be those that result in better identification of individuals, as measured by an
increased separation of genuine and impostor match score distributions.
In addition, the genuine and impostor distribution separation can be improved through the
use of a prior knowledge in large-scale biometric recognition systems. For example in
face recognition, knowledge such as demographics, face image resolution, pose,
expression, and eye/face detection confidence carry useful information that can improve
biometric performance. Using the a priori knowledge, new sets of localized genuine and
impostor distributions can be extracted out of the overall data. The Receiver Operator
Characteristics (ROC) Curve from the localized genuine and impostor distributions can
be used to determine its own optimal threshold. These local thresholds can then be
combined to improve upon the optimal threshold of the generalized ROC curve.
34
Therefore, the use of quality measures should help address issues in biometric
performance improvement encountered in large-scale biometric recognition systems. In
today’s literature, few works exist on automatic face quality measures. For example, a
method described in (Kryszcuk et al. (2005)) is developed where poorly illuminated face
image regions are segmented using statistical methods and the remaining face area left
after segmentation is used as a quality measure to find the optimal decision threshold.
Another method developed towards the use of quality measures for face verification is
presented in (Kryszcuk et al. (2006)) where two face quality measures are used as
evidence in the process of reliability estimation. The quality measures developed in that
work are based on image contrast and normalized 2-D correlation with an average face
template.
In section (5.3) of this thesis, a new biometric quality measure is proposed based on a
new information theoretic framework. The new approach is developed based on the
intuitive observation that a high quality biometric image is believed to have more useful
information in identifying an individual compared to a low quality image. This suggests
that the quantity of identifiable information decreases with a reduction in quality. Given a
way to measure the decrease in information caused by a given image degradation, one
can measure the associated decrease in biometric information.
2.6 Face recognition
Face recognition is one of the most common methods used for identifying humans due to
its non-intrusive nature, as acquiring face images can be done at a distance. Recognizing
35
faces and facial expressions is becoming very important in many practical applications,
such as in border control and airport security. The process flow in face recognition
consists of four phases: capture of samples, feature extraction, template creation and
matching. Several techniques exist for face recognition. Some techniques can be more
suited than others depending on the application. Recent surveys and reviews on face
recognition technologies are provided in (Zhao et al. (2003), Kong et al. (2005), Li and
Jain (2005)). Some of the most popular face recognition techniques are the Eigenfaces
(Turk et al. (1991)), local feature analysis (Chirillo et al. (2003) and Elastic Graph
Matching (Zhang et al. (1997)).
Eigenfaces method is used to efficiently represent facial images using principal
component analysis (PCA) where a set of images is represented using a set of
orthonormal basis vectors. Each eigenface is derived from the covariance analysis of the
face image population. A similarity measure is then created in order to measure
resemblance between faces. Two faces are considered to be similar if the distance
between their feature vectors is small in the eigenface feature space. The mathematical
properties of the eigenface representation and the matching process have shown to
perform well on face images taken in a certain controlled environment (Zhao et al.
(2003)). Like most facial recognition techniques, the eigenface methods perform better in
well-lit, frontal image capture scenarios (Turk et al. (1991)).
Local feature analysis (LFA) is a very popular method used in face recognition. It has the
ability to accommodate for changes in facial expression and aging (Chirillo et al. (2003)).
In LFA algorithms, a set of features such as geometrical metrics and distances are derived
from the facial images and used as the basis for image representation and comparison. In
36
general, the eyes, nose, mouth, jaw line and cheeks are the most common used features in
these methods and are usually represented by their position and dimensions. This
technique performs generally well except that it has few drawbacks such that it highly
depends on the environment where pictures are taken and on the quality of the images.
Another method used in face recognition is Elastic Graph Matching. This technique is
known for its ability to provide face recognition that is invariant to affine transformations
and changes in facial expressions (Wayman et al. (2005)). In this method, features are
locally extracted at specific locations in the face image. Afterwards, the distances
between these nodes are recorded for further processing. Based on the application that
this method is used for, some features (i.e. nodes) are more reliable and important for
face recognition than others. For this reason, the use of weights has been introduced in
(Duc et al. (1999)) where more important features are assigned higher weights. Zhang et
al. (1997) developed an extension of the Elastic graph Matching method where several
images of the same individual taken at different angles are used for feature extraction.
This technique (Elastic Bunch Graph Matching) showed better results because it is more
robust to differences in posture and facial expressions.
Today, face recognition technology has evolved and it is used on three-dimensional face
images (Philips et al. (2003)). This leads to significant improvement compared to the
two-dimensional face recognition technology. A common drawback with the 2D face
recognition technology is the inability for algorithms to accommodate for changes in
illumination and pose (Pan et al. (2003)). In (Pan et al. (2003)), the developed 3D face
37
recognition method performed well for images taken under different pose lighting
conditions however, the method doesn’t handle variations in facial expressions very well.
The work presented in this thesis develops new methods and solutions to improve the
performance of a face biometric system when dealing with low quality images. A variety
of algorithms are developed to deal with face detection, face tracking, feature extraction,
and pattern comparison using low quality face images. The developed methods make use
of two different face recognition algorithms based on PCA (Eigenface) and Fisher Linear
Discriminant (FLD) feature decompositions in order to measure biometric feature
information (section 5.2). A definition of biometric feature information is introduced and
an algorithm to measure it proposed, based on a set of population and individual
biometric features. In addition, section (5.2.6) develops a new approach to understand
and measure the changes in biometric sample quality resulting from image degradations.
We begin with the intuition that degradations to a biometric sample will reduce the
amount of identifiable information available. We then show that the biometric
information for a person may be calculated by the relative entropy theory between the
population feature distribution and the person’s feature distribution and that the biometric
information for a system is the mean for all persons in the population. In order to do this,
we simulate degradations of biometric images and calculate the resulting decrease in
biometric information. Results show a quasi-linear decrease for small levels of blur with
an asymptotic behavior at larger blur (section 5.3).
38
2.7 Face tracking algorithms
Face tracking is a crucial part of most face processing systems. It requires accurate target
(i.e. face) detection and motion estimation when an individual is moving. Generally, this
process is required to facilitate the face region localization and segmentation necessary
prior to face recognition. Accurate face tracking is a challenging task since many factors
can cause the tracking algorithm to fail. Some of the major challenges encountered by
face tracking systems are robustness to pose changes, lighting variations, and facial
deformations due to changes of expression and face occlusion. These factors might cause
the algorithm to lose track of the subject’s face and drift (i.e. lose face detection for
initialization). Face tracking techniques can be classified into two categories: tracking
cues and motion estimation (Li et al. (2008)). Tracking cues include knowledge-based
and learning-based techniques. The knowledge-based methods use prior information
about the face area such as face contour, intensity, shape and face models in order to
locate the face region in an image. These methods perform generally well, however, they
seem to fail in situations where the subject’s face is occluded by other objects, especially
those of circular shapes. On the other hand, learning-based techniques attempt to model
the face pattern using distribution functions in a probabilistic framework. These methods
perform better than the knowledge-based techniques since they are not limited to the
prior knowledge on face instead, they are known for their capability of developing
learning models using training sets which perform better in challenging situations (Li et
al. (2008)).
39
2.7.1 Knowledge-Based Classifiers
The following are some knowledge-based (i.e. tracking cue) methods applied for face
tracking.
2.7.1.1
Face color model
In face tracking, using skin color in knowledge-based techniques provides important
information that facilitates target detection based on the analysis of the previous image
frame. Color information is a solid feature due to its robustness to image scaling,
translation, rotation and stretching.
HSV color space model
HSV (Hue/Saturation/Value) space separates out hue (color) from saturation (color
concentration) and from brightness (value) (Bradski (1998)). HSV can be thought of as
describing colors as points in a cylinder (called a color solid) whose central axis ranges
from black at the bottom to white at the top, with neutral colors between them. The angle
around the axis corresponds to “hue”, the distance from the axis corresponds to
“saturation”, and the distance along the axis corresponds to “lightness”, “value” or
“brightness”. Color models are created by taking 1D histograms from the H-channel in
HSV space. For face tracking via a skin color model, face areas are sampled by
prompting users to center their face in an onscreen box to find skin areas for further color
sampling. The hues derived from skin pixels in the image are sampled from the Hchannel and binned into a 1D histogram. Once sampling completed, the histogram values
are stored for future use. More robust histograms can be developed by sampling skin
40
hues. During operation, the stored skin color histogram is used as a model, or lookup
table, to convert incoming video pixels to a corresponding probability value. This process
is done for each video frame. Using this method, probabilities range in discrete steps
from zero (probability 0.0) to the maximum probability pixel value (probability 1.0). For
8- bit hues, this range is between 0 and 255 (Dansheng (2009)).
Stochastic skin-color model
Skin color can be represented in the chromatic color space (Gejgu et al. (2003)).
Chromatic colors (
), also known as pure colors in the absence of brightness are
defined by the following normalization process:
(2.1)
(2.2)
Blue color is redundant after the normalization because
. In (Kjeldsen et al.
(1996)) it has been shown that chromatic skin-color has normalized Gaussian
distribution. Therefore, a face color distribution can be represented by a Gaussian model,
where
with
(2.3)
(2.4)
and
41
(2.5)
where
and
represent the mean and standard deviation, respectively.
In order to train the skin color model, a set of hand segmented images are required. From
this, the likelihood of skin color for any pixel in the face image can be obtained. For
example, the likelihood of skin color for a certain pixel transformed from RGB color
space to chromatic color space with chromatic pair value
can be computed as
follows:
(2.6)
where
. From this, a color image can be transformed into a grey scale image
using the skin color model. Each pixel grey value represents the likelihood of the pixel
belonging to the skin.
2.7.1.2
Texture model (LBP Cue)
Ojala et al. (1996) proposed a local binary pattern (LBP) operator as a powerful tool for
describing image textures. The LBP operator applies a
mask to the image from
which a binary number is computed by applying a threshold to the neighboring pixels
(Dansheng (2009), Wang et al. (2008)). Once the threshold values are computed for
every pixel in the image, the histogram of the labels is then used as a texture descriptor.
One limitation to the LBP operator is the fact that it cannot capture larger texture in an
42
image due to the fact that it is limited to a
pixels area. Therefore, the LBP operator
must be extended to different sizes and the use of circular masks becomes necessary.
2.7.1.3
Edge cue
Edged based models are other appearance-based classifiers that use face contours and
local structures (i.e. eyes, nose, mouth contours) in a face image (Swaminathan et al.
(2007)). In general, contours are extracted using the Difference of Gaussian (DOG)
filters applied to the grayscale image. One limitation of these methods is the fact that they
might fail when subject’s face is occluded.
2.7.1.4
Facial shape
Ellipse fitting algorithms are used for face image classification (Gejgu et al. (2003)). In a
face image, only blobs with elliptical shapes are retained for further processing. In
general, face detection is applied on the first frame in a sequence of images and then
blobs corresponding to faces are tracked further. Ellipses are classified using their center
of mass computed as follows:
(2.7)
(2.8)
43
where
is an
image with area
in pixels. Elliptical contour fitting can be applied
on images where faces are inclined. In order to improve the fitting accuracy, better results
are obtained by first rotating the face prior to the fitting operation. The inclination angle
is computed as follows:
(2.9)
where
(2.10)
(2.11)
(2.12)
and
(2.13)
(2.14)
Also, the moment of inertia can be used to compute the length of the major and minor
axis of the best-fitted ellipse.
(2.15)
(2.16)
44
where
and
are the least and the greatest moment of inertia of an ellipse with
orientation . From this, the length of the major axis ( ) and the length of the minor axis
( ) are computed as follows:
(2.17)
(2.18)
2.7.1.5
Motion segmentation
Also known as “Background subtraction” method, motion segmentation is performed
using a color change detector. This technique works assuming that a reference
background image is available for the segmentation to succeed. Some motion
segmentation methods use an adaptive background algorithm (Li et al. (2008)).
2.7.2 Learning-Based Classifiers
In this section, we review relevant learning-based (i.e. tracking cue) techniques used for
face tracking.
45
2.7.2.1
Haar-like features
Learning-based techniques attempt to model the face pattern using distribution functions
in a probabilistic framework. The Haar-like features, originally proposed by
Papageorgiou et al. (1998), are used to represent face features. Three types of rectangular
features exist that can be computed. For example, a rectangular feature can be computed
by taking the difference between the sums of two rectangular blocks. A fast technique
was proposed by Viola-Jones for computing these features used for face detection (Viola
et al. (2001)). Scanning through the picture, their one-pass method uses an intermediate
array to store a running sum of pixel above and to the left of the point
:
(2.19)
where
is the integral image, and
is the original image. Using
expression (2.19), the Haar features are easily computed. For example, Figure 2.4 shows
the rectangular region D that is computed by
(Viola
et al. (2001)).
46
Figure 2.4: Integral image used to calculate the sum of pixels in a rectangle. For example, the value of
the integral image at location 1 is the sum of the pixels in rectangle A. The value at location 2 is A+B,
at location 3 is A+C, and at 4 is A+B+C+D. The sum within D is computed as (4+1)-(2+3).
2.7.3 Motion estimation
Motion estimation is another category of face tracking techniques, other than tracking
cues (section 2.7.1), which deals with the major issues that occur in tracking algorithms.
In face tracking, the main research problem is “Drifting”. This issue is caused by the
following: 1) rapid variation in appearance of the target; 2) abrupt movement and poor
motion continuity. The first problem can be solved by creating proper tracking cue
models while the second issue entirely depends on the robustness of the motion
estimation algorithm. Dansheng (2009) classifies motion estimation algorithms as: 1)
Kernel-based trackers and 2) Bayesian filter estimators. The latter is much more accurate
and robust compared to the Kernel-based tracker and is considered the most suitable for
low frame rate video where severe “drifting” occurs.
47
2.7.3.1
Mean shift tracking
The mean shift tracking algorithm is a nonparametric statistical method that seeks the
nearest mode of a point sample distribution. In the mean shift tracking algorithms, a color
histogram is used to describe the target region. The Kullback-Leibler divergence,
Bhattacharyya coefficient and other theoretic similarity measures are commonly
employed to measure the similarity between the template region and the current target
region (Comaniciu et al. (2003)). Tracking is accomplished by iteratively finding the
local minima of the distance measure functions using the mean shift algorithm.
In Dansheng (2009), a kernel is defined as a function
function
if there exists a
called its profile, which is nonnegative, non-increasing,
piecewise continuous and with,
such that
where
is
an -dimensional Euclidean space.
Let
be a finite set called the sample data,
weight function. The sample mean with kernel
a kernel and
at a point
a
is defined as
(2.20)
From this, the mean shift is defined as
and the mean shift algorithm represents
the repeated movement of data points to the sample mean. Let
be a finite set, and
. The full mean shift procedure iterates and evolves
until it
48
finds a fixed point
or re-evaluated. In addition,
. After each iteration, the weights
and
can be either fixed
are usually different sets, with
a fixed set of
samples. Here are the steps to compute the Mean Shift Algorithm (Dansheng (2009)):
i.
Select a specific search window size.
ii.
Select the initial location of the search window.
iii.
Compute the mean location in the window.
iv.
Center the search window at the mean location computed in step (iii).
v.
Repeat steps (iii) and (iv) until the algorithm converges or until a predefined
threshold is met.
One limitation of the mean shift algorithm is the fact that the original formulation of this
method does not estimate the orientation of the tracked region. In addition, mean shift
tracking uses fixed color distribution which can become an issue in some applications
since color distribution can change due to rotation in depth. However, this issue is
resolved used continuous adaptive mean shift (CAMSHIFT) (Bradski (1998)) which
handles dynamically changing color distribution by searching the window size and
computing color distribution for each search window.
2.8 Iris Recognition
Iris recognition is another highly studied and evolved technology in biometrics. The iris
is known to contain a rich texture which means that unique information can be extracted
from the iris to identify users. It has been shown that iris features have been used to
49
obtain high recognition accuracy for security applications (Bonney et al. (2005),
Daugman (1993), Newton et al. (2007)). Even though iris recognition has shown to be
extremely accurate for user identification, there are still some issues remaining for
practical use of this biometric (Wayman et al. (2005)). For example, the fact that the
human iris is about
in diameter makes it very difficult to be imaged at high
resolution without sophisticated camera systems. Traditional systems require user
cooperation and interaction to capture the iris images. By observing the position of their
iris on the camera system while being captured, users adjust their eye positions in order to
localize the iris contour accurately (Newton et al. (2007), Wildes (1997)).
This step is crucial in iris recognition since iris features cannot be used for recognition
unless the iris region is localized and segmented correctly. Many iris localization
techniques exist and have been developed. Some of the classical methods for iris
localization are Daugman’s integro-differential operator (IDO) (Daugman (1993)) and
Wildes’ Hough transform (Wildes (1997)).
Figure 2.5 illustrates the major stages of an iris recognition system (Proenca (2006)). The
initial stage involves segmenting accurately the iris area from an eye image. This process
consists in localizing the iris inner and outer boundaries, assuming they have circular or
elliptical shapes. This process also requires detecting and removing any eyelash noise
from the image prior to segmentation. In order to compensate the variations in the pupil
size and in the image capturing distances, the segmented iris region is mapped into a
fixed length and dimensionless polar coordinate system (Daugman (2003)). In terms of
feature extraction, iris recognition approaches can be divided into three major categories:
50
phase-based methods (Daugman (1993)), zero-crossing methods (Boles et al. (1998)) and
texture analysis based methods (Wildes (1997)). Finally, the comparison between iris
templates is made, and a metric is measured. If this value is higher than a threshold, the
system outputs a non-match, meaning that each signature belongs to different irises.
Otherwise, the system outputs a match, meaning that both templates were extracted from
the same iris.
Enhanced Iris
Segmentation
Iris image
Iris
Normalization
Segmented iris image
Biometric template
10000101010100
10100001100101
Feature / Pattern
Comparison
Feature
Extraction
Normalized iris image
HD / Similarity Value
Enrolled
Signatures
Figure 2.5: Different stages in an iris recognition system
51
2.8.1 Iris structure
The iris has a multi-layered structure where the most posterior layer is comprised of
epithelial cells containing iris pigments and the anterior layer is comprised of two sets of
muscles, the sphincter muscle and the dilator muscle. The sphincter muscle serves to
contract the pupil and the dilator to open it. Further into the anterior side is the stromal
layer composed of collageneous tissues which generates the most part of an iris image.
2.8.2 Iris texture pattern and colors
The iris texture features are composed of various components within the iris such as
crypts, furrows, arching, collarette and rings of various shapes (Figure 2.6). In addition,
the iris pattern comes in different color pigments, such as green, red and blue. The natural
iris colors are generated mostly from a combination of melanin pigments, mostly in the
anterior layer and stromal layer. Visible light goes through the iris, and the absorption
and reflection of light depend on the wavelength. The iris color is further determined by
the variation in the pigment density and the amount of reflected light.
In iris recognition, near infrared (NIR) cameras are usually used instead of visible light
which tend to emphasize the texture pattern of the iris, especially for darker regions. The
NIR camera uses a wavelength spectrum that ranges from
to
(Iridian
Technology 2005). In addition, the fact that NIR illumination is a more comfortable
imaging modality for subjects compared to regular light source makes NIR imaging
widely used in the iris recognition field.
52
The iris contains a unique and rich texture which can be used in high security
applications. The unique and abundant texture patterns in the iris images are "determined
epigenetically by random events in the morphogenesis process" (Daugman et al. (2001)).
Cross-comparisons were performed between genetically identical images, taken from the
left and right eyes of the same persons and showed that the statistical distributions were
the same for iris images coming from genetically related and genetically unrelated
subjects. Nevertheless, in order for any emerging biometric technology to be accepted by
the community, an independent party is required to perform the evaluations, design the
protocols, collect the data sets, supervise the tests and analyze the results (Philips et.
(2000)). There exist many accepted standards and frameworks for how to design a
biometric testing protocol, how to collect data sets and minimize evaluation bias
(Mansfield et al. (2002)). In general, the FNMR, FMR and the detection error rate curve
(DET) are used as indicators of the level of accuracy of a typical iris recognition system.
For this reason, the International Biometric Group (IBG), the Authenti-Corp and the
National Institute of Standards and Technology (NIST) performed independent
evaluations with various iris recognition systems to analyze its accuracy and performance
levels.
Moreover, a project conducted by the University of Cambridge and the United Arab
Emirates interior ministry showed that for
and over
iris images acquired in the Middle East
billion cross-comparisons generated between different eyes that a false
match rate (FMR) of less than
in
billion was achieved using the Daugman
algorithm (Daugman (2005)).
53
One experiment named ITIRT was funded by the US Department of Homeland Security
for border control and security access consulting. The experiment was performed in July
2004 on several state-of-the-art iris recognition systems such as: “Iridian KnoWho OEM
SDK”, “LG IrisAccess 3000”, “Oki IRISPASS-WG” and “Panasonic BM-ET300”
(International Biometric Group (2005)). In this experiment, over 100,000 iris images
were acquired at different times, with different devices. These images were taken from
1224 people of different ethnic cultures and age groups. The image templates were
compared in order to generate the false accept rate, the false reject rate, the failure to
enroll and the failure to acquire rates. For the feature extraction and matching algorithms,
the generic version of the Daugman algorithm was implemented. The same algorithm is
also used to test all of the four systems. As an indicator of error rates, the FNMR at FMR
of 0.001 was used. The Panasonic BM-ET300 module achieved a FNMR of around
0.014. The Oki system achieved a FNMR of around 0.03 at an FMR of 0.001. The LG
unit achieved a FNMR of around 0.038 (Newton et al. (2007)). This shows that all the
tested iris recognition systems achieve very high identification rates, a strong interoperability and repeatability.
54
Eyelashes
Eyelid
Pupil
Iris
Crypts
Furrows
Collarette
Figure 2.6: Example iris image taken from CASIA database.
2.8.3 Imaging Systems
Two well-known iris imaging algorithms developed by John Daugman and the Richard
Wildes group are currently used (Daugman (2001), Masek (2003), Wildes (1996, 1997)).
Both the Daugman and the Wildes techniques achieve good identification rates by using
monochrome gray-scale images. The main difference between the two imaging systems is
the lighting source implementation where the Daugman system implementation consists
of a lateral light source, while the Wildes system applies a diffuser to illuminate the entire
eye region (Figure 2.7, Figure 2.8).
The Daugman system is shown in Figure 2.7 which demonstrates the position of the light
source to the side of the eye. The camera captures the reflected light after passing through
a beam splitter. The resulting iris image has a diameter ranging from
to
pixels,
55
which gives enough information for iris recognition. On the other hand, the Wildes
imaging system applies a series of light sources, illuminating the iris region through a
diffuser and a circular polarizer (Figure 2.8). Using this system, the captured iris image
has a diameter of around
pixels. The Wildes system generates iris images with
reduced specular reflections compared to a single light source system since it uses an
evenly distributed light illumination system (Wildes et al. (1996)).
Figure 2.7: The Daugman iris imaging system
56
Figure 2.8: The Wildes iris imaging system
2.8.4 Iris Localization and Segmentation
Iris localization and segmentation is a crucial step in iris recognition since it severely
affects the system’s performance. This section of the algorithm consists in segmenting
the specific iris region from an eye image by locating the exact iris boundary, the pupil
region, and the upper and lower eyelids. In some cases, artifacts can be found in the
resulting iris image which can be a combination of eyelash occlusion, eyelid occlusion
and/or noise. Advanced algorithms are required for successful removal of these artifacts
in order to generate a clean iris region for subsequent recognition. In fact, various
methods have been proposed to identify and eliminate artifacts in iris images, particularly
by detecting and removing eyelash occlusion and eliminating specular reflections. In
general, most algorithms perform reasonably well except that they tend to overestimate
57
eyelash occlusion (Xie (2007)). The following section describes few iris-pupil
segmentation techniques widely used in some iris recognition systems.
2.8.4.1
Pupil and Iris Localization
The integro-differential operator, proposed by Daugman, locates the pupil, the iris inner
and outer boundaries as well as the upper and lower eyelid boundaries (Daugman
(2004)).
(2.21)
(2.22)
where
represents the eye image,
circle of radius and center coordinates
Gaussian function with center
are parameters that correspond to a
, respectively.
and standard deviation
is a radial smoothing
. One application of this
function is that it searches the entire eye image for integrations along different circular
contours with center coordinates
and an increasing radius . The maximum
contour integral derivative found will then be classified as the most likely circle tracing
the iris. In a similar manner, the circular boundaries for the pupil and iris regions are
localized by searching through the entire iris image for the maximum integration along
various circular contours. In addition, Daugman approximates the upper and lower
eyelids with two open curves that are part of two different circles. Finally, the iris region
58
surrounded by the upper and lower eyelids as well as the extracted circular pupil and iris
boundaries are used for further feature extraction in the iris recognition process.
2.8.4.2
Hough transform
The circular Hough transform is a standard technique used in the machine vision field to
locate circular contours in images. The Hough transform is applied directly on an
intensity gradient edge map usually obtained through a gradient-based edge detector
(Wildes (1997)). The latter method is used in many iris recognition algorithms such as
the Wildes system (Wildes (1997)). First, the entire iris image
Gaussian filter
with centers (
) and a standard deviation
2.24). Then, the intensity gradient image map
image
is smoothed with a
of
(equation
is generated from the smoothed
, as shown in equation (2.25) using the gradient operation defined in (2.26).
Subsequently, the binary edge map is generated by setting a threshold on the intensity
gradient image
. The threshold is usually selected based on experimental data and
depending on the application. Finally, using the binary image map, the Hough transform
is performed to locate a circle with the largest number of edge points and with circular
parameters
calculated as shown in equation (2.27).
represents a circle
to be located within the iris image such that the circle is characterized by a radius , and
center coordinates
with possible edge point
.
(2.23)
(2.24)
59
(2.25)
(2.26)
(2.27)
From this, the Hough transform is then performed through the entire collection of the
edge points. Whenever equation (2.27) is satisfied, it means that the circular contour goes
through
, and one extra vote is added to the histogram count for possible
circular contours. Once the entire image is scanned for all possible contours, the contour
that obtained the highest amount of votes represents the most likely circle in the edge
map.
2.8.4.3
Discrete Circular Active Contour Model
The discrete circular active contour (DCAC) model can also be used to locate the pupil
and iris boundaries in the iris image (Ritter et al. (1999)). First, in order to localize the
pupil region in the iris image, the variance image must be computed from the original
image and then, an active contour model with a starting point in the center of the pupil, is
initiated and moved within the iris image under the influence of so called “internal and
external forces”. Along the active contour, the vertex
moves from time
to time
according to:
(2.28)
where
represents the position of the vertex at a specific time ,
and
represent the
internal and external forces, respectively.
60
The internal force is characterized by the continuity, and other prior knowledge about the
iris (i.e. circular contour) while the external force is directly related to the gray-scale
intensity values within and outside the vertex, which also includes the iris region. Finally,
the iris contour is segmented after an extensive iterative contour searching operation
which ends when equilibrium with minimum energy or minimum mean variance of the
annulus is attained. DCAC has shown good results but do suffer from few limitations
such that the performance of this method greatly depends on the iris image quality. For
instance, if the image contains severe noise, specular reflections or distortions, the
method will fail in locating the proper boundaries (Ritter et al. (2003)).
2.8.4.4
Noise and artifacts in iris images
In order to achieve higher system performance and better accuracy in the image
processing steps within an iris recognition system, noise and artifacts in iris images must
be reduced or if possible, completely eliminated. Such artifacts include the eyelash
occlusion, the eyelid occlusion and specular reflections. As seen in Figure 2.6, the iris
image includes severe eyelash and eyelid occlusion. Hence, a part of the iris is covered
by the top eyelid and some eyelashes are spread across the iris area which will affect the
system’s performance.
In addition, specular reflections are mirror-like reflections that occur during the iris
image acquisition procedure in such a way that the light source gets reflected and imaged
by the camera. Figure 2.9 shows one example of specular reflection seen on the iris
region as a “white spot” which results in high pixel values that deviate from the original
iris patterns and that constitute a major source of distortion.
61
In order to address all these issues with image artifacts, a new method is developed in
chapter 4 to detect and eliminate noise in iris images. This will then generate better
results when properly segmented iris images are used for further processing in the iris
recognition system.
Figure 2.9: Iris image showing severe specular reflections taken from Montgomery (2007)
62
The following table shows few examples of iris noise images (Proenca (2006)).
Type of iris noise
Addressed in this
thesis
Eyelid and eyelash
occlusions
YES
Lighting reflections
YES
Out-of-focus
YES
Examples
63
Off-angle iris
NO
Out-of iris images
YES
Motion blurred irises
NO
2.8.5 Size-invariant Unwrapping and Representation
This section describes a method to normalize the iris region once it is properly segmented
in order to obtain a size-invariant rectangular representation of the original iris pixels.
This method, referred to as “Daugman’s Rubber Sheet Model”, is developed by Daugman
(2003) to map the sampled iris pixels from the Cartesian coordinates to the normalized
64
polar coordinates in order to accomplish a size-invariant sampling of the original iris
points.
r
r
Figure 2.10: Unwrapping of the Iris using Daugman's Rubber Sheet Model
The Daugman’s rubber sheet model finds for every pixel in the iris, an equivalent
position on the polar axes
where
is the radial distance and
is the rotated angle at
the corresponding radius. The radial resolution is described as the number of data points
in the radial direction while the angular resolution is the number of radial lines generated
around the iris region. Using equation (2.29), the iris region is transformed to a 2D array
with horizontal dimensions corresponding to the angular resolution and the vertical
dimension to radial resolution (Figure 2.10).
(2.29)
where
corresponds to the iris region,
normalized polar coordinates, respectively.
and
and
ranges from
are the Cartesian and
to
and
from
to .
are defined as linear combinations of pupil boundary points. The
following equations perform the transformation:
65
(2.30)
(2.31)
(2.32)
(2.33)
(2.34)
(2.35)
where
and
respectively.
represent the pupil and iris coordinates along the
and
direction,
correspond to the pupil and iris center coordinates.
Two different cases can occur with the rubber sheet model after mapping the iris region
from the circular Cartesian scale to the rectangular polar scale. First, if the pupil and iris
boundary centers are located at the same pixel point, the sampled points are uniformly
distributed across the iris region, as shown in Figure 2.11. On the other hand, if the center
of the iris circular boundary and the center of the pupil circular boundary are different,
the feature points are then sampled non-uniformly within the iris region. In order to deal
with this situation, a series of sampling lines are emitted from the center of the pupil
circle, and rotated along the circumferential direction for
. Afterwards, a fixed
number of sampled points are taken inside the iris region along each sampling line, as
shown in Figure 2.12.
66
Figure 2.11: Uniform feature points sampling with the Daugman's rubber sheet model.
Figure 2.12: Feature points sampling with displaced pupil and iris centers.
2.8.6 Feature Extraction
The iris has a particularly interesting structure and provides abundant texture information.
Iris feature extraction is applied on the grayscale image intensity values after the iris
67
segmentation and noise removal steps. Feature extraction is a crucial part in any iris
recognition system since good identification rates are directly related to the uniqueness
and variability of the extracted features used to distinguish between different biometric
templates.
2.8.6.1
2D Gabor features
Gabor filter based methods have been widely used in computer vision, especially for
texture analysis. A Gabor filter is constructed by modulating a sine/cosine wave with a
Gaussian (Daugman (2003)). These filters provide optimum conjoint representation of a
signal in both space and spatial frequency since a sine wave is perfectly localized in
frequency, but not in space. A quadrature pair of Gabor filters is used to decompose a
signal, with a real part specified by a cosine modulated by a Gaussian, and an imaginary
part specified by a sine modulated by a Gaussian. The real part of the filter is known as
the even symmetric and the imaginary part as the odd symmetric components. The filter’s
centre frequency corresponds to the frequency of the sine/cosine wave, and the
bandwidth of the filter is specified by the width of the Gaussian.
In the Daugman’s iris recognition system, 2D versions of Gabor filters are used in order
to encode the iris pattern data in the normalized polar coordinates
. The filter
wavelet function can be described as follows:
(2.36)
68
where
and
are used to specify the multi-scale 2D wavelet size.
wavelet angular frequency.
represents the
are the center location of the frequency selective filter
bank.
The feature encoding process begins by sampling a collection of feature points from the
original iris image into the Cartesian coordinates. Afterwards, these feature points are
unwrapped into a matrix representation in the normalized polar coordinates using the
Daugman’s rubber sheet model described in section (2.8.5). Once this is done, a set of
Gabor filter banks is applied to the matrix which is then decomposed into a set of
complex coefficients
at location
as follows:
(2.37)
where
and
represent the dimensions along the radial and circumferential directions in
the normalized polar coordinates, respectively. Once the complex coefficients are
calculated using equation (2.37), the complex domain is divided into four phases or
quadrants, and each phase is represented by two binary bits (Figure 2.13).
After the Gabor feature extraction, Daugman demodulates the output of the Gabor filters
in order to compress the data. This is done through phase quantization into four levels,
for each possible quadrant in the complex plane. In general, phase information provides
more discriminating information compared to amplitude information since it discards
redundant information such as illumination variation within the image. Therefore, a
complex feature matrix is generated from the image and for each complex feature value
, two binary bits
are used to represent phase information at the pixel location in
69
the iris template. A compact
-byte iris template is then created, which allows for
efficient pattern comparison and decision making. The Hamming distance is calculated
between two binary feature templates to evaluate their match probability (Daugman
(2003)).
Im
Re
Figure 2.13: Phase quantization (taken from Masek (2003))
This process which includes Gabor feature extraction and phase encoding is repeated on
the entire iris image. One set of Gabor filter banks will extract one pair of complex
phasors for each feature point. For example, by applying k sets of Gabor filter banks on
an unwrapped image template of size
created with size
, a phase matrix of binary bits will be
. This will be the binary iris template that is used for the
Hamming distance calculation.
70
2.8.6.2
Log-Gabor filter
Brady et al. (2000) propose a model to explore the effectiveness for encoding the
information in raw biometric images. The Log-Gabor filter is examined to encode the
spatial, frequency and orientation information in an image. It was noticed that one
weakness of the Gabor filter is that the even symmetric filter produces a DC component
whenever the bandwidth is larger than one octave (Field (1987)). However, the use of a
Gabor filter that is Gaussian on a logarithmic scale will eliminate the DC component.
This is known as the Log-Gabor filter. The frequency response of a Log-Gabor filter is
given by:
(2.38)
where
and
represent the centre frequency and the filter bandwidth, respectively.
To encode iris information when working with an unwrapped iris matrix representation,
each row of pixel intensities corresponds to a ring of pixels centered at the pupil center.
In order to extract the phase feature templates, the Log-Gabor filter is applied to the 1D
image vectors. Since the normalization process involves unwrapping the iris region from
the circular shape to a rectangular matrix (i.e. from the Cartesian coordinates to the
normalized polar coordinates), the spatial relationship along the concentric sampling
rings and the radius become independent. Knowing this, the 2D Gabor filter feature
extraction mechanism will basically mix the relative spatial relationship when it
multiplexes over the normalized polar scale. Therefore, the 2D Gabor filter applies a
71
symmetric Gaussian envelope to the normalized polar image representation that is not
supposed to be treated evenly between radial and circumferential directions. On the other
hand, the 1D Log-Gabor filter extracts the feature vector from each row of the
normalized matrix representation, which avoids mixing the relative position information
between the radial and the circumferential directions.
2.8.6.3
2D Hilbert transform
The 2D Hilbert transform can be used on the Daugman rubber sheet model representation
in order to extract the features from the normalized rectangular iris images (Tisse et al.
(2002)). The analytical signal
becomes:
(2.39)
where
, and
represents the 1-D complex feature vector generated from original signal
stands for the Hilbert transform. In a similar fashion, the 2D version of the
Hilbert transformed image is computed in order to calculate the instantaneous phase and
frequency. From this, the complex phase information is taken as the iris feature template,
in the same way as the Gabor transformed phase template in the Daugman system.
72
2.8.7 Matching Algorithms and Distance Measure
Using one of the previously described feature extraction schemes, an iris image is
processed and transformed into a unique representation within the feature space. In order
to see if two iris templates match (i.e. extracted from the same eye) which involves
making an accept/reject decision, a distance measure is indeed necessary to measure the
closeness of a match. For example, some widely used methods in the iris recognition field
are the Hamming distance (HD), the normalized correlation (NC) and the weighted
Euclidean distance (WED). Other distance measures have also recently been proposed,
which aren’t used in this thesis.
2.8.7.1
Hamming distance
In order to measure the statistical independence between two iris templates, the Daugman
algorithm calculates the correlation between them. For instance, the Hamming distance is
used to measure the difference between encoded binary phase feature vectors. The
Hamming distance between two iris templates is defined as follows:
(2.40)
where
and
and
represent the two encoded iris feature matrices.
are two binary masks where the location of noise pixels is marked by “ ” and
the rest of the mask with binary “ ”. The operator
represents the logical
73
operation where it compares bit by bit, and
stands for the logical
operator which
takes the common area between the two matrices representing the valid iris regions.
Hence, the Hamming distance calculates the pattern difference between iris templates by
using a bit to bit comparison. For iris templates extracted from the same eye, the
statistical independence and the Hamming distance tend to approach zero, while two
different iris templates will have a Hamming distance close to
. For this reason, it is
important to set a proper threshold when computing the Hamming distance between two
iris templates in order to decide if both templates come from the same eye, thus achieving
accurate feature identification.
2.8.7.2
Normalized correlation
The normalized correlation method is implemented in the Wildes system in order to
measure the closeness of match between two encoded iris images. The normalized
correlation is defined as follows:
(2.41)
where
and
represent the two encoded iris templates of size
correspond to the mean of
of
and
and
, respectively.
and
,
and
are the standard deviations
, respectively.
74
2.8.7.3
Weighted Euclidean distance
Similar to the Hamming distance, the weighted Euclidean distance (WED) is another
distance metric that can be used in a biometric system. It defines the closeness of match
between two iris feature templates. For the weighted Euclidean distance, the norm
between two vectors is calculated. As described in (Zhu et al. (2000)), the weighted
Euclidean distance measure computed for a known and an unknown iris template is
calculated as follows:
(2.42)
where
and
respectively.
represent the unknown and known (i.e. in database) iris templates,
denotes the index of the features in the templates, and
deviation of the
is the standard
feature calculated from template .
2.8.8 Evaluation Metrics
2.8.8.1
Genuine-Impostor distributions
From the biometric templates computed in an iris recognition system, the genuine
and impostor
distributions are plotted showing a normalized histogram of the
distance measures (Figure 2.14). The histogram shows the matching measure according
to whether or not the biometric feature templates belong to the same subject. For
75
instance, the genuine histogram reflects all the distances measured within the same class
while the impostor histogram presents all the distances between different classes.
Therefore, the genuine distance measurements should be smaller than the impostor
distances since they represent the closeness of two iris templates extracted from the same
eye. Generally speaking, in an ideal biometric system, the genuine and impostor
distributions should be completely separated which signifies that all images are well
classified since they either come from the same or a different class. Unfortunately, this
doesn’t occur in most practical systems and the underlying genuine and impostor
distributions overlap at the tail (Figure 2.14). For this reason, a threshold must be selected
in order to obtain a good separation between both distributions. The latter are also used to
calculate the amount of overlap between the two curves which represent the false match
and non-match rate.
Figure 2.14: FAR and FRR seen from the overlap of the intra-class and inter-class distributions.
Also, as an example, the separation Hamming distance in this figure is 0.35 (Masek (2003)).
76
2.8.8.2
False Match Rate and False Non-Match Rate
The main goal of an iris recognition system is to be able to identify users by achieving a
good separation between the intra-class and inter-class Hamming distance distributions.
When the Hamming distance between two templates is less than a selected threshold
value, the processed templates are classified as coming from the same iris and a match is
found. On the other hand, if the Hamming distance is greater than the chosen threshold,
the templates are considered to come from different irises. The performance of the system
can be evaluated using the variability among the iris feature templates. Such a measure
also includes the within and the between subject variability where the between variability
sets the limit for minimum false match rate (FMR) and the within variability measure sets
the limit for minimum false non-match rate (FNMR) (Daugman (2004)). The intra-call
and inter-class comparisons between the biometric feature templates can be represented
using the genuine (
) and imposter (
) distributions. From this, the x-axis
represents the direction along which the WED between the iris features are plotted and
the y-axis shows the percentage of WED that fall into that range. Using this, the FMR
and FNMR are calculated as follows (Wayman (1999)):
(2.43)
(2.44)
A series of FMR and FNMR values are calculated using equations (2.43 and 2.44) for
different values of threshold . The match and non-match rate pairs are plotted in one
77
graph showing FNMR against FMR values. This graph is useful in a sense that it clearly
illustrates the tradeoff between the false reject to false accept rates when varying the
threshold
during the identification process. Clearly the separation point or threshold
will affect the false accept and false reject rates. For instance, a lower separation
Hamming distance will decrease FMR while increasing FNMR, while a larger threshold
will cause the opposite effect (Figure 2.14). For this reason, when choosing a separation
point it is important to consider both the false accept rate and false reject rate. FNMRFMR graphs have been used extensively to evaluate the performance of a biometric
system, and it is termed as detection error tradeoff (DET) in other occasions. The DET
curve has been used in this thesis to evaluate the performance of iris recognition
algorithm.
Another measure is the decidability metric which determines the performance of a
system. It helps determining the optimum parameters from which we can analyze the
system’s performance for optimal configuration (Masek (2003)). This metric is described
in the next section.
2.8.8.3
Decidability measure
A good metric to measure the separation between the distributions is the “decidability”,
which takes into consideration the mean and standard deviation of the intra-class and
inter-class distributions, respectively (Daugman (2002)). The decidability ( ) is
calculated as follows:
78
(2.45)
where
represent the mean of the intra-class and inter-class distributions,
respectively. In addition,
correspond to the standard deviation of the intra-class
and inter-class distributions, respectively. The decidability
corresponds to a distance
measured in standard deviation and is a function of the magnitude of the difference
between the means of the intra-class and inter-class distributions. Therefore, from
equation (2.45), it is seen that higher the decidability, the greater the separation of intraclass and inter-class distributions, which allows for more accurate recognition. Thus, a
system that has a higher decidability value tends to have a better intra and inter-class
distribution separation.
2.8.8.4
Rank-1 Identification rate
The identification process in a biometric system requires distance measurement (i.e.
Euclidean or Hamming distance) in order to determine the closeness of match between
two iris biometric templates. From this, matches are sorted according to distance
measures and the smallest distance obtained is considered to be the rank-1 match. Also,
the percentage of all the correct matches among all comparisons is the rank-1 match
score. In this work, the Hamming distance is used as a comparison metric.
79
2.8.8.5
Cumulative Match Curve
Another approach widely used to evaluate the performance of a biometric system is the
cumulative match curve (CMC). It is a plot of the cumulative match score against the
rank, which represents the percentage of images identified below the rank (Rukhin et al.
(2005)). This is generated by comparing each feature template against all the other
feature templates in the database. From this, a complete set of distance metrics is created
among which the smallest one is taken as the closest match in that specific class. Also,
among all the comparisons, the rank-1 match score would reflect the percentage of the
correct matches using the smallest distance to determine a correct match.
In a similar way, the rank-2 match score would be the percentage of correct matches if
using the second smallest distance as the correct match. Therefore, a series of match
scores could be calculated against the ranks such that the rank-n match score would
represent the percentage of correct matches when using the
smallest distance as a
match.
2.8.9 Image Database and Open Source Software
In this thesis, iris images from the CASIA database (Institute of Automation of the
Chinese Academy of Sciences, (2003)) are used to test and evaluate our system’s
performance. The CASIA database includes 689 iris images that are taken from 108
subjects (i.e. different eyes) using near-infrared cameras. The iris images are grayscale
bit-map with a resolution of
. In this work, a specific class of iris images
corresponds to one subject and each class consists of
or
iris images acquired from the
80
same eye. From the CASIA database, 327 low quality iris images, containing partial
eyelash occlusions and noise, are selected for further testing. Therefore, eyelash
detection, iris segmentation, circular localization, feature extraction, feature matching and
distance measurements are conducted on these images. Afterwards, biometric
information is calculated on the segmented iris region. Some of the code used in this
thesis such as the normalization, the Gabor filter and the Hamming distance was taken
from the open source Matlab code framework implemented by Masek (Masek et al.
(2003)) that is based on the Daugman’s iris recognition scheme. The rest of the newly
developed scheme creates a new methodology in order to improve the iris recognition
system’s performance in low quality images as well as to improve the match score using
enhanced segmentation.
2.9 Non-Cooperative Iris Recognition
One of the main objectives of this work is to develop new techniques that can be applied
in the biometrics field for non-cooperative face and iris recognition in an uncontrolled
environment. Non-cooperative iris recognition is the process of automatically recognizing
individuals using their eye/iris images captured at a distance and without requiring any
active participation. The problems that can occur in a non-cooperative environment are
related to image acquisition challenges, image quality, noise, lighting, pose and many
others. Figure 2.15 illustrates a non-cooperative face detection and iris recognition
biometric system where initially, subject’s face is captured in an uncontrolled
environment, then pre-processed for further tracking. Afterwards, we propose an
81
algorithm to capture an eye image, eliminate noise in it and then, use the resulting image
for further iris localization and automatic identification.
Some work exists related to non-cooperative iris recognition that attempts to solve this
challenging issue. For example, Dorairaj et al. (2005) developed an iris recognition
system that deals only with off-angle images. Their method consists in estimating the
gaze direction, through the Hamming distance between the Independent Component
Analysis of a frontal view image and the one that is actually captured. In addition, they
applied a projective transformation in order to bring the captured iris image to frontal
view. This then becomes similar to a standard frontal view iris recognition process. The
authors used images of the CASIA database, as well as few other images captured in their
institute.
Sung et al. (2002) addressed the challenges on non-cooperative iris recognition by
roughly identifying some potential problems that should be resolved for accurate
detection. They considered the problem of lighting conditions as being insoluble, unless
special lighting methods are introduced. The problem of off-angle images, when the gaze
of the subjects is not directed to the camera, motivated the development of a slightly
uncommon segmentation method composed by the initial inner eye corner detection
followed by a least square elliptical fit to the limbic edge pixels. The authors propose a
method based on wavelet packet maximum Shannon entropy reconstruction for
measuring the image information in order to identify the information degradation
resultant from the non-cooperative image capturing, especially on the acquisition of
82
defocused images. Afterwards, the authors applied a bank of complex-valued 2D Gabor
filters for the feature extraction. From this, the authors then concluded that the feature
comparison by means of correlation and classification through the nearest neighbor
outperforms the Hamming distance matching metric, although they used a small database
to test their method.
Fancourt et al. (2005) claim that it is possible to achieve iris recognition at up to 10
meters distance by using an imaging framework composed from a telescope and an
infrared camera. Images were captured at various distances, capture angles,
environmental lighting and eyewear. The authors concluded that minor performance
degradation was seen with an increasing distance, off-angle images and eyewear. They
used a local correlation matching metric for the pattern comparison process. However,
similarly to Du et al. (2005), their results were obtained using high quality images that do
not contain a significant amount of noise.
All the above proposed methods claim the possibility of capturing iris images with
enough quality in less cooperative biometric systems while achieving accurate human
recognition. However, these methods do not deal with some major issues such as
underexposed images, face occlusion, highly occluded iris regions, rapid head movement,
and many more.
83
Face Detection
and
Initialization
Pre-Processing
and
Image enhancement
Face Initialization
Face Tracking
Subject
Eyes Detection
Eyes Localization
Iris Localization
Enhanced Iris
Segmentation
Iris Capturing
Figure 2.15: Possible non-cooperative face detection and iris recognition system.
2.10
Summary
In this thesis, new methods and solutions are developed in order to improve the iris
recognition system’s performance when using low quality images. As mentioned in Jain
et al. (2000), an iris recognition system is a pattern recognition system which performs
personal identification by using unique and rich texture information extracted from an iris
image in order to establish the authenticity of a specific physiological characteristic
possessed by the user. The iris recognition system applied in this thesis makes use of
some of the existing techniques such as the log-Gabor filter for feature extraction, phase
84
quantization for template encoding and finally, the Hamming distance as a matching
metric.
Since one of the main issues with low quality iris images is eyelash and eyelid
occlusions, methods must be developed in order to eliminate this type of noise prior to
matching. A typical example of an occluded image is shown previously in Figure 2.6.
The upper eyelid and eyelashes have covered a significant portion of the iris, which
falsifies some the of the iris texture information. If this problem is not resolved before
proceeding with the feature extraction and matching steps, these erroneous pixel values
will cause the matching between two genuine biometric templates to be inaccurate which
might result in a false match.
For this reason, a new algorithm is developed in this thesis to detect most eyelash pixels
in a low quality iris image. When noise is properly removed, additional valid iris pixels
are then used and higher accuracy can then be achieved at the matching level. Further in
this thesis, system’s improvement due to proper eyelash detection and elimination is
shown through the ROC curve which demonstrates improved error rates in most of the
FMR range (Figure 4.20).
85
3 Chapter 3
Using infrared illumination to improve eye & face
tracking in low quality video images
3.1 Introduction
Recently, biometrics has been a major field of research that is indispensable for
authentication and identification of subjects and for increasing security. The use of
biometrics as a way to authenticate user’s identities has been a topic of research for years.
Eye tracking and face detection and recognition are a major branch of biometrics that is
employed in many areas, such as airport security and border management. The need of
fully automated systems that analyze the information contained in face images is
necessary and for this reason, robust and efficient face and eye detection algorithms are
required. Given a single image or a sequence of images, the goal of face detection is to
identify all image regions which contain a face regardless of its three-dimensional
position, orientation and lighting conditions. Such a problem is challenging because faces
and eyes are non-rigid and have a high degree of variability in size, shape, color, and
texture. The ability to detect faces and eyes in a scene is critical to modern surveillance
applications. While many image processing algorithms exist to detect faces in images,
86
their performance is not completely reliable, especially in situations with variable
lighting, and when dealing with low resolution images (Fromherz et al. (1997)).
Many authors used different techniques to detect eyes in images obtained under active IR
illumination. For example, appearance-based methods are developed in (Huang et al.
(2000), Pentland et al. (1994)). These methods use the bright pupil effect and its motion
characteristics to track the eye regions. However, these methods fail when eyes are closed
or occluded and when subjects show rapid head movement. Another method based on the
Hough transform is presented in (Nixon (1985)) to detect the eye region. This technique
is time-consuming and requires high quality eye images with a good contrast to succeed
in the detection process. Other papers developed methods using Kalman filtering and the
mean shift tracker (Comaniciu et al. (2000), Zhu et al. (2005)) to detect and track eyes in
an image. However, these methods might fail when applied on low quality images for eye
tracking.
In this work, we develop a new robust system for eye and face tracking in low quality
images using the bright pupil effect (Haro et al. (2000), Morimoto et al. (1998)). By
using IR illumination, it is possible to get information from which the eye positions in the
image can be calculated. Our system consists of three major parts: face localization, eye
detection and eye tracking. This is accomplished using traditional image-based passive
techniques such as shape information of the eye and active based methods which exploit
the spectral properties of the pupil under IR illumination. A frame differential templatebased technique (Ebisawa et al. (1993)) and a feature-based principal component analysis
87
method are used to search the image for valid eye regions. Afterwards, Kalman filtering
is applied to locate the bright pupil candidate of interest. If the processed image contains
weak reflections due to occlusion or eye closure, the algorithm uses the contour and
shape information of the eye to approximate the pupil location. This is achieved by using
adaptive thresholding techniques to extract the eye contour (Niblack (1986)).
The
developed method shows promising results in tracking subject eyes and face in low
quality images under variable illumination and different head orientations. The algorithm
is also tested on underexposed images where the subject shows large head movements.
3.2 Algorithm design
This section develops a system designed in the goal to automate the detection of faces
and eyes in images taken under low illumination with ON and OFF IR. The proposed
algorithm is designed to detect and track the subject’s eyes under challenging conditions.
Using the computed eyes location, face detection in the subsequent frames can be easily
accomplished. The developed algorithm consists of several steps in order to pre-process
the low quality and underexposed face images. For instance, our method uses some
existing image processing techniques with the objective of increasing the system’s
robustness and performance when dealing with challenging situations such as low and
non-uniform illumination, rapid head movement, occlusions, out-of-plane and in-plane
head rotations and many more.
88
3.3 Face detection
This section elaborates on eye and face detection using image enhancement, frame
differencing and adaptive template correlation techniques. The initialization step is
accomplished by using geometric and shape constraints of the eye to find the pupil
regions in the enhanced difference image. Some of the applied constraints are: the eye
separation distance, angle measurement, pupil size and eyes location within the face
boundary.
3.3.1 Experimental setup
Underexposed low quality images are taken using a single black and white camera,
sensitive to infrared light, with zoom lens of
and a NTSC output to the
frame grabber. In order to adjust the overall illumination of the area where images are
taken, a standard 60W bulb with variable illuminations was installed. IR illumination is
produced with IR diodes strobed by a frequency generator, allowing the acquisition of
experimental data with ON and OFF IR. The complete setup can be found in (Asfaw et
al. (2002)). The data are captured for different users under various experimental
conditions to simulate real life scenarios. Each volunteer is placed
away from the
camera. 24 test samples (5 seconds each) per volunteer are acquired with a combination
of ON and OFF IR. The pupil is then detected using the subtraction method (Zhu et al.
(2005)) using the dark and bright pupil images.
89
3.3.2 Non-linear image enhancement and denoising
This step is crucial for low quality images taken in low illumination (Figure 3.2). The
original image is low pass filtered with a 5× 5 Gaussian filter with N = 4 iterations, using
the non-linear edge and contrast enhancement algorithm described in (Deng et al. (1993,
1994)). We consider the image gray level digital representation in the [0, M) range, where
M = 255 for an 8-bit image. In order to avoid “clipping” (i.e. losing information),
arithmetic operations on image pixel values are defined in a logarithmical mapped space
where the forward mapping function between the image pixel space (F) and the real
number space ( ) is:
. Since vector addition, subtraction and
multiplication are bounded operations and well defined in the log space, it is possible to
derive non-linear equations that overcome the “clipping” problem caused by linear
methods (Bovik (2000)). The iterative technique shown in Figure 3.1 overcomes the
limitations of linear methods by performing a non-linear weighting operation on the input
pixels of the image. This requires the selection of parameters
high frequencies introduced in the solution. If
to control the amount of
, the solution will be smoothed;
otherwise it amplifies edges. The output of this system results in an enhanced image with
reduced high-frequency content and better contrast.
90
Figure 3.1: Multistage Algorithm block diagram showing three stages. At each stage, the input image
(F or A) is filtered using a Gaussian [5 × 5] low-pass filter. An image containing only high-frequencies
H(x,y) is obtained by subtracting the smoothed output from the input. The edge amplification
parameter si is selected at each stage, i, based on the level of high frequency noise in Hi(x,y). c is a
scalar controlling the contrast level in the enhanced image.
3.3.3 Histogram stretch
The histogram in Figure 3.2b is stretched in order to fill the entire available gray-scale
range. Lower and upper histogram threshold values tL and tH are calculated corresponding
to
% and
% of the total number of pixels in the histogram. This results in a more
visually distinctive image (Figure 3.2c) with a broad histogram (Figure 3.2d). Also, this
operation provides better edge delineation, which facilitates the extraction of the face
contour from the background in the initialization step.
a
b
c
d
Figure 3.2: (a) Low quality underexposed IR image showing shows that most pixels have low
intensity values due to poor illumination, (b) its corresponding histogram, (c) Noise reduction and
contrast enhancement using the log-ratio approach. The histogram stretch operation (d) presents
better edge delineation around the face region. Image a) shows that the pupil classification problem
is very challenging since subject’s eyes show very weak reflections.
91
3.3.4 Non-linear coarse edge enhancement
The image obtained in section (3.3.3) is low pass filtered with a
with
Gaussian filter
iterations, using the non-linear image enhancement algorithm described in
section (3.3.2). The output of this system results in a coarsely enhanced image with a
well-defined face boundary (Figure 3.3b).
(a)
(b)
Figure 3.3: (a) Coarsely enhanced image using non-linear enhancement algorithm, (b) binary image
using tH as threshold
3.3.5 Morphological image erosion operation and Edge detection
Morphological erosion is applied on the binary image in Figure 3.3b to reduce region
expansion caused by the blur effect from the non-linear edge enhancement operation in
section (3.3.4). A
disk-shaped structuring element is used for the morphological
image erosion operation. Using the eroded binary image, a Sobel operator is constructed
to perform a 2D spatial gradient measurement on an image and gives more emphasis to
high-frequency regions that correspond to edges (Figure 3.4a). The Sobel operator
consists of a pair of
convolution kernels, which are designed to find horizontal and
vertical edges in an image (Bovik (2005)).
92
3.3.6 Initial face contour extraction
Initial face contour extraction is performed as follows:
•
Compute all points on the contour in the image of Figure 3.4a obtained from
section (3.3.5).
•
Find an arbitrary point located in the face region by scanning the image row-wise
and by taking the mean of all computed edge points on the contour.
•
Starting at the approximated face location found in the previous step, search for
all points located on the inner face boundary that are not black (iteratively).
•
Create an intensity vector by summing all intensity values in Figure 3.4a columnwise. The intensity values corresponding to both maxima on the graph in Figure
3.4b and that are located on the inner contour are chosen as face proximities.
Intensity summation
200
180
50
160
140
# of pixels
100
150
120
100
80
60
40
200
20
0
(a)
50
100
150
200
250
(b)
0
50
100
150
200
Horizontal pixel position
250
300
Figure 3.4: (a) Inner face contour extracted using a Sobel operator and connected component
analysis, (b) plot of pixel intensity summation in the vertical direction.
93
3.4 Initial eye detection
The first step is to locate the eye position in the initial bright (i.e. image taken under IR
illumination where pupil show reflections) and dark eye images sequence.
This is
achieved by first subtracting the dark eye image from the bright eye image. Also, an
image obtained using the morphological opening operator (a disk structuring element of
size 2) is also subtracted from the bright pupil image. The resulting difference images are
then thresholded using a weak threshold in order to account for all the reflections in the
image. In order to reduce the background noise caused by non-uniform illumination, we
apply a logical AND operator to both images in order to keep the reflections common and
present in the binary images. The process is presented in Figure 3.5 as a block diagram
and the resulting images are shown in Figure 3.6. Afterwards, in order to locate the right
pupil region, the binary image (Figure 3.6h) is then processed using shape and geometric
information of the eye. Some of the measures that are used to segment the true pupil
regions are: the within-eye distance, angle measurement, pupil size and eyes location
within the face boundary. For example, any blob with area greater than
pixels is
neglected since it cannot correspond to eye reflections. The remaining blobs are
processed using distance and angle constraints as follows:
•
Compute a matrix (Λ) of size
as all possible distances in the vertical
direction between all eye regions such as
if
•
, and otherwise.
Repeat the process by computing a matrix ( ) of Euclidean distances between all
possible candidate regions as
94
where x and y are image coordinates in the horizontal and vertical direction,
respectively.
•
Search through
and
for two candidate regions that have the smallest distance
(greater than zero) in the vertical direction and which have an Euclidean distance
“
”
located
within
(
%
).
Bright pupil
image
and
%
of
the
face
width
is computed from section (3.3.6).
Image
Subtraction
Dark pupil
image
Image
Thresholding
Morphological
Opening
operator
Logical AND
operator
Image
Subtraction
Eyes located
Eye classification
Shape and geometric
constraints filters
Figure 3.5: Initial Eye detection block diagram
95
(a)
(b)
(c)
(d)
(e)
(f)
96
(g)
(h)
Figure 3.6: Image differencing process: (a) The original bright pupil image, (b) dark pupil image, (c)
image obtained after morphological opening with a disk structure of size 2, (d) Subtraction result of
images: [a-b], (e) Subtraction result of: [a-c], (f) Thresholded image (d) using a very small threshold
to account for most reflections in the image, (g) Thresholded image (e) using the same small
threshold, (h) Image obtained using the logical AND operator which keeps the bright regions which
appear in both thresholded images.
3.5 Eye and Face tracking
After computing the initial eye position in section (3.4), an eye and face detection
algorithm is initiated for eye tracking. A Kalman filter is activated in order to track bright
pupils in subsequent frames. In case the images contain weak reflections or if the subject
exhibits rapid head movement, the Kalman filter might fail in tracking the exact location
of the pupil which is then approximated using the extracted eye contour.
3.5.1 Template correlation
Using the initial eyes location computed in section (3.4), two eye templates are created
for further matching. The template size is chosen to be
in our experiment, which
has shown to provide accurate tracking and correlation scores. Templates are updated
after each frame using the previously computed pupil positions. The correlation scores
97
are then calculated for every frame using the normalized correlation coefficient (Chau et
al. (2005)):
(3.1)
where
is the intensity value of the video frame at point
value in the current search window,
location
,
is the average
is the intensity value in the template image at
, is the mean value of the template image. In order to reduce computation
time, this operation is not performed on the entire image, but instead only on the
extracted face region found in the frame at time
. This computation gives correlation
scores between - and where indicates a probable match between the search region of
interest and the correlation template. This similarity measure has the advantage that it is
insensitive to changes in lighting conditions which can then be used for images taken in
low illumination.
3.6 Eigen-eyes
In case the template matching method does not provide high correlation scores, a
principal component analysis technique (Turk et al. (1991)) is used to search for possible
eye regions. The previously computed eye templates
with centers located at
. Using
of size
are normalized
, the principal component analysis is
performed as follows:
98
•
Represent
•
Compute the mean
in a vector format
of the previous templates
(3.2)
•
Subtract the mean
from the data (eye templates)
(3.3)
•
Compute the covariance matrix
as
(3.4)
where
•
Compute the eigenvectors and eigenvalues of the covariance matrix
(3.5)
•
Determine the most dominant K eigenvalues with their eigenvectors.
•
The template eye image can be represented as a linear combination of the K
eigen-basis vectors such that:
(3.6)
where
and
represent the eigen-eyes
The normalized eye template can be represented in the developed basis by the following
vector:
(3.7)
99
Eye region detection:
•
Each candidate eye image
in the extracted face region is projected into the
corresponding eigen-space
If
•
Compute
•
Compute
•
Compute
, where
region is possibly an eye.
measures the distance from the eye space, then the selected
is selected based on the image database that is used
and it is chosen through experimentation.
3.7 Face detection using the previous face template
In case eye regions are not detected possibly due to occlusions or head rotation, the
algorithm must keep track of the face region. This is achieved by using previous face
template extracted using preceding eye location. The intensity distribution of face region
around the eyes is calculated in order to locate the exact contour of the face in the
succeeding frames. Using the previously calculated face template at time
and the
current dark pupil image, both images are downsampled prior to correlation in order to
reduce the processing time. Afterwards, correlation scores are computed using equation
(3.1). The sub-region with the highest correlation score (close to 1) is assumed to contain
face information. This process is illustrated in Figure 3.7.
100
1
Prior face template
obtained using frame
(t-1)
Dark pupil
image frame (t)
Upsampling
↑2
-1
Downsampling
↓2
Normalized
correlation
coefficient
Face region
extraction
Figure 3.7: Block diagram showing face detection using prior face template. This algorithm is
activated when the processed image does not contain valid eye regions possibly due to out-of-plane
head rotation or occlusions.
3.8 Pupil candidate regions computation and eye tracking
Once the eye and face regions are approximately located, the current bright and dark
pupil images are preprocessed as described in section (3.4), and all possible eye candidate
regions computed. The elimination of non-eye blobs is achieved by first imposing the
same geometric and shape constraints in the selection process and then activating a multistage eye tracking module based on Kalman filtering and eye contour segmentation. The
classification and tracking process is described as follows:
•
Compute the centroid of all the remaining blobs in the binary image obtained
using connected component analysis and mark them all as possible pupil
candidates (Bovik (2005)).
•
Eliminate all regions with centroids located outside the face contour computed in
section (3.3.6).
101
•
Eliminate all regions with centroids located outside the eye regions computed
through template matching. If all blobs in the binary image are eliminated, go
back to the previous step and skip this one. This might happen if the subject
shows rapid head movement in which case all candidate regions will be located
far from the prior eye location, outside the estimated boundary.
•
Find the region that minimizes the following equation which calculates the closest
centroid to the prior eye location computed in images at instance
.
(3.8)
where
represent the prior pupil location and (
) are the centroid
position of blob i. (This operation is done for the left and the right eye separately).
•
After detecting possible pupil candidates, the Kalman filter (section (3.9)) is then
used to compute the exact location of the bright pupil in the difference image. In
case none of the binary regions are classified as possible pupil candidates, go to
the next step.
•
If none of the blobs are classified as possible pupil candidates (probably because
of weak reflection, eye closure or occlusion), an eye contour extractor (section
(3.10)) is then activated to give a more accurate approximation of the pupil
location.
102
•
After computing the pupil coordinates, update the eye and face templates to
process the successive frames.
The entire process is shown in Figure 3.8 as a block diagram.
103
Frames at time t
&
Computed Eyes
position at t-1
Template
correlation
Eigen-space
decomposition
Eye region
extraction
Success
No
No
Eye region
located
Face detection
using previous
face template
Yes
Yes
Face & Eye region
approximation
Bright/dark
images
Image preprocessing
&
Connected component analysis
Blobs
Geometric & Shape
constraints filters
Approximated
Eye & Face
location
Prior eye
location
Candidate
pupil regions
Eyes
located
Distance
minimization
Yes
Success
Kalman filter
No
Update eye
target model
Eye contour
&
Centre extraction
Figure 3.8: Block diagram showing eye and face tracking algorithm. The algorithm is initialized with
the previously computed eyes and face location at time t-1. Subsequent frames are then processed
using a Kalman-based and adaptive thresholding techniques to successfully track eyes in low quality
images.
104
3.9 Kalman filtering
Kalman filtering is applied on the thresholded difference image obtained using the dark
and bright pupil images. The Kalman tracker has proven to work well when the pupil
shows bright reflections and when the subject’s head is slowly moving (Zhu et al.
(2005)). We start by characterizing the motion of the pupil in the eye image at each time
instance by its position and velocity. The pixel position or centroid in the binary image is
represented by (
) at time . (
) represent the pixel velocity in the
and
directions, respectively. The system model (predictor) can then be written as follows:
(3.9)
where the state vector at time t is
is the system perturbation.
,
is the state transition matrix and
is normally distributed as
where
represents the process noise covariance.
Using our feature extractor estimates
of the pupil position in each frame, the
measurement model (updater) can be written as follows:
(3.10)
where
is a matrix relating the current state information to the current measurement and
represents the measurement uncertainty and is normally distributed as
105
where
is the measurement noise covariance. The system and measurement models are
used to provide an estimate of the state and its covariance matrix at time
experiment
1 0 0 0
=
 since
0 1 0 0 
. In this
deals with the pupil position information only.
Assuming a bright pupil effect, the estimate of the pupil position obtained using the
geometric and shape constraints filters is combined with the Kalman estimate to provide
us with the final pupil position.
3.10
Eye contour extractor
In a situation where the subject’s eyes show weak reflections or are occluded, the Kalman
filter will not predict the correct pupil location in the difference image. For this reason, an
alternative way of approximating the pupil location should be developed to complement
the first approach. The selected method processes the possible eye regions computed in
section (3.4) and automatically detects the eyelids using an adaptive thresholding
technique described in (Niblack (1996)). Global thresholding methods fail in extracting
the eye contour since the eye template does not necessarily have a bi-modal distribution
where an optimal threshold can be easily selected to separate the desired region and the
foreground. For this reason, we use a local adaptive thresholding technique that selects a
unique threshold based on the local neighborhood of each sub-region in the image where
illumination is assumed to be uniform (Niblack (1996)). Local thresholds are chosen
using the following equation:
(3.11)
106
where
and
are the local mean and variance of the corresponding local
region. In this experiment, a
window size is chosen with
based on our
experiment. This operation allows thresholding images that do not have an intensity
histogram with two major peaks and that might contain strong and non-uniform
illumination.
This method shows good contour extraction in complex eye images where the eyes are
closing or in the presence of rotation and head movement. Once the eyelids are extracted,
the pupil position is then assumed to be located at the center of this region. A distance
minimization method is then applied to select the candidate blob with its centroid located
the closest to the estimated pupil location using the eye contour and that is nearby the
prior pupil location computed in frame at time
. Once these conditions are satisfied,
the binary blob is then marked as the possible pupil location in order to update the eye
template model in the subsequent frame. The entire process is described in Figure 3.8.
3.11
Experimental Results
The algorithm is tested on two different eye image databases where images are taken for
moving subjects (males and females of different ethnicity) under low and variable
illumination where subjects show rapid head movement, eye closure, various facial
expressions and in-plane/out-of-plane head rotation. The first set is composed of the low
quality images of section (3.3.1) and the other is the RPI ISL IR Eye Database (Wang et
al. (2005)). The image sequences are taken with IR cameras. Each image sequence is
107
decomposed into dark pupil image sequence and its corresponding bright pupil image
sequence. We processed
image sequences taken from the ISL database and
images from the low quality image set. The selected frames display various facial
expressions, and show small and large head movements and long eye closure.
For the initialization part, the
parameters were set to
and to
in section (3.3.2)
in section (3.3.4). These choices were motivated
by the requirement in section (3.3.2), to reduce noise and to slightly brighten the image
by shifting the histogram of the image towards higher pixel intensity values. When the
images show bright pupil reflection, the Kalman filter succeeds in tracking the pupil
location in the difference image. However, when the subject shows significant head
movement and out-of-plane head rotations, the bright pupil effect disappears causing the
Kalman filter to fail. This will then activate the eyelid extractor in order to get a better
approximation of the pupil location. A
window size is chosen for the local
adaptive thresholding technique of section (3.10) with
.
Figure 3.9 and Figure 3.10 show eye detection results in weak and strong reflection
scenarios, respectively. By treating the human-identified parameters as ground truth, we
estimated the pixel and standard deviation results for eye detection in the horizontal and
vertical directions. Table (3.1) shows the detection and tracking results for the two sets of
images. From this, we can conclude that the proposed approach detects and tracks the
eyes accurately. The detection rate is
% using the ISL database and
% using the
low quality images. It is worth mentioning that 6 images in the low quality image data set
108
contained only one eye due to a
degree out-of-plane head rotation. For this reason, the
algorithm does not detect the eye region which is not appearing on the image. However,
the proposed method is able to re-locate the eye target region once it re-appears in the
subsequent frames. The average and standard deviation pixel offset error in the horizontal
and vertical direction are shown in table (3.2). This illustrates the efficiency and accuracy
of our method since in general, an average pixel offset error <5 pixels means that the
detected pixel position is still located within the pupil region that has a diameter of
approximately 5 pixels in both the ISL and the low quality images databases.
Figure 3.9: Eye detection results obtained using low quality images taken in very low illumination
conditions.
109
(a)
(b)
(c)
(d)
Figure 3.10: Eye detection results obtained using images from the ISL IR EYE database. Frames
(b,d) show eye detection under challenging conditions where the subject has his eyes closed.
Database
Subjects
# of images per
subject
Total processed
images
Detection and
tracking results
ISL EYE IR
4
300
1200
1200/1200
Underexposed
low quality
images
3
200
600
591/600
Table 3.1: Table showing eye detection and tracking results. 1800 images are processed in total using
two different databases. A pupil detection rate of 100% and 98.5% were achieved using 1200 images
of the ISL database and 600 underexposed images, respectively. A pupil is located successfully in an
image when the estimated pupil position is contained within the pupil area that has a 5 pixels
diameter.
110
Database
ISL:
1200 frames
Low quality:
600 frames
µerror in pixels
(x-direction)
σerror in pixels
(x-direction)
µerror in pixels
(y-direction)
σerror in pixels
(y-direction)
1.0976
0.9090
2.2352
1.9794
1.4402
1.3297
3.0829
1.9628
Table 3.2: Table showing average pixel offset error and standard deviation results for eye detection
in the horizontal and vertical directions. µerror is average pixel offset error and σerror is the standard
deviation of the offset error, respectively. x and y are the horizontal and vertical directions,
respectively. It is worthwhile mentioning that the pupil diameter is approximately 5 pixels in both
image databases.
3.12
Discussion
This work presents a new face and eye detection system that combines several image
processing techniques in the goal to extract and track face and eye positions from
surveillance type images with IR strobe taken under poor illumination. For example, in
the case where many reflections (blobs) occur, the algorithm will find all possible eye
locations and presents the best solution using multi-stage classification techniques. In
order to reduce the set of possible eye region candidates, shape and geometric constraints
are imposed in the classification process. A Kalman tracker is used to approximate eye
location in bright pupil images. If the image contains weak reflections, a local adaptive
thresholding technique is used to extract the eye contour in order to estimate the pupil
location. This improves the performance and accuracy of the system when dealing with
faces at different orientation and with eye closure. The algorithm achieves a pupil
detection rate of 100% and 98.5% using 1200 images of the ISL database and 600
underexposed images, respectively. A pupil is located successfully in an image when the
estimated pupil position is contained within the pupil area that has a 5 pixels diameter.
111
As a comparison, Zhu et al. (2005) developed an eye tracking technique that locates
pupils using IR illumination. Their technique achieved a 96.7% detection rate when
tested on 1600 frames taken from the ISL EYE IR database. However, their method does
not address the challenge of tracking eyes in underexposed images.
Our developed algorithm consists of several steps in order to pre-process the low quality
and underexposed face images. For instance, this multi-stage algorithm uses some
existing image processing techniques with the objective of increasing the system’s
robustness and performance when dealing with challenging situations such as low and
non-uniform illumination, rapid head movement, occlusions, out-of-plane and in-plane
head rotations and many more. This algorithm is able to adapt to different situations, can
track eyes and faces in challenging situations using advanced image enhancement
techniques when dealing with low quality images.
3.13
Summary
This work proposes an algorithm to automatically detect and track faces and eye
locations in IR images taken under poor illumination. The algorithm detects face region
in the image, extracts the face contour and tracks eye location in subsequent frames using
a Kalman filter. In case the images show weak reflections, a local adaptive thresholding
technique is used to approximate the eye location. Tested on
proposed system achieves a
IR images, the
% detection rate.
112
4 Chapter 4
Improved Identification of Iris and Eyelash Features
4.1 Introduction
Proper Iris segmentation is essential for various security applications using iris
recognition technology for personal identification (Wang et al. (2002)). Irises are
occluded by the eyelid and eyelashes as well as from specular reflections from the
(typically infra-red) illumination system. In order to accurately process the image, it is
important to identify such occluded regions in order to remove them from further
processing. There exist some algorithms that try to bypass the eyelash occlusion by
selecting an iris boundary within the real iris region, such as the one presented by
Poursaberil et al. (2006). This method is not very efficient since it avoids eyelashes by
eliminating both the eyelashes and certain iris regions, therefore, it underestimates the iris
region. Furthermore, it changes the relative sampling distribution of the feature points.
For this reason, inaccurate detection of these occlusions considerably reduces the
performance of an iris-based identification system when subject cooperation is not
possible. Cooperative users can be asked to stand still for multiple image acquisitions,
while for Iris On the Move (http://www.sarnoff.com/) or covert surveillance applications
(i.e. in airport security) such cooperation is not available. This will greatly affect the
localization of the iris inner and outer boundaries as well as it will degrade the iris feature
extraction process. For this reason, exact eyelash detection and segmentation is required
113
to improve the entire biometrics system’s accuracy and improve the recognition
performance. In this work, we develop an algorithm for accurate iris segmentation in
images where the major portion of the iris is occluded. Our algorithm detects separable
and multiple eyelashes, respectively. Separable eyelashes are first detected using a local
intensity variation based algorithm while multiple eyelashes are found using the block
mean and variance approach. Various methods have been proposed for eyelash detection
(Kong et al. (2001), Huang et al. (2004), Yuan et al. (2004), Kovesi (1999)) which uses
1-D Gabor filter, intensity variance, phase congruency, template mean and standard
deviation for multiple eyelash detection and a local intensity minimum method for
separable eyelash detection. All these methods perform generally quite well but do suffer
from some limitations such as computational complexity, inexact iris boundary
segmentation, false eyelash detection and improper eyelash segmentation over the iris
region. Specifically, all previous approaches tend to overestimate the occluded regions,
and thus lose iris information that could be used for identification. In this chapter, a new
algorithm is developed that improves the iris segmentation process by providing a
detailed iris image area while eliminating distortion as much as possible. This enhanced
iris segmentation approach results in iris images with fewer artifacts and distortion
leaving more iris pixels for the recognition process. The masked images are normalized
and encoded by the Log-Gabor filter described in section (2.8.6.2). This is followed by
encoding the extracted iris features using phase quantization and then classifying them
using the Hamming distance measure previously implemented in the Daugman’s system
(Daugman (2004), Xie (2007)). The proposed method addresses most of the noise and
distortion issues within the iris images using a collection of image processing techniques
114
such as: logarithmic (i.e. non-linear) image enhancement, edge detection, morphological
operators, Hough transform, intensity gradient based algorithm and a block mean and
variance method using region’s local statistics. Finally, our results show that the
enhanced segmentation decreases the genuine distances and the error rates, which
effectively increases the robustness of the recognition system. Our main algorithm is
presented in section (4.2) and some experimental results are presented in section (4.3).
Finally, section (4.4) concludes this work.
4.2 Enhanced Segmentation
The following section develops an algorithm to automate the detection and segmentation
of eyelash features in an eye image. The design criterion is to determine the detailed
eyelash regions without overestimation (falsely detecting iris regions in the image as
eyelash). In performing this calculation, it localizes and segments the pupil-iris region
using the Hough transform technique (Gonzalez (2002), Ballard (1981), Xie (2007)) and
the non-linear image enhancement algorithm applied on the iris region in order to
facilitate eyelash detection.
4.2.1 Pupil-Iris region localization and boundary extraction
Assuming that the pupil and iris regions have circular shapes, the Hough transform,
combined with standard edge detection techniques, is used to detect the circular
boundaries in iris images. This section describes an algorithm for accurate iris boundary
115
detection and contour extraction (Figure 4.1) based on a combination of image processing
techniques.
4.2.1.1
Non-linear image enhancement
This recursive filtering technique overcomes the limitations of linear methods by
performing a non-linear weighting operation on the input pixels of the image. This
requires the selection of parameters
introduced in the solution. If
to control the amount of high frequency content
< 1, the solution will be smoothed otherwise, it amplifies
edges. The output of this system results in a binary enhanced image with sharper edges
and better contrast (Figure 4.14c). Since this method attempts to remove distortions in the
images, it makes the subsequent localization of the circular boundaries in iris images
easier.The original eye image
goes through a recursive smoothing operation
(Figure 3.1). At each iteration,
is low pass filtered with a
Gaussian filter
using the non-linear edge and contrast enhancement algorithm described in (Deng et al.
(1993), Deng et al. (1994)). In each step, the output image, obtained from the previous
iteration, becomes the input to the next Gaussian low-pass filter (LPF).
represents the smoothed image at step
,
represents the subtraction of the
images before and after the smoothing operation and
is a scalar that controls the
contrast level in the final enhanced image
. The whole process is repeated until
the following condition
is met.
116
In order to avoid “clipping”, arithmetic operations on image pixel values are defined in a
logarithmical mapped space where the forward mapping function between the image
pixel space ( ) and the real number space ( ) is
(4.1)
Using equation (4.1), the addition, subtraction and multiplication operations in the
logarithmic space are defined as follows:
(4.2)
(4.3)
(4.4)
where
and
represent the two grayscale images and a scalar, respectively. The
Gaussian filtering is computed in the grayscale intensity space while the addition
subtraction
and multiplication
,
are defined in the logarithmic mapping space. The
images are first mapped into the logarithmic space before the addition, subtraction and
multiplication operations and then inversely mapped back into the grayscale intensity
space after these operations (equations 4.2 to 4.4). These images are used for the
segmentation, the feature extraction and the pattern comparison. In this case,
is
used to search for the circular boundaries of iris and pupil regions. Once the parameters
of these boundaries are located, the original iris image is used for the feature extraction
and subsequent processing.
117
4.2.1.2
Edge detection
The edge and contrast enhanced image
obtained from the previous section is
used to identify the spatial gradient map with the Sobel edge detector. The Sobel operator
performs a 2D spatial gradient measurement on an image and gives more emphasis to
high-frequency regions that correspond to edges. The Sobel operator consists of a pair of
convolution kernels, which are designed to find lines in an image. The edge map is
shown in Figure 4.15b.
4.2.1.3
Hough transform
The Hough transform is a technique which can be used to isolate features of a particular
shape within an image. In this section, it is used to locate the iris outer boundary. The
Hough transform is applied directly on an edge map (calculated in section (4.2.1.2)) to
reduce processing time. The Hough transform represents an image in terms of a threedimensional accumulator array. For example, circles correspond to the equation
which defines a circle of center
and radius
space. For this specific feature, the accumulator array will contain the
which are updated for each edge pixel
in the
parameters
. After updating the parameters for all pixels
in the edge map, peaks in the accumulator array indicate the location of the desired
feature (i.e. circle). From this information, the iris boundary is located (Figure 4.15c) and
the iris-pupil region is segmented for further processing.
118
4.2.1.4
Enhanced iris segmentation
The iris contour given using the Hough transform technique in section (2.8.4.2) is an
approximation of the proper iris boundary which is not always circular. To correct the
offset on the iris contour, we recalculate a new contour using the edge map obtained in
section (4.2.1.2). Starting at the center position of the approximated contour, scan
outwards for the first set of pixels, different from black, that form a closed contour near
the boundary computed in section (4.2.1.2). This process is shown in Figure 4.15b. The
exact pupil-iris segmented region is shown in Figure 4.15d.
4.2.2 Eyelash detection
In this section, separable and multiple eyelashes are detected using an intensity gradient
based algorithm and a block mean and variance method (Figure 4.3). The iris and non-iris
images (Figure 4.15 (d, f)) are processed independently for improved image enhancement
(Figure 4.15(e, g)) and precise eyelash detection based on the local region statistics and
finally, the computed eyelash points are combined for exact iris region extraction.
Original image
Non-linear image
enhancement
Iris
segmentation
Image
Binarization
Iris-pupil
localisation
Edge
Detection
Hough
Transform
Figure 4.1: Iris segmentation algorithm based on local image enhancement
119
4.2.2.1
Local image enhancement
The iris region (Figure 4.15(d)) and the non-iris image (Figure 4.15(f)) are enhanced
separately using the method described in section (4.2.1.1). This will improve eyelash
detection since it depends on the local image statistics. The iris region tends to contain
higher intensity variation than the overall eye image. The enhancement results are shown
in Figure 4.15(e, g), respectively.
4.2.2.2
Separable eyelashes
In order to detect separable eyelashes in the horizontal, vertical and diagonal direction,
the enhanced iris image obtained in section (4.2.2.1) is convoluted with the developed
masks (Figure 4.2) as shown in equation (4.5). An image with all possible eyelash points
is created. A possible eyelash candidate point is set to “ ” when the mask response
is
negative and to “ ”, otherwise.
A final eyelash map is created by selecting strong eyelash candidates only. This is
obtained by taking
where
is a threshold selected to be
in our
experiment. The following operation is used in order to locate the eyelash pixels:
(4.5)
where
is a
is the original image,
is the mask response at position
and
convolution mask. This condition also satisfies the connective
120
criterion (Huang et al. (2004)) which states that the eyelash is a continuous line instead of
unconnected points. If a detected eyelash point is singled out as independent with no
connection to another eyelash or eyelid pixel, then this point is removed from the
candidates of potential eyelash pixels.
Using the masks shown in Figure 4.2, a negative mask response is obtained if and only if
the center pixel is located between two adjacent eyelash points which satisfy the
following connectivity criterion: if the center pixel is surrounded by non-eyelash points,
the convolution operation will result in a value greater than the selected threshold which
indicates that the pixel is not an eyelash point.
Figure 4.2: Four different masks for detecting horizontal, vertical and diagonal edges, respectively.
4.2.2.3
Multiple eyelashes
For regions containing multiple eyelashes, the mean and variance of an
region is
taken to detect eyelash candidates. These regions are generally composed of lower
intensity pixels with a higher variance. In order to find eyelash candidates in these
regions, the computed block mean
compared to different thresholds. If
(equation 4.6) and variance
or
(equation 4.7) are
then the center pixel in block is
considered to be an eyelash point. In our implementation, the block size used is
.
121
(4.6)
(4.7)
Iris
region
Non-linear image
enhancement based
on local region
statistics
Non-iris
region
Multiple eyelash detection
Separable eyelash detection
Final iris region
segmentation
Eyelash data points
Fusion
Figure 4.3: Eyelash detection algorithm and ideal iris region segmentation.
4.2.2.4
Specular Reflection
One way to locate the specular reflection points is by setting a threshold
on the iris
pixel values.
(4.8)
For the transition area from the strong reflection to the surrounding iris region, the
reflection intensities are relatively weaker, and they are still considered as noise instead
122
of genuine iris pixels. The detected area is considered noise and will not be used for
further processing.
4.2.3 Enhanced Iris Recognition
This section describes the iris recognition scheme developed and used on the enhanced
iris images where most of the eyelash and eyelid occlusions are eliminated. This section
also describes the feature extraction method applied on the clean iris area followed by the
phase quantization and pattern comparison using the Hamming distance matching metric.
First of all, the enhanced iris image is unwrapped into a normalized polar scale template,
according to the Daugman rubber sheet model described in section (2.8). Using the
method developed in section (2.8.7), a mask indicating the location of unwanted artifact
pixels is created in order to locate only the good iris pixels. This mask is also mapped
from the Cartesian coordinates to normalized polar coordinates. Finally, the performance
of our developed technique is evaluated and compared to the Masek’s approach, which
does not perform any eyelash detection prior to iris recognition (Masek (2003), Xie
(2007)).
4.2.3.1
Enhanced Segmentation
First of all, it is important to mention that the traditional way of segmenting the iris image
consists of locating the iris and pupil boundary circles, and the parabola of the upper and
lower eyelids, using the Hough transform. For example, Figure 4.4 shows an image taken
123
from the CASIA database where the iris region is segmented using the Masek’s algorithm
(Masek et al. (2003)). Clearly, the image shows that the iris region is not well segmented
since eyelash occlusions are overestimated, hence losing major iris information. This is
shown in Figure 4.5 where the shaded regions covering the top and bottom of the iris
image discard valid underlying iris pixels, classifying them as noise. The discarded pixels
are replaced by the average pixel intensity value of the valid iris region for subsequent
feature extraction.
On the other hand, the proposed method accurately detects the eyelash occlusions and
classifies them as noise. It aims on improving the result of Figure 4.5 by eliminating the
eyelashes outside the mask, and identifying all the valid iris pixels that were discarded as
noise in the Masek’s method. Using the proposed eyelash detection technique, iris
segmentation results with enhanced eyelash detection are shown in Figure 4.6. The
eyelashes are accurately detected, and the iris pixel locations occluded by the eyelashes
are discarded as noise. In addition, valid iris pixels located between the eyelashes are kept
for the subsequent feature extraction and the pattern matching. Hence, eyelash occlusion
is not overestimated in our proposed technique since it reduces unwanted iris artifacts
while revealing the iris pixels previously ignored in the Masek's segmentation.
124
Figure 4.4: Iris image taken from the CASIA database.
Figure 4.5: Iris segmentation using the Masek’s algorithm. As seen in the image, the iris region
includes some eyelash occlusion. In addition, some of the valid iris pixels are discarded as noise.
125
Figure 4.6: Iris segmentation using the developed enhanced eyelash detection algorithm.
4.2.4 Iris Unwrapping
Once the eyelashes are detected and iris area accurately segmented, it is necessary to
unwrap the iris region using the rubber sheet model from the image space to the
rectangular polar space matrix described in section (2.8). The unwrapping operation maps
each point in the Cartesian coordinates
to the polar coordinates
where
represents the radial distance from the center of pupil, and describes the angular shifting
from
to
. The feature points are selected from the iris image with a specified
sampling resolution of
direction and
where
feature points are selected along the radial
along the circumferential direction. The rubber sheet model
representation makes it easier to use the Gabor filter in order to create the iris feature
templates for pattern comparison. The mapping function, described in (Masek (2003)),
can be written as follows:
126
(4.9)
(4.10)
where
and
and
represent the Cartesian and polar coordinates, respectively.
correspond to the Cartesian coordinates of the two iris radius end points. An
example of iris unwrapping is shown in Figure 4.7. Each row in the unwrapped iris
template corresponds to one circle in the iris region in the Cartesian scale. The
unwrapped and normalized polar iris representation is robust since it compensates for
pupil-iris center displacement and iris size. It provides uniform sampling of the feature
points across the iris region where
points are taken along the radial direction and
points along the circumferential direction. This is shown as white dots in Figure 4.8 and
Figure 4.9 where the sampling process was completed using the Masek and the enhanced
iris segmentation techniques, respectively. The iris feature points covered by eyelids,
eyelashes, and other artifacts are assigned an interpolated value before encoding. This
corresponds to an average pixel value that corresponds to the average of all other valid
iris pixel values within the segmented iris image (Masek (2003)). Some examples of
unwrapped iris images are shown in Figure 4.10 and Figure 4.11. Once the iris region is
unwrapped using the rubber sheet model, the rectangular template is then convolved, row
by row, with a 1D Log-Gabor filter in order to extract iris feature vectors. The output of
the convolution produces a complex value for each pixel point that is located in one of
the four complex phase domains. In order to create a binary template, at each pixel
location, phase quantization is applied as described in section (2.8). This will generate a
binary pair for each complex number (i.e. every pixel). In a similar way, all the iris Log-
127
Gabor feature complex values are encoded with a binary pair in order to create the feature
template that is used afterwards for pattern matching using the Hamming distance.
However, before proceeding with the pattern comparison operation, the noise pixels are
discarded using a binary mask template that is of the same size as the iris features
template. In the binary mask, calculated in section (4.2.3.1), a binary ‘ ’ (i.e. black pixel)
indicates the location of an unwanted pixel (i.e. noise) in the iris template (Figure 4.12
and Figure 4.13). Therefore, the quadrature phase information for each pixel is given by
, a complex feature
two binary values. Hence, for each iris template of size
template of the same size is generated, which is then represented by a binary phase
template of size
. The phase quantization process is repeated for all the pixels
in the complex iris template in order to discard all the feature points related to noise
and/or eyelash occlusions during the Hamming distance calculation.
r
Pupil
r
Sclera
Upper eyelid
Lower eyelid
Figure 4.7: Example of the iris region unwrapping using the rubber sheet model described in Xie
(2007).
128
Figure 4.8: Iris feature sampling points using the Masek’s iris segmentation scheme.
Figure 4.9: Iris feature sampling points using the enhanced iris segmentation scheme.
129
Figure 4.10: Iris region unwrapping using the Masek’s technique.
Figure 4.11: Iris region unwrapping using the enhanced technique.
Figure 4.12: Example of a binary mask applied on the unwrapped iris image in the Masek’s
segmentation technique. The binary 0s (black pixels) indicate noise and are used to discard the
underlying pixel values in the iris template. The pixel value corresponding to 1s (white) are used for
the Hamming distance calculation.
Figure 4.13: Example of a binary mask applied on the unwrapped iris image in the enhanced
segmentation technique. The mask shows accurate eyelash and noise detection.
130
4.2.5 Pattern Matching
Pattern matching is crucial in order to determine if two biometric iris templates
correspond to the same eye (i.e. intra-class comparison) or if they come from two
different eyes (i.e. inter-class comparison). Pattern matching involves the selection of a
distance measure and a decision threshold. Since intra-class comparison involves distance
calculation between two iris templates taken from the same subject, it generally results in
smaller distances compared to the inter-class comparisons. In this work, the Hamming
distance is used as a matching metric (section (2.8.7.1)) which measures the closeness of
match between the intra-class and inter-class. The Hamming distance uses the binary iris
template and its corresponding binary mask in the distance calculation. Once all the noise
pixels are identified by the binary “ ” in the mask, the rest of the binary pixels are used in
the XOR operation between the two valid iris feature templates. The Hamming distance
is then computed as the average of the XOR resulting operation. After computing all the
Hamming distances between the iris templates, the smallest distance is used to determine
a match and to select the right class for which that subject belongs. In order to obtain
smaller Hamming distances, circular shifting between the encoded iris templates is
required. This will compensate for rotation invariance by row-wise circularly shifting
(left and right) the encoded templates. This operation will compensate for circular
rotation of the iris in the original Cartesian coordinates since each row in the binary iris
template corresponds to one circular contour in the original iris region (Masek (2003)).
131
4.3 Results
The algorithm was tested on iris images taken from the CASIA database which
comprised
gray-scale iris images in bit-map format. From the entire database,
iris images, partially occluded by eyelashes, were used to test the developed algorithm.
Figure 4.16(d,e,f) show accurate detection of eyelashes in different images. Figure
4.16(g,h,i) illustrates an accurate segmentation of the iris region using the enhanced
segmentation method proposed in this thesis. The following parameter values were
chosen in this process:
, and
was
set to the variance of the entire region. These values were selected after running few tests
with the algorithm in order to determine the sensitivity of the results to the parameter
choices, and results did not vary significantly for a wide range of parameter choices near
the values used. Figure 4.17 shows normalized iris regions for (a) an iris image without
segmentation, (b) an iris with eyelash segmentation using Masek’s technique and (c) an
iris image obtained with our enhanced segmentation algorithm. In the normalized iris
region obtained using the enhanced technique, fewer iris noise and eyelash occlusion is
present in the image compared to the other two. From this, Figure 4.18 shows the
genuine-impostor distributions computed using 327 iris images for (a) the Masek’s
algorithms and (b) the enhanced algorithm. This graph is obtained by plotting the intraclass and inter-class distributions. An overlap exists at the tail from which the FAR and
FRR can be calculated. In order to quantify the improvement that the enhanced method
brought to the developed system, the decidability measure described in section (2.8.8.3)
is calculated. The latter method measures the separation between the intra-class and
inter-class Hamming distance distributions (Figure 4.19). It is seen that a higher
132
decidability measure is obtained for the enhanced algorithm which implies that the iris
recognition system implementing the enhanced eyelash detection scheme tends to have a
better intra and inter-class distribution separation (Masek (2003)). This also implies a
lower false accept rate and false reject rate since it is easier to select a threshold that
provides a good genuine-impostor distribution separation.
Figure 4.20 shows the detection error trade-off curve (DET) plotting the false non-match
rate against the false match rate. The red curve represents the DET curve of the Masek's
segmentation and the green curve represents the DET curve of the enhanced
segmentation using the enhanced eyelash detection technique. From this curve, it is
noticed that the enhanced segmentation technique outperforms the Masek’s method for
most of the cases except in the very low FMR range where it becomes very close.
Therefore, the match score using the enhanced technique is better within the entire
comparison range since it uses more of the valid iris region in the matching operation by
discarding most of the eyelash occlusions and other iris distortions. Masek’s code
neglects some valid iris pixels since it overestimates eyelash occlusion and noise pixels.
This will result in a loss of valid iris pixels which will eventually decrease the system’s
performance. Thus, the enhanced algorithm can be applied when the iris images are
acquired in a more challenging or uncontrolled environment (i.e. at a distance) where
more noise and distortion can occur in the images as a result of camera tilting or lack of
client cooperation.
In addition, it is also shown that the identification rate is improved from 93.67% (Masek)
to 95.25% as illustrated in the cumulative match curve in Figure 4.21. The green curve
133
represents the cumulative match curve without any eyelash detection and the red curve
represents the one with eyelash detection. It is clearly shown that the Rank-1
identification rate is improved.
(a)
(b)
134
(c)
Figure 4.14: Image enhancement result: (a) Original image of the eye, (b) Non-linear image
enhancement, (c) Binarized image
(a)
135
(b)
(c)
(d)
(e)
136
(f)
(g)
Figure 4.15: Accurate iris boundary extraction and enhancement: (a) Approximated location of the
iris outer boundary using the Hough transform, (b) Edge map and accurate iris boundary
calculation, (c) Accurate pupil-iris boundary extraction, (d) Exact Pupil-iris region segmentation, (e)
Pupil-iris local region enhancement, (f)Non-iris eye image, (g) Non-iris local image enhancement.
137
(a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
(i)
Figure 4.16: Eyelash detection and iris segmentation examples: (a, b, c) Original eye images, (d, e, f)
Computed candidate eyelash points using our algorithm, (g, h, i) Accurate segmentation of the iris
regions without eyelash occlusions.
138
(a)
5
10
15
Radial resolution (pixels)
20
20
40
60
80
100
120
20
40
60
80
100
120
20
40
60
80
100
120
(b)
5
10
15
20
(c)
5
10
15
20
Circumferential resolution (pixels)
Figure 4.17: Normalized iris images in the polar space. (a) shows a normalized iris region without
eyelash detection, (b) shows the result of iris normalization using the Masek’s algorithm and (c)
represents normalization using the enhanced segmentation algorithm. The y-axis represents the
radial resolution and the x-axis corresponds to the circumferential (circular) resolution.
139
(a)
Genuine Impostor Distribution for masek Segmentation
0.25
GENUINE
IMPOSTOR
Percentage
0.2
0.15
0.1
0.05
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Hamming distance
(b)
Genuine Impostor Distribution for enhanced Segmentation
0.25
GENUINE
IMPOSTOR
Percentage
0.2
0.15
0.1
0.05
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Hamming distance
Figure 4.18: (a) Plot of the intra-class and inter-class distribution using the Masek’s segmentation
algorithm, (b) Genuine-impostor distributions using the enhanced segmentation algorithm.
140
3.67
3.66
Enhanced method
Decidability Measure
3.65
3.64
3.63
3.62
Masek's method
3.61
3.6
0
10
20
30
40
50
60
# of samples
70
80
90
100
Figure 4.19: The decidability ( ) measure showing a higher decidability measure for the enhanced
algorithm.
141
(a)
(b)
Figure 4.20: (a) DET curve and (b) log-log plot showing the FRR vs FAR for the Masek's and the
enhanced segmentation methods, respectively. An incremental improvement is seen when using the
enhanced segmentation technique which results in a smaller EER.
142
Figure 4.21: Cumulative Match Curve comparison. A 95.25% rank-1 identification rate is obtained
using our proposed method while 93.67% is calculated using Masek’s technique.
4.4 Discussion
The proposed algorithm shows promising results for eyelash noise detection, accurate iris
boundary extraction and ideal iris segmentation. The results obtained using the enhanced
algorithm do not show particularly large increases in performance compared to the
existing iris segmentation schemes since the latter existing segmentation/identification
algorithm is already extracting enough iris information to achieve high recognition rates.
This explains the small increase in our enhanced identification rate but still shows an
important improvement. The proposed algorithm locates the iris region using logarithmic
image enhancement and the Hough transform techniques, locates the iris boundary,
extracts the exact iris contour, detects eyelash based on the local image statistics and
143
block intensity and finally, proposes an iris model for accurate iris recognition. The
developed method overcomes the limitations encountered in other iris segmentation and
eyelash detection techniques such that our method detects accurately separable and
multiple eyelashes, extracts the exact iris contour and is illumination invariant. In
addition, the developed method does not overestimate the eyelash noise in the eye image,
maximizing iris information. Using our technique, fewer iris noise and eyelash occlusions
are found in the normalized iris region compared to the Masek approach. Also, the
decidability metric is calculated for both methods in order to measure the separation
between the intra-class and inter-class Hamming distance distributions. It is seen that a
higher decidability measure is obtained for the enhanced algorithm which implies that the
iris recognition system implementing the enhanced eyelash detection scheme tends to
have a better intra and inter-class distribution separation. In addition, this will result in
lower false accept rate and false reject rate since it is easier to select the decision
threshold that provides a good genuine-impostor distribution separation. Furthermore, the
match score using the enhanced technique is better within the entire comparison range
since it uses the most of the iris region in the pattern matching process. Masek’s code
neglects some valid iris pixels since it overestimates eyelash occlusion and noise pixels.
This will result in a loss of valid iris pixels which will probably decrease the system’s
performance.
144
4.5 Summary
This work proposes a new system for eyelash noise detection, accurate iris boundary
extraction and ideal iris segmentation. The results obtained using the enhanced algorithm
show incremental but important improvement in performance compared to the existing
iris segmentation schemes. The proposed algorithm locates the iris region using
logarithmic image enhancement and the Hough transform techniques, locates the iris
boundary, extracts the exact iris contour, detects eyelash based on the local image
statistics and block intensity and finally, proposes an iris model for accurate iris
recognition. Using our technique, fewer iris noise and eyelash occlusions are found in the
normalized iris region compared to the Masek approach. In addition, a higher decidability
measure is obtained for the enhanced algorithm which implies that the iris recognition
system implementing the enhanced eyelash detection scheme tends to have a better
genuine-impostor distributions separation. Finally, it is also shown that the rank-1
identification rate improved from 93.67% (Masek) to 95.25% (enhanced method) as
illustrated in the cumulative match curve in Figure 4.21.
145
5 Chapter 5
Measuring Information Content in Biometric Features
5.1 Introduction
How much information is there in a face, or a fingerprint? This question is related to
many issues in biometric technology. For example, one of the most common biometric
questions is that of uniqueness, e.g. to what extent are fingerprints unique? From the
point of view of identifiability, one may be interested in how much identifying
information is available from a given technology, such as video surveillance. In the
context of biometric fusion (Ross et al. (2003)) one would like to be able to quantify the
biometric information in each system individually, and the potential gain from fusing the
systems. Additionally, such a measure is relevant to biometric cryptosystems and privacy
measures. Several authors have presented approaches relevant to this question. For
example, Wayman (2004) introduced a set of statistical approaches to measure the
separability of Gaussian feature distributions using a “cotton ball model”. Another
approach is developed by (Daugman (2003)) to measure the information content of iris
images based on the discrimination entropy (Cover et al.(1991)), which is calculated
directly from the match score distributions. Also, Golfarelli et al. (1997) showed that the
most commonly used feature representations of hand geometry and face biometrics have
a limited number of distinguishable patterns, on the order of
and
, respectively, as
146
measured by a theoretical estimate of the equal error rate. In this work, we elaborate an
approach to address this question based on definitions from information theory. We
define the term “biometric information” as follows:
Biometric information (
): the decrease in uncertainty about the identity of a
person due to a set of biometric features measurements.
In order to interpret this definition, we refer to two instants: 1) before a biometric
measurement,
, at which time we only know a person
is part of a population , which
may be the whole planet; and 2) after receiving a set of measurements, , we have more
information and less uncertainty about the person’s identity.
may be used to answer
two different types of questions. First, given a set of measurements from a specific
person, we want to know how identifiable that individual is in a population. This is the
individual biometric information (
). Second, given a system which makes biometric
measurements, such as fingerprint minutiae or eigenfaces, we want to know, on average,
how distinguishable people are in the population, using those biometric features. This is
the system biometric information (
an individual’s features and
). The difference is that
is the information of
is the average information over the population. In order
to motivate our approach, we initially consider the properties that such a measure should
have. Consider a soft biometric system which measures height and weight; furthermore,
assume all humans are uniformly and independently distributed in height between
cm and weight between
-
-
lb. If a person’s features were completely stable and
could be measured with infinite accuracy, people could be uniquely identified from these
147
measurements, and the biometric features could be considered to yield infinite
information. However, in reality, repeated biometric measurements give different results
due to measurement inaccuracies and to short- and long-term changes in the biometric
features themselves. If this variability results in an uncertainty of ± 5 cm and ±5 lb, one
simple model would be to round each measure to 105, 115, ..., 195. In this case, there are
equiprobable outcomes, and an information content of
. Such
an analysis is intrinsically tied to a choice of biometric features. Thus, our approach does
not allow us to answer “how much information is in a fingerprint?”, but only “how much
information is in the position and angle data of fingerprint minutiae?”. Furthermore, for
many biometrics, it is not clear what the underlying features are. Face images, for
example, can be described by image basis features or landmark based features (Zhao et
al. (2003)). To overcome this, we may choose to calculate the information in all possible
features. For example, we may provide height in inches as well as cm; however, in this
case, a good measure of information must not increase with such redundant data.
This work also develops a new approach to measuring Biometric image quality.
Biometric sample quality is a measure of the usefulness of a biometric image (ISO JTC1
SC37 (2007)). One recent development is the significant level of interest in standards for
measurement of biometric quality. For example, ISO has recently established a biometric
sample quality draft standard (ISO JTC1 SC37 (2007)). According to (ISO JTC1 SC37
(2007)), biometric sample quality may be considered from the point of view of character
(inherent features), fidelity (accuracy of features), or utility (predicted biometrics
performance). A general consensus has developed that the most important measure of a
quality metric is its utility – images evaluated as higher quality must be those that result
148
in better identification of individuals, as measured by an increased separation of genuine
and impostor match score distributions. The nature of biometric sample fidelity has seen
little investigation, although for specific biometric modalities, algorithms to measure
biometric quality have been proposed. For example, the NFIQ algorithm (Tabassi et al.
(2004)) is a widely used measure for fingerprint image quality.
One current difficulty is that there is no consensus as to what a measure of biometric
sample fidelity should give. In this work, we propose a new approach to measure this
quantity, based on an information theoretic framework. We begin with the intuitive
observation that a high quality biometric image is more useful to identify the individual
than a low quality image. This suggests that the quantity of identifiable information
decreases with a reduction in quality. Given a way to measure the decrease in information
caused by a given image degradation, one can measure the associated decrease in
biometric information. In this chapter, we develop a mathematical framework to measure
biometric feature information in a given system for a set of biometric features. We
address ill-conditioning in the measurements using distribution modeling and
regularization. We then use this algorithm to analyze the biometric information content of
two different face recognition algorithms and then define the information loss due to
degradation in image quality.
In addition, this chapter presents a new method to calculate the iris biometric feature
information using the system described in section (2.8) and the relative entropy measure.
The developed algorithm is divided into the following sections: i. distribution modeling
of iris biometric features, ii. relative entropy calculations, iii. ICA iris feature extraction
and biometric information calculation. The iris regions used in the entropy calculation are
149
obtained using the Masek and the enhanced iris segmentation techniques described in
section (4.2).
5.2 Theoretical framework
In this section we develop an algorithm to calculate biometric information based on a set
of features, using the relative entropy measure (Cover et al. 1991). We then measure the
effect of an image degradation model on biometric image quality. We explain our method
in the following steps: 1) measure requirements, 2) relative entropy of biometric features,
3) Gaussian models for biometric features and relative entropy calculations, 4)
regularization methods for degenerate features, 5) regularization methods for insufficient
data, 6) information loss due to degradation.
5.2.1 Requirements for biometric feature information
In order to elaborate the requirements that a good measure of biometric feature
information must have, we consider the system that measures height and weight. These
values differ within the global population, but also vary for a given individual, both due
to variations in the features themselves and to measurement inaccuracies. We now wish
to consider the properties a measure of biometric feature information should have:
1. If an intra-person distribution
is exactly equal to the inter-person
distribution, then there is no information to distinguish a person, and biometric
feature information is zero.
150
2. As the feature measurement becomes more accurate (less variability), then it is
easier to distinguish someone in the population and the biometric information
increases.
3. If a person has unusual feature values (i.e. far from the population mean), they
become more distinguishable, and their biometric feature information will be
larger.
4. The biometric information of uncorrelated features should be the sum of the
biometric information of each individual feature.
5. Features that are unrelated to identity should not increase biometric
information. For example, if a biometric system accurately measured the direction
a person was facing, information on identity would be unchanged.
6. Correlated features such as height and weight are less informative. In an
extreme example consider the height in inches and in cm. Clearly, these two
features are no more informative than a single value (except perhaps a reduction
in noise from the averaging of repeated measurements).
Based on this definition, the most appropriate information theoretic measure for the
biometric feature information is the relative entropy (
between the intra- (
) and inter-person (
) (Cover et al. (1991))
) biometric feature distributions.
,
or the Kullback-Leibler distance, is defined as the measure of the information gain in
moving from a prior distribution
to a posterior distribution
bits” of information needed to represent
with respect to
, or to be the “extra
.
is defined to
be
151
(5.1)
where the integral is over all feature dimensions, .
or distribution of features of an individual and
A comment on notation: we use
is the probability mass function
is the overall population distribution.
to refer to both an individual person, and the
distribution of the person’s features, while
represents the population and the
distribution of its features. This measure can be motivated as follows: the relative
entropy,
, is the extra information required to describe a distribution
on an assumed distribution
(Cover et al. (1991)).
differs from the entropy,
, which is the information required, on average, to describe features
.
based
distributed as
is not in itself an appropriate measure for biometric feature information, since it
does not account for the extent to which each feature can identify a person
in a
population . An example of a feature unrelated to identity is the direction a person is
facing. Measuring this quantity will increase
of a feature set, but not increase its ability
to identify a person. The measure
corresponds to the requirements: given
knowledge of the population feature distribution , the information in a biometric feature
set allows us to describe a particular person .
5.2.2 Distribution modeling
In a generic biometric system,
feature vector
biometric features are measured, to create a biometric
for each person. For person , we have
features samples,
152
while we have
samples for the population. For convenience of notation, we sort ’s
measurements to be the first grouping of the population. Defining
as an instance of
random variable , we calculate the population feature mean
(5.2)
where the feature mean of person ,
population feature covariance
, is defined analogously, replacing
by . The
is
(5.3)
The individuals feature covariance,
, is again defined analogously. One important
general difficulty with direct information theoretic measures is that of data availability.
Distributions are difficult to estimate accurately, especially at the tails; and yet
will give large absolute values for small
or
. Instead, it is
typical to fit data to a model with a small number of parameters. The Gaussian
distribution is the most common model; it is often a good reflection of the real world
distributions, and is analytically convenient in entropy integrals. Another important
property of the Gaussian is that it gives the maximum entropy for a given standard
deviation, allowing such models to be used to give an upper bound to entropy values.
Based on the Gaussian model, which seems to be the simplest and appropriate for
and
, we write:
153
(5.4)
(5.5)
From which we can calculate
as:
(5.6)
where
and
.
This expression calculates the relative entropy in bits for Gaussian distributions
and
. This expression corresponds to most of the desired requirements for a biometric
feature information measure introduced in the previous section:
1. If person’s feature distribution matches the population,
; this yields
, as required.
2. As feature measurements improve, the covariance values,
resulting in a reduction in
, and an increase in
, will decrease,
.
154
3. If a person has feature values far from the population mean,
will be larger,
.
resulting in an larger value of
4. Combinations of uncorrelated feature vectors yield the sum of the individual
measures. Thus, for uncorrelated features
represents
concatenation
of
and
the
, where
feature
5. Addition of features uncorrelated to identity will not change
feature will have an identical distribution in
uncorrelated features,
and
and . If
for
. Such a
is the set of such
and
. Under these conditions,
vectors,
while
will be identical
to its value when excluding the features in . One way to understand this criterion
is that if the distributions for
and
differ for features in
, then those features
can be used as a biometric to help identify a person.
6. Correlated features are less informative than uncorrelated ones. Such features
will decrease the condition number (and thus the determinant) of both
This will decrease the accuracy of the measure
perfectly correlated features,
and
.
. In the extreme case of
becomes singular with a zero determinant and
is undefined. Thus, our measure is inadequate in this case.
In the next section, we develop an algorithm to deal with this effect.
155
5.2.3 Regularization Methods for degenerate features
In order to guard against numerical instability in our measures, we wish to extract a
mutually independent set of
“important” features (
). To do this, we use the
principal component analysis (PCA) (Draper et al.(2003)) (Grother (2000)) to generate a
mapping (
space
), from the original biometric features
of size
(
) to a new feature
. The PCA may be calculated from a Singular Value
Decomposition (SVD) (Alter et al. (2000)) of the feature covariance matrix, such that
Since
is positive definite,
is orthonormal and
is diagonal. We choose to perform
the PCA on the population distribution , rather than , since
is based on far more data,
and is therefore likely to be a more reliable estimate. The values of
significance of each feature in PCA space. A feature , with small
indicate the
will have very
little effect on the overall biometric feature information. We use this analysis, in order to
regularize
, and to reject degenerate features by truncating the SVD. We select a
truncation threshold of
truncated to be
where
, and
is truncated to
. Based on this threshold,
. Using the basis
is
calculated from
the population, we decompose the individual’s covariance into feature space
:
(5.7)
where
is not necessarily a diagonal matrix. However, since
similar data, we expect
and describe somewhat
to have a strong diagonal component comparable to
, as seen
in Figure 5.4.
156
Based on this regularization scheme, equation (5.6) may be rewritten in the PCA space
as:
(5.8)
where
and
5.2.4 Regularization Methods for insufficient data
The expression developed in the previous section solves the ill-posed nature of
However,
may still be singular in the common circumstance in which only a small
images of an individual
number of samples of each individual are available. Given
from which
in
.
features are calculated,
diverging to
will be singular if
, which will result
. In practice, this is a common occurrence, since most biometric
systems calculate many hundreds of features, and most biometric databases contain far
less samples for each person. In order to address this issue, we develop an estimate which
may act as a lower bound. In order to do this, we make the following assumptions:
1. Estimates of feature variances are valid
2. Estimates of feature covariances
important
features, where
for all .
for
are only valid for the most
. Features which are not considered valid
based on these assumptions, are set to zero by multiplying
by a mask
,
where
157
(5.9)
This expression regularizes the intra-person covariance,
, and assures that
not diverge. To clarify the effect of this regularization on
feature covariance will decrease
diverging to
does
, we note that intra-
toward zero, leading a differential entropy estimate
. We thus consider this regularization strategy to generate a lower bound
on the biometric feature information. The selection of
is a compromise between using
all available measurements (by using large ) and avoiding numerical instability when
is close to singular (by using small ).
5.2.5 Average information of a biometric system
This section has developed a measure of biometric feature information content of a
biometric feature representation of a single individual with respect to the feature
distribution of the population. As discussed, the biometric feature information will vary
between people; those with feature values further from the mean have larger biometric
feature information. In order to use this approach to measure the biometric feature
information content of a biometric system, we calculate the average biometric feature
information for each individual in the population (weighted by the probability of needing
to identify that person, if appropriate). This is a measure of the system biometric
information (
) which can be calculated by the average
over the population .
158
(5.10)
5.2.6 Information loss due to degradation
In this section, we explore the effect of image degradation and the resulting decrease in
biometric quality on the relative entropy measure. Intuitively, it is expected that image
degradation changes the intra and inter person distribution of the face features resulting in
a loss of biometric information. Given a degradation process, we wish to measure how
much
is lost in the degraded images,
, versus the original images,
to measure the severity of a degradation process. Features
degraded images
. This allows us
are then extracted from the
using three feature extraction methods given. We then compute the
biometric information for the non-degraded distributions
degraded distributions
and for the
using equation (5.8). Here
represents the relative entropy between the individual and population distribution prior to
degradation while
is the relative entropy measure between the degraded
individual and population distributions, respectively. From this, we calculate the
normalized mean square distance characterizing the loss of information caused by the
degradation model on the underlying features as:
159
(5.11)
where
is the variance of
.
measures the relative distance offset
between the original and degraded distributions.
interpreted as the fractional loss in
is a unitless measure, and may be
due to a given image degradation. In order to
motivate this calculation, we initially considered calculating
as a function
of degradation. Surprisingly, this measure increases with decreasing quality. The
reasoning behind this stems from the fact that when a person’s ( ) feature distribution is
degraded and compared to a high quality population feature distribution, the algorithm
seems to be saying: Aha! I can recognize . He always has a blurry face!. Therefore, it is
necessary to compare a degraded person’s image to the degraded population
in order to compensate for this effect.
5.3 Face recognition
Information in a feature representation of faces is calculated using our described method
for different individuals. In order to test our algorithm, it is necessary to have multiple
images of the same individual. Using the Aberdeen face database (Belhumeur et al.
(1997)), we chose 18 frontal images of 16 persons, from which we calculate the PCA
(eigenface) features using the algorithm of (Grother (2000)) and the FLD face features
components using the algorithm described in (Xiang et al. (2004)). Initially, all face
images were registered by rotation and scaling to have eye positions at
and
160
. Images were then cropped to
cover the intensity range
pixels and histogram equalized to
. The same set of operations is applied to all images using
the same thresholds. This results with the same effect on all images when computing the
biometric feature information. Features are calculated from a set of
images using
different component analysis methods such as Principal Component Analysis (PCA, also
referred to as Eigenface features) (Grother (2000)) (Turk & Pentland (1991)) and Fisher
linear discriminant (FLD) (Li et al. (2005)).
and individual mean distributions, while
and
and
are
are
vectors of the population
matrices of the individual
and population covariance matrices.
The feature decomposition process was conducted on
giving
images of each of
total images. For PCA and Fisher feature decompositions,
vectors were computed, and the most significant
persons,
separate
features per image used for
subsequent analysis. Figure 5.1 and Figure 5.2 illustrate PCA and FLD features,
respectively. From this,
which assumes that
and
is computed for each of
persons using equation (5.8),
have Gaussian distributions. In order to test the validity of
the Gaussian model for our data, we use the following normality tests:
• Kolmogorov-Smirnov test: compares the distributions of values in the two data vectors
and
, where
represents random samples from the underlying distribution and
follows an ideal Gaussian with zero mean and variance. The null hypothesis is that
and
are drawn from the same continuous normal distribution. We reject the null
hypothesis at
.
161
• The Lilliefors test (Conover (1980)): evaluates the hypothesis that
has a normal
distribution with unspecified mean and variance, against the alternative that
have a normal distribution. This test compares the empirical distribution of
normal distribution having the same mean and variance as
hypothesis at
does not
with a
. We reject the null
.
Figure 5.1: An example of PCA (Eigenface) face features. From left to right, PCA features number 3,
15, 35, 55 are shown. The PCA features are othonormal and fit the data in a least squares sense.
Figure 5.2: An example of FLD face features. From left to right, FLD features number 7, 10, 30, 50
are shown. FLD attempts to maximize class separation while minimizing the within class scatter.
Using these tests, an average of
and
for the FLD and PCA features marginal
distributions are normally distributed.
162
5.3.1 Biometric information calculations
After fitting the distributions of
and
to a Gaussian model, we initially analyze
the biometric feature information in each PCA and FLD feature separately. PCA features
are shown in Figure 5.3, and show a gradual decrease from an initial peak at feature 2.
The form of the curve can be understood from the nature of the PCA decomposition,
which tends to place higher frequency details in higher number features. Since noise
tends to increase with frequency, the biometric information in these higher numbered
PCA features will be less. A sum of biometric feature information over the first
features for one individual gives
PCA
bits. This does not assume statistical independence
nor uncorrelatedness of PCA coefficients. Biometric feature information calculated using
FLD features seems to be similar to PCA features such that most biometric feature
information is computed for the most dominant fisherfaces. In order to calculate
for all features, we are limited by the available information. Since
for more than
used to calculate the covariances, attempts to calculate
will fail, because
images are
features
is singular. This effect is seen in the condition number (ratio of the
largest to the smallest singular value) which was
The relatively small condition number of
for
and
for
.
indicates that no features are degenerate for
PCA and FLD face recognition features. However,
is severely ill-conditioned. To
overcome this ill-conditioning, we introduced a regularization scheme based on a mask
(equation 5.9) with a cut-off point . This scheme is motivated by the diagonal structure
of
, as shown in Figure 5.4. To ensure convergence, the mask size
smaller than
is set to a value
.
163
We solve this singularity of equation (5.8) using a mask for
To further explore the effect of parameters
and
based on a parameter .
, we artificially reduce the
randomly eliminating some images from individuals. Results for
features for each person as a function of
are shown in Figure 5.5 for
by
for PCA
,
,
and
. In these curves, we observe a “hockey stick” shape. The relative entropy measure
remains stable when
, but if
, we observe a dramatic increase in
the algorithm approaches a singularity of
and the ill-conditioning of
is stable with a lower and upper bounds between
,
to
as
.
,
bits. However, when
estimates start diverging and reach very large values. Clearly, points for
greater than the knee in the hockey stick do not represent accurate estimates of
We also argue that when
approaches
our algorithm over-estimate
estimate
.
, the inherent ill-conditioning of
makes the
. On the other hand, small values of
will under-
, since these values will mask inter-feature correlations. This effect
increases | | as
decreases. However, the results suggest that this effect is minor,
especially in Figure 5.5a and Figure 5.5b, where the “base” of the hockey stick is more
flat. In order to produce an unique and stable estimate for
, it is necessary to
choose a compromise between these effects. We recommend choosing
larger value of
puts the estimate in an unstable region of Figure 5.4. Using this
algorithm and value of
, we calculate the overall biometric feature information for
different face recognition algorithms. For PCA features, the average
and for FLD features
(making
, since a
is
features in all), average
is
bits,
bits. If PCA and FLD features are combined
is
bits (Figure 5.6). This combination
164
of features illustrates that a biometric fusion of similar features may offer very little
information above that of the individual underlying features. It is initially somewhat
surprising that FLD feature information is measured to be lower than that from PCA.
This result may be understood because PCA features retain unwanted information due to
variations in facial expression and lighting, which are measured to contain useful
information, while FLD ”projects away” variations in lighting and facial expression while
maintaining the discriminant features. In order to understand the difference in the
discriminating effect between PCA and FLD features, the DET curves (Figure 5.7)
showing the FRR as a function of FAR are shown for PCA, FLD and PCA+FLD features,
respectively. Results show a smaller EER is obtained by fusing PCA and FLD features.
This effect is consistent with the calculation of the highest BI (
PCA (
bits) and FLD (
bits) compared to
bits) in Figure 5.6. Similarly, better system performance
was obtained when using PCA features compared to FLD which is also seen in the DET
curves plot of Figure 5.7. Therefore, the increase in BI calculation for a specific set of
features is reflected in a DET curve as a relative decrease in the EER. Higher biometric
feature information indicates that the feature set used in the biometrics system under test
contains additional discriminating information which should reduce the FAR and FRR
which is shown in Figure 5.7.
Moreover, feature decomposition using independent component analysis (ICA)
(Golfarelli et al. (1997)) was also conducted on the same set of faces. ICA has the
advantage that it does not only decorrelate the signals but also reduces higher-order
statistical dependencies in order to make the signals as statistically independent as
possible (Lee (1999)). Since ICA maximizes non-Gaussianity, it fits less well to the
165
assumptions of our model. For ICA features, an average of
bits was computed for
.
PCA
2
1.5
Biometric information (bits)
1
0.5
0
0
20
40
60
80
100
60
80
100
FISHER
2
1.5
1
0.5
0
0
20
40
Features
Figure 5.3: Biometric information (bits) as a function of number of features for (A) PCA (Eigenface)
and (B) FLD (bottom) face feature decomposition.
Figure 5.4: The regularized intra-person covariance matrix Sp showing dominant components along
its diagonal. Since Σp represents similar information to Σq it is reasonable to expect the matrices
have similar eigenvectors, resulting in strong diagonal components in Σp.
166
160
140
Biometric information (bits)
120
100
80
60
40
20
0
5
10
15
(b)
5
10
15
5
10
15
(d)
5
10
15
Biometric information (bits)
(a)
(c)
Mask size (L)
Mask size (L)
Figure 5.5: Biometric information (in bits) (y-axis) vs. the mask size (L) (x-axis) for each person.
Each subfigure represents a different value of Np (images of the same person): (A) 8, (B) 12, (C) 16
and (D) 18. The curves show that
diverges as
becomes singular (L ≥ Np). The relative
entropy increases with the size of the mask.
167
90
PCA
FISHER
PCA+FISHER
80
Average BI (bits)
70
60
50
40
30
20
0
2
4
6
8
10
12
14
16
18
Mask size (L)
Figure 5.6: Average
vs L (x-axis) for Np = 18. Each line represents the average of information
calculated for a population of 16 individuals with 18 images each using PCA (middle), FLD (bottom)
and a fusion of PCA and FLD features (top).
1
PCA
Fisher
PCA + Fisher
0.9
0.8
0.7
FRR
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.2
0.4
FAR
0.6
0.8
1
Figure 5.7: DET curves showing the FRR vs FAR for PCA, Fisher and fusion of PCA, and Fisher
features, respectively. The relatively large error rates are a consequence of the use of a particularly
difficult database, and the fact that the choice of difference metric is not a high performance face
recognition algorithm. The purpose of this comparison is to show the effect of features fusion and
classification on the EER for a specific system.
168
5.3.2 Degraded features
In this section,
is computed for degraded features and information loss measured with
respect to the original image. Equation (5.12) represents the blur degradation model used
to generate degraded features where
and
,
is a space invariant Gaussian operator of size
is the original image and
is the resulting degraded image.
(5.12)
Using the degradation model described by equation (5.12), two different sets of images
(
and
) are generated. Each set of images is composed of
per individual for a total of
.
is a set of images obtained as a result of
blurring the entire face region. An example of images in
and
images
is obtained by degrading half of each individuals face
using different Gaussian operators while
Using
people with
, new PCA, FLD and ICA features
and
are seen in Figure 5.8.
are extracted using the original
(non-degraded) principal component vectors. From the degraded features,
is
computed for the degraded individual and population distributions using equation (5.11).
This measure represents the amount of information lost as a function of the degradation
level. Figure 5.9 shows
taken from
and
computed as function of the blur level for different images
. The x-axis represents
different levels (in increasing order) of
Gaussian blur. As seen in Figure 5.9, the relative information loss in an image increases
with the amount of system degradation. Interestingly,
tends to reach a steady state
after some level of degradation (Figure 5.9b). This suggests that some features are
169
unaffected by the degradation process and represent a lower bound of information
measure of an individual distribution. PCA features extracted using the most dominant
eigenvalues of the system tend to be more robust against blur since they preserve
valuable information at a large degradation level. It is important to note that in Figure
5.8b, ICA seems to be me robust to blur which is not a true representation of the actual
information loss due to degradation at higher blur level. The reason behind this is the fact
that ICA isn’t a good scheme to use on Gaussian data such as the ones developed in this
work since ICA maximizes non-Gaussianity and prefers fine features. Hence, this
explains the unusual robustness of ICA against degradation at higher blur levels (Figure
5.9b).
(a)
(b)
(c)
Figure 5.8: Degraded image obtained by applying a Gaussian blur to (b) a section of the original
image
and to (c) the entire image
.
170
0.03
PCA
FISHER
ICA
0.025
0.02
∆BI
0.015
0.01
0.005
0
−0.005
0
2
4
6
8
10
Blur level
(a)
0.08
PCA
FISHER
ICA
0.07
0.06
∆BI
0.05
0.04
0.03
0.02
0.01
0
0
2
4
6
(b)
8
10
Blur level
Figure 5.9: ∆BI as a function of an increasing blur level for images taken from (a)
and (b)
171
5.4 Biometric Iris Features Information
In this section we develop an algorithm to calculate biometric information based on a set
of features, using the iris recognition system described in section (2.8) and the relative
entropy measure developed in the section (5.2). The developed algorithm in this section
is divided in the following steps: i. distribution modeling of iris biometric features, ii.
relative entropy calculations, iii. ICA iris feature extraction and biometric information
calculation. The iris regions used in the entropy calculation are obtained using the Masek
and the enhanced iris segmentation techniques described in section (4.2).
In a generic biometric system,
feature vector
biometric features are measured, to create a biometric
for each iris. For a person iris
feature samples, while we have
in a subset of irises
, we have
samples for a set of irises. Defining
as an
instance of random variable , we calculate the population feature mean
(5.13)
where the feature mean of an iris ,
feature covariance matrix
, is defined analogously, replacing
by . The iris
can be written as follows:
(5.14)
172
The individuals iris feature covariance,
, is again defined analogously. One important
general difficulty with direct information theoretic measures is that of data availability.
Based on the Gaussian model, we can write:
(5.15)
(5.16)
.
From which we can calculate
(5.17)
where
and
.
This expression calculates the relative entropy in bits for Gaussian distributions
and
. This expression corresponds to most of the desired requirements for a biometric
feature information measure introduced in the section (5.2.2). The regularization scheme
used for this method is the same as the one described in sections (5.2.3 and 5.2.4).
173
In order to measure the biometric feature information content of a biometric iris feature
representation, we calculate the average biometric feature information for each iris in
group of irises. This is a measure of the system biometric information (
be calculated by the average
) which can
over the set of irises .
(5.18)
Information in a feature representation of an iris is calculated using our described method
for different irises. In order to test our algorithm, it is necessary to have multiple images
of the same iris. For this reason, we used the CASIA database which includes
images taken using
subjects where
or
iris
images were presented per class (i.e.
subject’s eye). The iris images were processed using the iris recognition and eyelash
detection schemes described in section (4.2) from which we calculate the PCA (eigeniris) features using the algorithm of (Grother (2000)) and the ICA iris features
components using the algorithm described in (Xiang et al. (2004. For PCA and ICA
feature decompositions,
separate feature vectors were computed and used for
subsequent analysis. Figure 5.10 and Figure 5.11 illustrate the amount of biometric
information calculated per PCA and ICA iris feature, respectively. Using the biometric
information calculation procedure described in section (5.2), the sum of the biometric
information over the PCA iris features extracted from the set of irises taken from the
CASIA database give approximately
hand, we computed
bits using Masek’s algorithm. On the other
bits of information for the PCA iris features when using the
174
developed enhanced iris segmentation algorithm. The difference in bits can be explained
by the fact that Masek’s technique doesn’t completely eliminate eyelash noise over the
iris region and these pixels are falsely used as iris information. This will increase the
biometric information since the algorithm takes eyelash as iris texture which then,
compared to the rest of the irises in the set, makes it more unique or different.
In addition, Gabor features were used for the biometric information calculation. These
features were decomposed using the ICA technique described in (Xiang et al. (2004)) in
order to have
independent feature vectors. ICA has the advantage that it does not only
decorrelates the signals but also reduces higher-order statistical dependencies in order to
make the signals as statistically independent as possible. Since ICA maximizes nongaussianity, it fits well to the assumptions that iris features tend to be non-gaussian. For
the ICA features, an average of
segmentation algorithm and
bits was computed for
using the Masek’s
bits using the enhanced segmentation algorithm. As
noticed, the amount of information per iris feature is very close for PCA and ICA
features. ICA features tend to contain more information since they fit the iris feature data
model better.
175
D(p2||q2): "Masek Eigen features"
2.5
BI (bits)
2
1.5
1
0.5
0
0
50
100
150
200
250
300
350
250
300
350
D(p3||q3): "Enhanced Eigen features
2.5
BI (bits)
2
1.5
1
0.5
0
0
50
100
150
200
Iris features
Figure 5.10: Biometric Eigen iris feature information computed for 327 iris features. The y-axis
represents the biometric information for each feature (in bits) and the x-axis is the feature number.
The top graph is calculated using the Masek’s algorithm while the bottom graph is generated using
the enhanced technique. The standard deviation is also plotted at the bottom of each graph.
D(p||q) : "MASEK ICA GABOR PHASE FEATURES (Constant Angle, varying radius)"
BI (bits)
1.5
1
0.5
0
0
50
100
150
200
250
300
350
D(p||q) : "ENHANCED ICA GABOR PHASE FEATURES (Constant Angle, varying radius)"
BI (bits)
1.5
1
0.5
0
0
50
100
150
200
250
300
350
Iris features
Figure 5.11: Biometric ICA iris feature information computed for 327 iris features where the
features are extracted from the iris region at a constant angle/varying radius. The y-axis represents
the biometric information for each feature (in bits) and the x-axis is the feature number. The top
graph is calculated using the Masek’s algorithm while the bottom graph is generated using the
enhanced technique. The standard deviation is also plotted at the bottom of each graph.
176
D(p||q) : "MASEK ICA GABOR PHASE FEATURES (Constant Radius, varying angle)"
BI (bits)
2
1.5
1
0.5
0
0
50
100
150
200
250
300
350
D(p||q) : "ENHANCED ICA GABOR PHASE FEATURES (Constant Radius, varying angle)"
BI (bits)
1.5
1
0.5
0
0
50
100
150
200
250
300
350
Iris features
Figure 5.12: Biometric ICA iris feature information computed for 327 iris features where the
features are extracted from the iris region at a varying angle/constant radius. The y-axis represents
the biometric information for each feature (in bits) and the x-axis is the feature number. The top
graph is calculated using the Masek’s algorithm while the bottom graph is generated using the
enhanced technique. The standard deviation is also plotted at the bottom of each graph.
5.5 Discussion
This work describes an approach to measure biometric feature information and the
changes in biometric sample quality resulting from image degradations. A definition of
biometric feature information is introduced and an algorithm to measure it proposed,
based on a set of population and individual biometric features, as measured by a
biometric algorithm under test. Biometric information is defined in terms of the reduction
in uncertainty of the identity of a person resulting from a set of biometric feature
measurements. Based on this definition, we show that this concept matches the
information theoretic concept of relative entropy
, where
is the probability
distribution of the person’s features, and is the distribution of features of the population.
177
Examples of its application were shown for two different face recognition algorithms
based on PCA (Eigenface) and FLD feature decompositions. Subsequently, we
introduced a measure of information loss as a function of image degradation. It is shown
that the normalized mean square distance measure (
), based on the relative entropy,
increases with the blur level but reaches a steady state after some amount of degradation
which suggests that some features are unaffected by this degradation process.
Clearly, the framework developed in this work depends on accurate estimates of the
population distributions . Developing a good estimate of the “world model” is known to
be a hard problem; in this work, we use the typical approach of assuming our database is
an adequate representation of the population. The result of biometric feature information
calculations (approximately
bits per face) is compatible with previous analyses of face
recognition accuracy. From the FRVT results, we extrapolate the gallery size for an
identification rate of
(Philips et al. (2003, 2007)). This is taken to be a rough model of
the population for which the algorithm can reduce the identity uncertainty to
. For the
top three face recognition algorithms described in Philips et al. (2003, 2007), the gallery
sizes were
,
, and
, corresponding to
,
, and
bits. These values correspond to over half the ones we calculated in this work which
seems reasonable, since the FRVT database appears to be significantly more difficult
than the one used here (Craw et al. (1999)), and current face recognition algorithms are
not yet considered to be close to optimal. They seem to use approximately 1/2 to 2/3 of
the available feature information.
178
In addition, this work describes an approach to measure biometric feature information for
iris images segmented using the Masek and enhanced algorithms. It was shown that
Masek’s segmentation overestimates the eyelash noise in an iris image which tends to
falsify the iris data by classifying eyelash noise as valid iris pixels (Xie (2007)). This will
have an effect on the biometric information calculation since eyelash features add
information to the iris image by making it more unique compared to the rest of the
images.
Examples of its application were shown for two different iris recognition
algorithms based on PCA and ICA feature decompositions. The result of biometric
feature information calculations (approximately
iris features using the Masek’s code while obtaining
bits for PCA and
bits for ICA
bits for PCA and
bits for
ICA iris features using the enhanced segmentation algorithm) is compatible with previous
analyses of iris recognition accuracy. Daugman (2002) states that the combinatorial
complexity of the phase information of the iris across different persons spans about
degrees of freedom. Expressing this variation as discrimination entropy (Cover and
Thomas (1991)) and using typical iris and pupil diameters of
and
respectively, the observed amount of statistical variability among different iris patterns
corresponds to an information density of about
on the iris. From this, we
can easily calculate that an average iris should have an average iris area of
which ideally would give
bits per
iris. Using our developed biometric information calculation scheme, we found that,
depending on the feature decomposition and iris segmentation technique, we obtain on
average
bits of information for the iris decomposition features using the Masek’s
segmentation technique and
bits of iris information when applying the enhanced
179
technique prior to feature extraction. These results obtained using our algorithm supports
Daugman’s theory since our iris images have an average iris diameter ranging
approximately from
to
which explains the difference in bits between
. For instance, a
our method and Daugman’s who assumes the iris diameter to be
positive difference of
in the iris diameter size results in an increase of
bits
using Daugman’s discrimination entropy measure. Hence, this explains the difference in
numbers between our current results and Daugman’s assumptions.
As an exploration of the implications of this work, an analogy can be made between a
biometric system and a traditional communication system in terms of information
capacity (Cover & Thomas (1991)). The signal source transmits one symbol from an
alphabet; this corresponds to one person from a population to be identified. The symbol is
encoded and sent across a channel and is subject to channel noise; similarly, biometric
features from a person are measured, and are subject to variability due to noise in the
measurement system and to inherent feature variability. Thus the biometric feature
measurement system corresponds to the communication channel. The communications
system receiver detects a signal and must decide which symbol was sent, corresponding
to the role of the biometrics identification process. In this context,
differential information of a single signal, and the average
is the
, weighted by the
probability of each signal , is the channel capacity. Based on this analogy, we can say
that biometric feature information is the channel capacity of a biometric measurement
system.
180
In a general biometric system, the following issues associated with biometric features
must be considered:
• Feature distributions vary. Features, such as minutiae ridge angles may be uniformly
distributed over
, while other features may be better modeled as Gaussian. In this
work, all features are modeled as Gaussian. This is valid model for most PCA and FLD
features, but is not valid for any ICA features (since ICA is designed to maximize nonGaussianity). On the other hand, a Gaussian model may be considered to estimate an
upper bound for the entropy.
• Raw sample images need to be processed by alignment and scaling before features can
be measured. Any variability in registration will dramatically increase the variability in
measured features and decrease the biometric feature information measure.
• Feature dimensionality may not be constant. For example, the number of available
minutiae points varies. The method presented in this work does not address this issue,
since the dimensions of
and
must be the same. Generalized Entropy measures
exist which may allow an extension of this approach to non-constant dimensional
features. It is interesting to note that the biometric entropy is larger for some faces. Figure
5.5 shows a range of biometric information (from 32 to 47 bits) for different individuals,
which may help explain why some people are potentially easier to recognize than others.
This is perhaps some evidence for the “biometrics zoo” hypothesis (Doddington (1998))
which classifies users, in the context of a speaker recognition system, into different
181
groups based on their tendency to affect the FAR and FRR of a biometric system. In
general, it states that some individuals possess more reliable/recognizable features (i.e.
subjects with features that are well separated from others in the database) compared to
other users who are intrinsically difficult to recognize and who can degrade the
performance of a biometric system by increasing the FRR or FAR. In order to explore
this effect, we plot the biometric feature information as a function of average feature
variance for each person (Figure 5.13). A significant correlation (
) is calculated
for those features indicating that they contain less variability in those subjects with higher
biometric feature information. The
measure may help address many questions in
biometrics technology, such as the following:
• Uniqueness of biometric features: A common question is “are fingerprints really
unique?”. While Pankanti et al. (2002) have recently provided a sophisticated analysis of
this problem based on biometric feature distributions directly, a general approach based
on information content would help address this question for other biometric modalities.
• Inherent limits to biometric template size requirements. A maximum compression of
biometric features will be limited to the biometric feature information. This theoretical
lower limit may be of use for ID card applications with limited data density.
• Feasibility of biometric encryption: Proposed biometric encryption systems use
biometric data to generate keys (Uludag et al. (2004)), and thus the availability of
182
biometric feature information limits the security of cryptographic key generation (Ballard
et al. (2007), Dodis et al. (2004)).
• Performance limits of biometric matchers: While some algorithms outperform others, it
clear that there are ultimate limits to error rates, based on the information available in the
biometric features. In this application, the biometric feature information is related to the
discrimination entropy (Daugman (2003)).
• Biometric fusion: Systems which combine biometric features are well understood to
offer increased performance (Ross et al. (2003)). It may be possible to use the measure of
biometric feature information to quantify whether a given combination of features offers
any advantage, or whether the fused features are largely redundant. The example of
fusion of FLD and PCA (200 features) given here clearly falls into the latter category
since it does not necessarily offer double the amount of information.
• Novel biometric features: Many novel biometric features have been suggested, but it is
often unclear whether a given feature offers much in the way of identifiable information.
Biometric information measurement may offer a way to validate the potential of such
features.
• Privacy protection: It would be useful to quantify the threat to privacy posed by the
release of biometric feature information, and also to be able to quantify the value of
183
technologies to preserve privacy, such as algorithms to de-identify face images (Newton
et al. (2005), Zhu et al. (2007)).
50
Average BI (bits)
45
40
35
30
10
15
20
25
Average feature variance
30
Figure 5.13: Average
as a function of the mean feature variance (arbitrary units) (x-axis) for
16 different persons. The mean feature variance is computed by summing all the diagonal
components of
matrix for each person. The correlation coefficient is
, which is significant at
.
5.6 Summary
This work describes an approach to measure biometric feature information and the
changes in biometric sample quality resulting from image degradations. A definition of
biometric feature information is introduced and an algorithm to measure it proposed,
based on a set of population and individual biometric features, as measured by a
biometric algorithm under test. Biometric information is defined in terms of the reduction
in uncertainty of the identity of a person resulting from a set of biometric feature
measurements and is measured using the relative entropy. Examples of its application
184
were shown for two different face recognition algorithms based on PCA and FLD feature
decompositions. Subsequently, we introduced a measure of information loss as a function
of image degradation. It is shown
, increases with the blur level but reaches a steady
state after some amount of degradation which suggests that some features are unaffected
by this degradation process. The result of biometric feature information calculations
showed approximately an average of
bits of BI per face. In addition, this work
describes an approach to measure biometric feature information for iris images
segmented using the Masek and enhanced algorithms. Examples of its application were
shown for two different iris recognition algorithms based on PCA and ICA feature
decompositions. Biometric feature information calculations of approximately
PCA,
bits for ICA iris features using the Masek’s code and
bits for
bits for PCA,
bits for ICA iris features using the enhanced segmentation algorithm were obtained.
185
6 Chapter 6
Conclusion and Future Work
6.1 Discussion
Biometrics is a rising field of information technology that uses a subject’s identifying
biological traits in the goal to identify them. By calculating the physiological and
behavioural characteristics using the individual’s biological samples, it has been shown
that information identifying each individual can be extracted in order to verify the
identity of that individual in a population. The reason why biometrics is an attractive field
is due to the fact that biometric traits cannot be forgotten or lost, they are difficult to
copy, share and distribute and they require the person to be present at the time of
authentication. Iris recognition technology is ranked as one of the best performing
approaches to biometric recognition as compared with other biometric technologies such
as fingerprint, hand geometry, speaker recognition and many more (Bolle et al. (1998)).
Even though iris recognition is a relatively high performing technology, there are still
some issues to be solved in order to further improve its accuracy and robustness.
In this thesis, new algorithms are developed to deal with low quality images for various
applications in biometrics. A collection of techniques that can be applied for face and iris
recognition in a non-cooperative environment is also presented. A novel scheme and new
186
algorithms are proposed for automatically detecting and recognizing human subjects via
their face and eye traits at a long distance without cooperation. Face and iris recognition
algorithms are widely used except that in most of them, images are taken from a
cooperative individual under a controlled environment in order to provide satisfactory
results (Ma et al. (2004)). Therefore, these techniques have limited capability of
identifying non-cooperative subjects for applications such as surveillance, where the
observed individuals are non-cooperating and/or non-habituated to the technology.
Images taken from a non-cooperating individual tend to include more distortions and
noise (i.e. low quality) hence the need of advanced algorithms to process low quality face
and iris biometric images. The thesis objectives are to (i) use infrared illumination to
improve eye and face tracking in low quality video images, (ii) develop improved
identification of iris and eyelash features in low quality images, (iii) measure biometric
sample quality in terms of biometric information for face features, (iv) measure
information content in biometric iris and face features.
For the face tracking and eye detection algorithm, the first step is to locate the face in the
image, detect the eyes and track the face based on pupil detection using IR illumination.
The new algorithm for face and eye tracking extracts and tracks face and eye positions
from surveillance type images with IR strobe taken under poor illumination. In the case
where many reflections occur, the algorithm will find all possible eye locations and
presents the best solution using multi-stage classification techniques. In order to reduce
the set of possible eye region candidates, shape and geometric constraints are imposed in
the classification process. A Kalman tracker is used to approximate eye location in bright
187
pupil images. If the image contains weak reflections, a local adaptive thresholding
technique is used to extract the eye contour in order to estimate the pupil location. This
improves the performance and accuracy of the system when dealing with faces at
different orientation and with eye closure. Tested on
system achieves a
IR images, the proposed
% detection rate.
The second proposed algorithm is used on low quality eye/iris images where a major
section of the iris region is occluded. The developed technique is intended to increase the
biometric system’s performance when dealing with low quality iris images, particularly
when eyelash occlusions become severe. The developed iris recognition scheme is
compared to the approach of Masek (1999) in order to observe the effect of the amount of
valid iris image processed after noise detection on the system’s performance assuming
that more accurate iris segmentation (i.e. enhanced noise detection) results in better
identification performance. The enhanced iris segmentation scheme is presented in
chapter 4. The proposed method shows promising results for eyelash noise detection,
accurate iris boundary extraction and ideal iris segmentation. This algorithm locates the
iris region using logarithmic image enhancement and the Hough transform techniques,
locates the iris boundary, extracts the exact iris contour, detects eyelash based on the
local image statistics and block intensity and finally, proposes an ideal iris model for
accurate iris recognition. The developed method overcomes the limitations encountered
in other iris segmentation and eyelash detection techniques such that our method detects
accurately separable and multiple eyelashes, extracts the exact iris contour and is
illumination invariant. The developed method does not overestimate the eyelash noise in
188
the eye image, maximizing iris information. Using our technique, less iris noise and
fewer eyelash occlusions are found in the normalized iris region compared to the
approach of Masek (1999). The evaluation showed that this algorithm improves the error
rates in DET curves except in very low FMR range. In addition, the decidability metric is
calculated for both methods in order to measure the separation between the intra-class
and inter-class Hamming distance distributions. It is seen that a higher decidability
measure is obtained for the enhanced algorithm which implies that the iris recognition
system implementing the enhanced eyelash detection scheme tend to have a better intra
and inter-class distribution separation. This will result in lower false accept rate and false
reject rate since it is easier to select the decision threshold that provides a good genuineimpostor distribution separation. The match score using the enhanced technique is better
within the entire comparison range since it uses the most of the iris region in the pattern
matching process. Masek’s code neglects some valid iris pixels since it overestimates
eyelash occlusion and noise pixels. This will result in a loss of valid iris pixels which will
probably decrease the system’s performance.
Finally, this work describes an approach to measure biometric feature information and
changes in biometric sample quality resulting from image degradations. A definition of
biometric feature information is introduced and an algorithm to measure it proposed,
based on a set of population and individual biometric features, as measured by a
biometric algorithm under test. Biometric information is defined in terms of the reduction
in uncertainty of the identity of a person resulting from a set of biometric feature
measurements. Based on this definition, we show that this concept matches the
189
information theoretic concept of relative entropy
, where
is the probability
distribution of the person’s features, and is the distribution of features of the population.
Examples of its application were shown for two different face recognition algorithms
based on PCA (Eigenface) and FLD feature decompositions. Subsequently, we
introduced a measure of information loss as a function of image degradation. We selected
blur as a degradation process to test our algorithm. It is important to note that other type
of degradations and distortions can also occur in a covert biometric system such as
motion blur, head rotation, face occlusion, eyewear, non-uniform lighting, out-of-focus
images and many more. It is shown that the normalized mean square distance measure
(
), based on the relative entropy, increases with the blur level but reaches a steady
state after some amount of degradation which suggests that some features are unaffected
by this degradation process.
Clearly, the framework developed in this work depends on accurate estimates of the
population distributions . Developing a good estimate of the “world model” is known to
be a hard problem; in this work, we use the typical approach of assuming our database is
an adequate representation of the population. The result of biometric feature information
calculations (approximately
bits per face) is compatible with previous analyses of face
recognition accuracy. From the FRVT results, we extrapolate the gallery size for an
identification rate of
(Philips et al. (2003, 2007)). This is taken to be a rough model of
the population for which the algorithm can reduce the identity uncertainty to
top three algorithms, the gallery sizes were
corresponding to
,
, and
,
, and
. For the
,
bits. This value is over half that calculated here, and
is reasonable, since the FRVT database appears to be significantly more difficult than the
190
one used here (Craw et al. (1999)), and current face recognition algorithms are not yet
considered to be close to optimal. They seem to use approximately 1/2 to 2/3 of the
available feature information.
Moreover, this work describes an approach to measure biometric feature information for
iris images segmented using the Masek and the enhanced algorithms (Masek (1999)). It
was shown that Masek’s segmentation overestimates the eyelash noise in an iris image
which tends to falsify the iris data by classifying eyelash noise as valid iris pixels. This
will have an effect on the biometric information calculation since eyelash features add
information to the iris image by making it more identifying compared to the rest of the
images.
Examples of its application were shown for two different iris recognition
algorithms based on PCA and ICA feature decompositions. The result of biometric
feature information calculations (approximately
iris features using the Masek’s code while obtaining
bits for PCA and
bits for ICA
bits for PCA and
bits for
ICA iris features using the enhanced segmentation algorithm) is compatible with previous
analyses of iris recognition accuracy. Daugman (2002) states that the combinatorial
complexity of the phase information of the iris across different persons spans about 249
degrees of freedom. Expressing this variation as discrimination entropy (Cover and
Thomas (1991)) and using typical iris and pupil diameters of
and
respectively, the observed amount of statistical variability among different iris patterns
corresponds to an information density of about
on the iris. From this, we
can easily calculate that an average iris should have an average iris area of
which ideally would give
per
191
iris. Using our developed biometric information calculation scheme, we found that,
depending on the feature decomposition and iris segmentation technique, we obtain on
average
bits of information for the iris decomposition features using the Masek’s
segmentation technique and
bits of iris information when applying the enhanced
technique prior to feature extraction. These results obtained using our algorithm supports
the value from Daugman (2002) since our iris images have an average iris diameter
ranging approximately from
to
which explains the difference in bits
between our method and Daugman’s who assumes the iris diameter to be
instance, a positive difference of
approximately
. For
in the iris diameter size results in an increase of
using Daugman’s discrimination entropy measure.
6.2 Future Work
Based on the results of this thesis, there are several promising avenues of research that
are likely to yield useful results. For example, sophisticated methods can be developed to
enhance circular detection by accurately locating the iris and pupil boundaries in the least
amount of time, minimizing the identification error rates. It will be also interesting to
study the biometric information theory in a way to analyze the information content with
progressive image segmentation. It can give further insight into how the segmentation
affects the biometric entropy for recognition. Does more iris mean additional
discriminative information (in bits)?
Also, it is suggested to compare additional iris recognition algorithms in order to obtain
new valuable information that will help improving the recognition accuracy of a typical
192
iris recognition system, according to different recognition strategies. Moreover,
additional types of noise should be subject of further work, since this work dealt with low
quality eye and face images but not necessarily with off-angle iris images which will
obviously introduce new challenges to the recognition process.
Another idea will be to develop an accurate biometric identifier in a non-cooperative
system. The algorithms and methods developed in this thesis can be used to understand
the requirements to be able to physically implement a non-cooperative prototype system
and to improve real-time algorithms for face and eye detection, face tracking and facial
and iris recognition. Furthermore, since it is difficult to capture high-quality iris images
with minimal user cooperation, the proposed methods in this thesis can be used to design
a non-cooperative access control system for applications requiring high security. This
system will identify individuals at a distance in an uncontrolled environment. When the
subject is located within a capture zone, face and eye images will be captured using
surveillance cameras. The subject can be located and tracked using advanced face
detection and tracking techniques. In order to obtain high quality eye/iris images, a
zoomed near infrared iris camera can be used to collect images of the targeted subject.
The iris camera lens will automatically adjust its focal distance depending on the subject
distance from the camera platform. From the acquired face and eye video frames, the
proposed algorithms can be used to detect the subject’s face, locate the eyes, reduce iris
noise, segment the iris, generate a template and then identify the subject through typical
pattern matching algorithms.
193
Another possible line of research is to use the entropy theory to analyze the information
content with the progressive image segmentation proposed by Xie (2007). This will
facilitates the understanding of the effects of iris noise segmentation on the biometric
entropy for recognition.
Overall, all the developed techniques in this thesis could be used to improve an iris
recognition system performance with low quality iris images. Such applications include
security surveillance; face and eye tracking in poor lighting conditions, unsupervised
capture of iris image under poor lighting conditions, iris capture for subjects on the move
and many more. The results obtained in this thesis help to clarify future research
directions in the effort to improve non-cooperative iris and face detection/recognition
performance for low quality face and eye images.
194
Bibliography
Adler, A., Youmaran, R., Loyka, S. (2009).Towards a Measure of Biometric Feature
Information. Pattern Analysis and Applications, 12:261−270.
Adler, A. (2005). Vulnerabilities in biometric encryption systems. Audio- and Videobased Biometric Person Auth. NY, USA, 1100-1109.
Adler, A., Youmaran, R., Loyka, S. (2005). Information content of biometric
Features. Biometrics Consortium Conference Washington, DC, USA.
Adler, A., Youmaran, R., Loyka S. (2005). Information content of biometric features.
Biometrics Consortium Conference, Sep. 19-21, Washington, DC, USA.
Alter O, Brown PO, Botstein D.(2000). Singular value decomposition for genomewide
expression data processing and modeling. Proc Natl. Acad. Sci., 97:10101–10106.
Asfaw, Y., Chen, B., Adler, A. (2002). Face detection using red-eye effect. Department
of Electrical engineering, Montréal, Canada, 1211-1214.
Ballard, D.H. (1981). Generalizing the Hough transform to detect arbitrary shape, Pattern
Recognition, 13: 111-122.
Ballard, L.,Kamara, S.,Monrose, F., Reiter, M. (2007). On the Requirements of Biometric
Key Generators, Technical Report TR-JHU-SPAR-BKMR-090707, John Hopkins
University.
Boles, W., Boashash, B. (1998). A human identification technique using images of the
iris and wavelet transform. IEEE Transactions on Signal Processing, 46 (4): 1185–1188.
Bolle, R., Pankanti, S. (1998). Biometrics, Personal Identification in Networked Society:
Personal Identification in Networked Society. Kluwer Academic Publishers, Norwell,
MA, USA.
Bovik, A.C. (2005). Handbook of Image and Video Processing. Elsevier Science &
Technology.
Bradski, G.R. (1998). Computer vision face tracking for use in a perceptual user
interface. Intel Technology Journal, 2nd Quarter.
Brady, N., Field, D. (2000). Local contrast in natural images: normalization and coding
efficiency. Perception, 29 (9): 1041-1055.
195
Belhumeur, P.N., Hespanha, J.P., Kriegman, D.J. (1997). Eigenfaces vs. Fisherfaces:
recognition using class specific linear projection, Pattern Analysis and Machine
Intelligence, IEEE Transactions on, 19:711 - 720.
Bovik, AL. (2000) “Handbook of Image and Video Processing”, Academic Press Texas.
Bradski, G.R., (1998). Computer Vision Face Tracking For Use in a Perceptual User
Interface. Workshop on Applications of Computer Vision.
Chau, M., Betke, M. (2005). Real time eye tracking and blink detection with USB
cameras. Technical Report, Boston University Computer Science.
Chinese Academy of Sciences Institute of Automation. (2004).Casia iris image database.
http://www.sinobiometrics.com
Chirillo, J., Blaul, S. (2003). Implementing biometric Security. Wiley Publishing, Inc.
Comaniciu, D., Ramesh, V., Meer, P. (2000). Real-time tracking of non-rigid objects
using mean shift. IEEE Conf. on Computer Vision and Pattern Recognition, Hilton Head
Island, South Carolina.
Comaniciu, D., Ramesh, V., Meer, P. (2003). Kernel-based object tracking. IEEE Trans.
Pattern Analysis and Machine Intelligence. 25(5):564–577.
Craw, I., Costen, N.P., Kato, T., Akamatsu, S. (1999). How should we represent faces for
automatic recognition, Pattern Analysis and Machine Intelligence, IEEE Transactions on,
21:725–736.
Conover, W.J. (1980). Practical Nonparametric Statistics, Wiley.
Cover, T.M., Thomas, J.A. (1991). Elements of Information Theory New York:Wiley.
Dansheng M. (2009). Face Tracking Algorithms: A Survey. Panasonic Singapore Lab,
Singapore, Research Report.
Daugman, J. (1993) High confidence visual recognition of persons by a test of statistical
independence. IEEE Trans. Pattern Analysis and Machine Intelligence, 15:1148–1161.
Daugman, J. (2000). Biometric decision landscapes. Technical Report No. TR482,
University of Cambridge Computer Laboratory.
Daugman, J. (2003). The importance of being random: Statistical principles of iris
recognition, Pattern Recognition, 36:279–291.
Daugman, J. (2004). How Iris Recognition Works. IEEE Transactions on Circuits and
Systems for Video Technology, 14 (1): 21-304.
196
Daugman, J. (2004). Recognizing Persons by Their Iris Patterns. Lecture Notes in
Computer Science, 3338: 5-25.
Daugman, J. (2005). Technical Report of Results from 200 Billion Iris Cross
Comparisons. University of Cambridge. UCAM-CL-TR-635.
Daugman, J., Downing, C. (2001). Epigenetic Randomness, Complexity and Singularity
of Human Iris Patterns. Biological Sciences, vol. 268, no. 1477: 1737-1740.
Deng, G., Cahill, L.W. (1993). Multiscale image enhancement using the logarithmic
image processing model. Electronics Letters, 29:803-804.
Deng, G., Cahill, L.W. (1994). Image Enhancement Using the Log-ratio Approach,
Signals Systems and Computers, 1:198-202.
Dessimoz, D., Richiardi, J., Champaud, C., Drygajlo, A. (2005). Multimodal biometrics
for identity documents. Technical report, University of Lausanne.
Doddington, G., Liggett, W., Martin, A., Przybocki, M., Reynolds, D. (1998). Sheep,
Goats, Lambs and Wolves: An Analysis of Individual Differences in Speaker
Recognition Performance, Proc. Int. Conf. Auditory-Visual Speech Processing, Sidney,
Australia.
Dodis, Y., Reyzin, L., Smith, A. (2004). Fuzzy Extractors and Cryptography, or How to
Use Your Fingerprints, Proc. Eurocrypt’04. http://eprint.iacr.org/2003/235/
Dorairaj, V., Schmid, N., Fahmy, G. (2005). Performance evaluation of non-ideal iris
based recognition system implementing global ICA encoding. In Proceedings of the
IEEE International Conference on Image Processing (ICIP 2005, 285–288.
Draper, B.A., Baek, K., Bartlett, M.S., Beveridge, J.R. (2003). Recognizing
faces with PCA and ICA, Computer Vision and Image Understanding, 91:115-137.
Du, Y., Bonney, B., Ives, R., Etter, D., Schultz, R. (2005). Analysis of partial iris
recognition using a 1D approach. In Proceedings of the IEEE International Conference on
Acoustics, Speech and Signal Processing (ICASSP’05), 961–964.
Duc B., Fischer, S., Bigun, J. (1999). Face Authentication with Gabor Information on
Deformable Graphs. IEEE Transactions on Image Processing, 8(4).
Ebisawa, Y., Satoh, S., (1993). Effectiveness of pupil area detection technique using two
light sources and image difference method. Proceedings of the 15th Annual Int. Conf. of
the IEEE Eng. in Medicine and Biology Society, San Diego, CA, 1268–1269.
197
Fancourt, G., Bogoni, L., Hanna, K., Guo, Y., Wildes, R., Takahashi, N., Jain, U. (2005).
Iris recognition at a distance. In Proceedings of the 2005 IAPR Conference on Audio and
Video Based Biometric Person Authentication, 1–13.
Field, D. (1987). Relations between the statistics of natural images and the response
properties of cortical cells. Journal of the Optical Society of America.
Freund, Y., Schapire, R.E. (1996). Experiments with a New Boosting Algorithm. 13th
Conf. on Machine Learning, 148-156.
Fromherz, T., Stucki, P., Bichsel, M. (1997). A Survey of Face Recognition”, MML
Technical Report, No 97.01, Dept. of Computer Science,Zurich.
Gejgu, P., Perka, M. (2003). Face tracking in color video sequences, Proceedings of the
19th spring conference on Computer.
Golfarelli, M., Maio, D., Maltoni, D. (1997). On the Error-Reject Tradeoff in
Biometric Verification Systems, Pattern Analysis and Machine Intelligence, IEEE
Transactions on, 19:786-796.
Gonzalez, R.C., Woods, R.E. (2002). Digital Image Processing, Addison-Wesley
Publishing Company.
Grother, P. (2000). Software Tools for an Eigenface Implementation, National
Institute of Standards and Technology, http://www.nist.gov/humanid/feret/
Haro, A. ,Flickner, M., Irfan, E. (2000). Detecting and Tracking Eyes By Using Their
Physiological Properties, Dynamics, and Appearance. Computer Vision and Pattern
Recognition Conference, June 13-15.
Huang C., Ai, H., Li, Y., Lao, S. (2007). High-Performance Rotation Invariant Multiview
Face Detection. Pattern Analysis and Machine Intelligence, IEEE Transactions on. 29
(4).
Huang, W., Mariani, R. (2000). Face detection and precise eyes location”, in: Proc. Int.
Conf. on Pattern Recognition.
Huang, J., Wang, Y., Tan, T., Cui, J. (2004). A New Iris Segmentation Method for
Recognition, Proceedings of the 17th International Conference on Pattern Recognition,
1051-4651.
Hyvarinen, A. (1999). Fast and Robust Fixed-Point Algorithms for Independent
Component Analysis. IEEE Trans. Neural Networks 10:626–634.
198
Institute of Automation of the Chinese Academy of Sciences. (2003). CASIA V1.0.
CASIA Iris Image Database, Beijing, China.
International Biometric Group (2005). Independent test of iris recognition technology.
http://www.biometricgroup.com/reports.
ISO JTC1 SC37 Biometrics (2007). ISO 29794-1 Biometric Sample Quality, Committee
Draft 1.
Jain, A.K., Ross, A., Prabhakar, S., (2004). An Introduction to Biometric Recognition.
IEEE Transactions on circuits and systems for video technology, 14(1).
Jain, A., Hong, L., Pankanti, S. (2000). Biometric Identification. Communications of the
Association for Computing Machinery, 43 (2): 90-98.
Kong, S.G., Heo, J., Abidi, B.R., Paik, J., Abidi, M.A.. (2005). Recent advances in visual
and infrared face recognition: a review. Computer Vision and Image Understanding,
97,1.
Kong, W.K., Zhang, D. (2001). Accurate Iris Segmentation Based on Novel Reflection
and Eyelash Detection Model, Proceedings of International Symposium on Intelligent
Multimedia, Video and Speech Processing, 263-266.
Kosmerlj, M. (2004). Passport of the Future: Biometrics against Identity Theft? MSc
thesis, Gjovik University College, NISlab.
Kovesi, P. (1997). Image Features From Phase Congruency, Journal of Computer Vision
Research, 1:1-27.
Kryszcuk, K., Drygajlo, A. (2005). Gradient-based image segmentation for face
recognition robust to directional illumination. In Visual communications and image
processing 2005: 12-15, Beijing, China.
Kryszcuk, K., Drygajlo, A. (2006). On combining evidence for reliability estimation in
face verification. In. Proc. Of the EU-SIPCO 2006, Florence.
Kuncheva, L.I., Whitaker, C.J., Shipp, C.A., Duin, R.P.W. (2000). Is independence good
for combining classifiers?. Int’l Conf. on Pattern Recognition (ICPR), 2:168–171.
Lee, T.W. (1999). Nonlinear approaches to independent component analysis, Proc.
American Institute of Physics.
Li, S., Jain A. (Eds.). (2005). Handbook of Face Recognition. Springer.
199
Li, Y., Ai, H., Yamashita, T., Lao,S., Kawade, M. (2008). Tracking in Low Frame Rate
Video: A Cascade Particle Filter with Discriminative Observers of Different Life Spans,
Pattern Analysis and Machine Intelligence, IEEE Transactions on 30 (10).
Ma, L., Tan, T., Zhang, D., Wang, Y. (2004). Local Intensity Variation Analysis for Iris
Recognition. Pattern Recognition, 37:1287-1298.
Mansfield A.J., Wayman, J.L.(2002). Best Practices in Testing and Reporting
Performance of Biometric Devices, Version 2.01. Centre for Mathematics and Scientific
Computing, National Physical Laboratory, Queens Road, Teddington, Middlesex, TW11
OLW.
Masek, L., Kovesi, P. (2003). MATLAB Source Code for a Biometric Identification
System Based on Iris Patterns. School of Computer Science and Software Engineering.
University of Western Australia, Perth, Australia.
Masek, L. (2003). Recognition of Human Iris Patterns for Biometric Identification.
Master's thesis, School of Computer Science and Software Engineering, University of
West Australia, Perth, Australia.
Montgomery, T.M. (2007). Iris image, “Anatomy, Physiology and Pathology of the
Human Eye”. Available:
http://www.tedmontgomery.com/theeye/eyephotos/IrisNevus.html
Morimoto, C.H. ,Koons, D.,Amir, A.,Flickner,M. (1998). Pupil Detection and Tracking
Using Multiple Light Sources. Technical Report RJ-10117, IBM Almaden Research
Center, San Jose, Ca. http://domino.watson.ibm.com/library/cyberdig.nsf/Home.
Newton, E.M., Phillips, P.J. (2007). Meta-Analysis of Third-Party Evaluations of Iris
Recognition. Technical Report NISTIR 7440, National Institute of Standards and
Technology, Gaithersburg, MD.
Newton, E.M., Sweeney, L., Malin, B. (2005). Preserving Privacy by De-Identifying
Face Images, IEEE Trans. Knowledge Data Eng. 17: 232–243.
Niblack, W. (1986). An Introduction to Image Processing”, Prentice-Hall, Englewood
Cliffs, NJ, 115–116.
Nixon, M. (1985). Eye spacing measurement for facial recognition”, in: Proc. of the
Society of Photo-Optical Instrument Engineers.
Ojala, T., Pietikäinen, M., Harwood, D. (1996). A Comparative Study of Texture
Measures with Classification Based on Feature Distributions. Pattern Recognition
19(3):51-59.
Pan, G., Wu, Y., Wu, Z., Liu, W. (2003). 3D face recognition by profile and
200
surface matching. Neural Networks, IEEE, 3:2169–2174.
Pankanti, S., Prabhakar, S., Jain, A.K. (2002). On the Individuality of Fingerprints,
Pattern Analysis and Machine Intelligence, IEEE Transactions on, 24:1010–1025.
Papageorgiou, C., Oren, M., Poggio, T. (1998). A general framework for object detection.
International Conference on Computer Vision.
Pentland, A., Moghaddam, B., Starner, T. (1994). View-based and modular eigenspaces
for face recognition. Proc. IEEE Conf. on Computer Vision and Pattern Recognition,
Seattle, WA.
Phillips, P.J., Martin, A., Wilson, C.L., Przybocki, M. (2000). An Introduction to
Evaluating Biometric Systems," Computer, 33(2):56-63.
Phillips, J. The Face Recognition Grand Challenge (FRGC). http://www.frvt.com/FRGC/.
Phillips, P.J., Grother, P., Micheals, R.J., Blackburn, D.M., Tabassi, E., Bone, J.M.
(2003).
FRVT
2002:
Evaluation
Report,
NIST.
http://www.frvt.org/DLs/FRVT_2002_Evaluation_Report.pdf
Phillips, P.J., Scruggs, T.W., OToole, A.J., Flynn, P.J., Bowyer, K.W., Svhott, C.L.,
Sharpe,
M.
(2007).
FRVT
2006:
Evaluation
Report,
NIST.
http://www.frvt.org/FRVT2006/docs/FRVT2006andICE2006LargeScaleReport.pdf
Poursaberil, A., Araabi1, B. N. (2006). Iris Recognition for Partially Occluded Images:
Methodology and Sensitivity Analysis. EURASIP Journal on Advances in Signal
Processing, 25-38.
Proenca, H.P. (2006). Towards Non-Cooperative Biometric Iris Recognition. University
of Beira Interior. Department of Computer Science.
Ritter, N., Owens, R. (1999). Location of the Pupil-Iris Border in Slit-Lamp Images of
the Cornea. International Conference on Image Analysis and Proceedings, 00: 740-745.
Ritter, N., Cooper, J.R. (2003). Locating the iris: A first step to registration and
identification. In Proc. 9th IASTED International Conference on Signal and Image
Processing, 507-512.
Ross, A., Jain, A. (2003). Information Fusion in Biometrics, Pattern Recognition Letters,
24:2115-2125.
Rukhin, A., Malioutov, I. (2005). Fusion of Biometric Algorithms in the Recognition
Problem. Pattern Recognition Letters, 26 (5): 679-684.
Sandström, M. (2004) Liveness Detection in Fingerprint Recognition Systems. MSc
thesis, Linköpings tekniska högskola.
201
Sarnoff Corporation. Iris On The Move, 2005. http://www.sarnoff.com/
Schuckers, S. A. C. (2002). Spoofing and Anti-Spoofing Measures. Information
Security Technical Report, 7(4), 56–62.
Soutar, C., Roberge, D., Stoianov, A., Gilroy, R., Vijaya, B. (1998). Biometric
Encryption using image processing, Proc. SPIE Int. Soc. Opt. Eng., 3314 178-188.
Sung, E., Chen, X., Yang, J. (2002). Towards non-cooperative iris recognition systems.
In Proceedings of the seventh International Conference on Control, Automation, Robotics
and Vision (ICARV’02), 990–995.
Swaminathan, G., Venkoparao, V., Bedros, S., Rowley, H. (2007). Multiple appearance
models for face tracking in surveillance videos. Advanced Video and Signal Based
Surveillance. IEEE Conference AVSS 2007.
Tabassi, E., Wilson C.R., Watson, C.I. (2004). Fingerprint Image Quality, NISTIR 7151.
Tisse, C.L., Torres, L., Robert, M. (2002). Person Identification Technique using Human
Iris Recognition. International Conference on Vision Interface. Calgary, Canada, 294299.
Tajbakhsh, N., Araabi, B.N., Soltanianzadeh, H. (2008). Feature fusion as a practical
solution toward non-cooperative iris recognition. Information Fusion, 11th International
Conference on. Control & Intell. Process. Center of Excellence, Univ. of Tehran. 1-7.
Turk, M., Pentland, A. (1991). Eigenfaces for recognition, J. Cognitive Neuroscience,
3:71-86.
Turk, M., Pentland, A. (1991). Face Recognition Using Eigenfaces. Proceedings of the
IEEE Computer Society Conference on Computer Vision and Pattern Recognition, June
Maui, Hawaii, 586-591.
Uludag, U., Pankanti, S., Prabhakar, S., Jain, A.K. (2004). Biometric Cryptosystems:
Issues and Challenges, Proc. IEEE 92:948–960.
Viola, P., Jones, M. (2001). Robust Real-time Object Detection. Second International
Workshop on Statistical and Computational Theories of Vision - Modeling, Learning,
Computing and Sampling.
Wang, C., Li, Z. (2008). A New Face Tracking Algorithm Based on Local Binary Pattern
and Skin Color Information. Computer Science and computational Technology,
ISCSCT’08. International Symposium on. 2: 20-22.
202
Wang, Y., Zhu, Y., Tan, T. (2002). Personal Identification Based on Iris Recognition,
Acta Automatica Sinica, 28:1-10.
Wang, P., Ji, Q. (2005). Learning Discriminant Features for Multi-View Face and Eye
Detection. IEEE International Conference on Computer Vision and Pattern Recognition
(CVPR), San Diego, CA.
Wayman, J. (1999). Error Rate Equations for the General Biometric System. IEEE
Transactions on Robotics and Automation Magazine, 6: 35-48.
Wayman, J.S. (2001). A Definition of Biometrics.
http://www.engr.sjsu.edu/biometrics/nbtccw.pdf.
Wayman, J.S. (2004). The cotton ball problem, Biometrics Conference, Washington
DC, USA.
Wayman, J., Jain, A., Maltoni, D., Maio, D. (2005). Biometric Systems: Technology,
Design and Performance Evaluation. Springer-Verlag London Limited.
Wildes, R., Asmuth, J.C., Green, G.L. (1996). A machine Vision System for Iris
Recognition. Machine Vision and Applications, 1(9): 1432-1769.
Wildes, R. (1997) Iris recognition: An emerging biometric technology. Proc. IEEE,
85:1348–1363.
Xiang, C., Fan, X.A., Lee, T.H. (2004). Face recognition using recursive Fisher linear
discriminant, Communications, Circuits and Systems international Conference on.
Xie, L. (2007). Algorithms for Recognition of Low Quality Iris Images. Master’s thesis,
Carleton University, Ottawa.
Adler, A., Youmaran, R., Loyka, S. (2009). Towards a Measure of Biometric Feature
Information. Pattern Analysis & Applications, 12:261−270.
Youmaran, R., Adler, A. (2008). Measuring Information Content in Biometric Features.
Book chapter in Biometrics: Theory, Methods, and Applications N.V. Boulgouris, K.N.
Plataniotis, and E.Micheli-Tzanakou (Eds), In press: Wiley/IEEE 2008.
Youmaran, R., Adler, A. (2008). Using Infrared Illumination to Improve Eye & Face
Tracking in Low Quality Video Images. 24th Queen's Biennial Symposium on
Communications, QBSC 2008, Kingston, Canada.
Youmaran, R., Lipeng, X., Adler, A. (2008). Improved Identification of Iris and Eyelash
Features, 24th Queen's Biennial Symposium on Communications, QBSC 2008, Kingston,
Canada.
203
Youmaran, R., Adler, A. (2006). Measuring Biometric Sample Quality in Terms of
Biometric Information. Biometrics Consortium Conference 2006, Sep. 19-21, Baltimore,
MD, USA.
Youmaran, R., Adler, A. (2006). Towards a measure of biometric information. Can.
Conf. Computer Elec. Eng. (CCECE), Ottawa, Canada, May 7-10.
Yuan, X., Shi, P. (2004). An Iris Segmentation Procedure for Iris Recognition,
Sinobiometrics, 546-553.
Zhao, W., Chellappa, R., Philips, P.J., Rosenfeld, A. (2003). Face Recognition: A
Literature Survey”, ACM Computing Surveys.
Zhang, J., Yan, Y., Lades, M. (1997). Face Recognition: Eigenface, Elastic Matching,
and Neural Nets. Proceedings of the IEEE, 85(9),1423–1435.
Zhu, Y., Tan, T., Wang, Y.W. (2000). Biometric Personal Identification based on Iris
Patterns. 15th International Conference on Pattern Recognition, 2: 801-804.
Zhu, Z., Ji, Q. (2005). Robust real-time eye detection and tracking under variable lighting
conditions and various face orientations. Computer Vision and Image Understanding, 1:
124-154.
Zhu, Y., Dass, S. C., Jain, A. K. (2007). Statistical Models for Assessing the Individuality
of Fingerprint, IEEE Transactions on Information Forensics and Security, 2, No. 3, 391401.
204
Download PDF
Similar pages