Upload
ramiz-raja
View
99
Download
1
Tags:
Embed Size (px)
Citation preview
“Iris Recognition System”
Submitted By:
Samiullah (2007-CSE-0609)
RamizRaja (2007-CSE-0640)
Atta-Ur-Rehman (2007-CSE-0683)
Supervisor:
Engr. Sumayyea Salahuddin
i
Department of Computer System Engineering, KPK University of Engineering
and Technology Peshawar
“Iris Recognition System”
A Project submitted to the Department of Computer System Engineering, KPK
University of Engineering & Technology Peshawar in partial fulfillment of the
requirements for the degrees of
“B.Sc. Computer Systems Engineering”
Approved by:
Project Supervisor _________________________
External Examiner. ________________________
Chairman, Dept. of CSE. _________________________
ii
DEPARTMENT OF COMPUTER SYSTEMS ENGINEERING,
KPK UNIVERSITY OF ENGINEERING & TECHNOLOGY,
PESHAWAR
iii
AUTHOR'S DECLARATION
It is hereby notified that all the intellectual component of this report is original
and has been created by the group members themselves. All the aspects of
copyrighting are acknowledged in the making of the report.
Name. Signature.
Samiullah. ________________.
Ramiz Raja. ________________.
Atta-Ur-Rehman. ________________.
iv
Abstract
Iris recognition systems capture an image from an individual's eye. The iris in the
image is then segmented and normalized for feature extraction process. The performance of
iris recognition systems highly depends on segmentation and normalization. For
segmentation purpose we use Circular Hough Transform to localize the circular iris region
and pupil region, occluding eyelids and eye lashes. The segmented iris region is then
normalized into rectangular block. For this purpose we use Rubber Sheet Model having
constant dimension. Finally we use 1D log-Gabor Filter to extract features’ bit-wise pattern.
Then Hamming Distance (HD) and Support Vector Machine (SVM) are employed for the
classification of iris templates.
v
Acknowledgements
First of all we are thankful to ALMIGHTY ALLAH, who enabled us to
complete this project.
We acknowledge with gratitude the guidance and advice provided by
Engr. Sumayyea Salahuddin at every crucial state of our project regarding the
study of the project and all guidance and helping material are provided by her.
We pay special regards to her for pushing her efforts with us for the completion
of our project.
We would like to express our profound thanks to Dr. K. M. YAHYA,
Chairman Department of Computer Systems Engineering for inspiring guidance
and help.
vi
Dedication
Success always Solicitous
For Two Things
Exertion and Fortune
If we are successful then our exertion is the efforts of our parents, which
they made to fulfill our wishes and our fortunes is due to their prayers, hence
we dedicate our success to our BELOVED PARENTS, TEACHERS and
FRIENDS.
vii
Table of Contents
Author’s Declaration iii
Abstract iv
Acknowledgements v
Dedication vi
Table of Contents vii
List of Figures ix
List of Tables x
Chapter 1 Introduction 11
1.1 What is Biometric?......................................................................................................................11
1.1.1 How a Biometric System works?.........................................................................................11
1.2 BIOMETRICS AUTHENTICATION........................................................................................11
1.3 Applications of Biometrics..........................................................................................................12
1.3.1 Forensic(Law Enforcement Department).............................................................................12
1.3.2 Government..........................................................................................................................12
1.3.3 Commercial..........................................................................................................................13
1.4 Biometric Characteristics............................................................................................................13
1.5 Why Iris Recognition?.................................................................................................................13
1.6 IRIS RECOGNITION.................................................................................................................14
1.6.1 HISTORY OF IRIS RECOGNITION..................................................................................14
1.6.2 THE HUMAN IRIS..............................................................................................................14
1.7 Iris Recognition System Methodology........................................................................................16
1.7.1 Image Acquisition................................................................................................................16
1.7.2 Preprocessing........................................................................................................................16
1.7.3 Feature Extraction................................................................................................................16
1.7.4 Matching...............................................................................................................................17
Chapter 2 Literature Review 18
2.1 Our Approach..............................................................................................................................25
viii
Chapter 3 Methodology 27
3.1 Image Acquisition:......................................................................................................................27
3.2 Pre-processing.............................................................................................................................29
3.2.1 Iris Localization....................................................................................................................29
3.2.2 Normalization.......................................................................................................................34
3.3 Feature Extraction.......................................................................................................................38
3.3.1 Overview..............................................................................................................................38
3.3.2 Implementation of feature extraction...................................................................................40
3.4 Matching......................................................................................................................................41
3.4.1 Hamming Distance...............................................................................................................41
3.4.2 Weighted Euclidean Distance..............................................................................................42
3.4.3 Normalized Correlation........................................................................................................42
3.4.4 Support Vector Machine......................................................................................................42
3.4.5 Implementation of Hamming Distance................................................................................43
3.4.6 Implementation of Support Vector Machine........................................................................45
3.4.7 Random Forest.....................................................................................................................47
Chapter 4 Results and Discussion 49
4.1 Tests and Results of Hamming Distance.....................................................................................50
4.1.1 Performance Measures.........................................................................................................50
4.1.2 Tests and Results..................................................................................................................50
4.2 Tests and results of SVM............................................................................................................54
Chapter 5 Conclusion 56
5.1 Summary of Work.......................................................................................................................56
5.2 Summary of Findings..................................................................................................................57
5.3 Future Work................................................................................................................................57
REFERENCES 59
ix
List of Figures
Figure 1.1 Simple biometric system.....................................................................................................................12
Figure 1.2 A front-on view of the human eye......................................................................................................15
Figure 1.3 Iris recognition system methodology..................................................................................................16
Figure 2.1 The phase demodulation process used to encode the iris patterns....................................................19
Figure.2.2 Specular reflection removal: (a) An original iris image occluded by reflections.(b) Thereflection map
and the envelop points, denoted with bright and grey points, respectively........................................................24
Figure 2.3 Overlapping angular patches with their various parameters [15]......................................................25
Figure 2.4 Showing a Log-Gabor function...........................................................................................................26
Figure 3.1 Typical stages of iris recognition........................................................................................................27
Figure 3.2 Two samples of same individual: (a) Sample taken in session 1, (b) Sample taken in session 2.........28
Figure 3.3 Iris Segmentation Process..................................................................................................................30
Figure 3.4 (a) original image, (b) Iris localization, (c) segmenting by CHT, (d) Removal of eyelashes.................32
Figure 3.5 Image 001_1_1 of CASIA segmented accurately................................................................................33
Figure 3.6 Image 003_1_1 of CASIA database, Eyelashes removal.....................................................................34
Figure 3.7 Dougman’s Rubber Sheet Model........................................................................................................35
Figure 3.8 Outline of the normalization process with radial resolution of 10 pixels, and angular resolution of 40
pixels...................................................................................................................................................................37
Figure 3.9 a) An image from CASIA Database b) its Normalized iris image using the Cartesian to polar
transformation....................................................................................................................................................38
Figure 3.10 Iris coding using circular Hough transform: (a) With eyelashes, (b) Without eyelashes...................40
Figure 3.11 Bit patterns representing a template...............................................................................................41
Figure 3.12 An illustration of the shifting process...............................................................................................44
Figure 3.13 SVM with Linear separable data......................................................................................................45
Figure 3.14 Hierarchical decomposition of a Random Forestsclassifier on a data set.........................................48
Figure 4.1 FRR and Accuracy of Hamming Distance............................................................................................53
x
List of Tables
Table 1- Formulation for kernel function________________________________________________________47
Table 2-Different Thresholds for HD and their corresponding FRR and FAR values________________________51
Table 3-HD values for 15 different individuals____________________________________________________52
Table 4 Experiments and Results of HD_________________________________________________________53
Table 5-Experiments and Results of SVM________________________________________________________55
xi
Chapter 1
Introduction
1.1 What is Biometric?
Biometric authentication is the task of verifying the identity of someone using some
biometric information. Biometric characteristics can be divided in two main classes:
Behavioral. These are related to the behavior of a person like typing rhythm, gait,
and voice. Some researchers have coined the term behaviometrics for this class of
biometrics.
Physiological. These are related to the shape of the body like fingerprint, face
recognition, DNA, and Palm print, hand geometry, iris recognition.
1.1.1 How a Biometric System works?
Biometric systems work by first capturing a sample of the feature, such as
recording a sound signal for voice recognition, or taking a color image for face
recognition. This sample is then transformed using some sort of mathematical
function into a biometric template. The biometric template provides a normalized,
efficient and highly discriminating representation of the feature, which can then be
objectively compared with other templates in order to determine identity. Most
biometric systems allow two modes of operation. An enrolment mode for adding
templates to a database, and an identification mode, where a template is created for an
individual and then a match is searched for in the database of pre-enrolled templates.
1.2 BIOMETRICS AUTHENTICATION
The Biometric authentication system includes three layered architecture:
Enroll
Verify
12
Identify
Figure 1.1 Simple biometric system.
1.3 Applications of Biometrics
There are many applications of biometrics; some of these are categorized as follows.
1.3.1 Forensic (Law Enforcement Department)
Forensic applications are helpful to the court of law.
i. Dead body identification like DNA test etc.
ii. Criminal Investigation like Fingerprint of a killer on the killing tool etc.
iii. Missing children identification.
1.3.2 Government
These applications are helpful for the government of any country.
i. NIC or PASSPORT etc.
ii. Driver license.
iii. Welfare payment.
13
1.3.3 Commercial
These applications are helpful for both organization and customers. These include:
i. ATM or Online Banking etc.
ii. Accessing devices like Login to a computer etc.
iii. Cellular Phones.
iv. Security system for a trade center.
1.4 Biometric Characteristics
A good biometric is characterized by use of a feature that is; highly unique – so that the
chance of any two people having the same characteristic will be minimal, stable – so that the
feature does not change over time, and easily capturable – in order to provide convenience to
the user, and prevent misrepresentation of the feature.
There are several biometric identification systems, like Retina Scan Identification
System, Face Identification System, Signature Identification System, Voice Identification
System, Finger Print Recognition System and Iris Recognition System etc.
1.5 Why Iris Recognition?
We have chosen iris recognition because of following reasons:
The iris has the great mathematical advantage over the other biometrics that
is; its pattern variability among different persons is enormous.
It is believed to be stable throughout the person’s life.
Acquiring iris image is easy.
Template size is small so making it timely efficient.
14
As iris is internal organ so highly protected and rarely changed and damaged.
Iris pattern possesses high degree of randomness because the natural
phenomenon does not repeat.
Image encoding and matching is fast because of small template size.
1.6 IRIS RECOGNITION
In Iris recognition system, an individual is identified based on the Iris contained in
eye image. Thus people are carrying with them a live key or password that will never be
forgotten and will always be there. There is no burden or key management or fear of
identity theft. People feel relax with implementation of iris recognition system because it
is the most convenient and the most secure biometric than the others.
1.6.1 HISTORY OF IRIS RECOGNITION
Ophthalmologists Alphonse Bertillon and Frank Burch were one among the first to
propose that iris patterns can be used for identification systems. In 1992, John Daugman [1]
was the first to develop the pupil identification software. Other important contribution was by
R.Wildes et al [11]. Their method differed in the process of iris code generation and also in
the pattern matching technique. The Daugman system has been tested for a billion images
and the failure rate has been found to be very low. Resultant system is patented by the Iris
can Inc. and is also being commercially used in Iridian technologies, UK National Physical
Lab, British Telecom.
1.6.2 THE HUMAN IRIS
The iris is a thin circular diaphragm, which lies between the cornea and the lens of the
human eye. Frontal view of the iris is shown in Figure 1.2. The iris is perforated close to its
centre by a circular aperture known as the pupil. The function of the iris is to control the
15
amount of light entering through the pupil, and this is done by the sphincter and the dilator
muscles, which adjust the size of the pupil. The average diameter of the iris is 12 mm, and
the pupil size can vary from 10% to 80% of the iris diameter [30].
Figure 1.2 A front-on view of the human eye.
The iris consists of a number of layers; the lowest is the epithelium layer, which
contains dense pigmentation cells. The stromal layer lies above the epithelium layer, and
contains blood vessels, pigment cells and the two iris muscles. The density of stromal
pigmentation determines the color of the iris. The externally visible surface of the multi-
layered iris contains two zones, which often differ in color. An outer ciliary zone and an
inner pupillary zone, and these two zones are divided by the collarette – which appears as a
zigzag pattern [31].
Formation of the iris begins during the third month of embryonic life. The unique
pattern on the surface of the iris is formed during the first year of life, and pigmentation of
the stromal takes place for the first few years. Formation of the unique patterns of the iris is
random and not related to any genetic factors. The only characteristic that is dependent on
genetics is the pigmentation of the iris, which determines its color. Due to the epigenetic
16
nature of iris patterns, the two eyes of an individual contain completely independent iris
patterns, and identical twins possess uncorrelated iris patterns.
1.7 Iris Recognition System Methodology
Iris recognition system has following main steps:
1.7.1 Image Acquisition
For image acquisition a monochrome CCD camera is required. As we are not going
to implement a practical system that is why we have used online available iris image
database provided by CASIA (Chinese Academy of Sciences Institute of Automation).
1.7.2 Preprocessing
Preprocessing of iris includes Noise removal, segmentation (Isolation of the iris region
from the eye image) and Normalization.
1.7.3 Feature Extraction
The unique features of the iris are extracted and then converted to biometric templates that
can be stored in database.
Figure 1.3 Iris recognition system methodology
17
1.7.4 Matching
In this step, two iris templates are compared. During this process a test of statistical
independence is performed. If the test fails, this means two irises are different otherwise they
are same. At this stage the system has to make decision based on the outcome of matching
whether to except or reject.
18
Chapter 2 Literatur
e Review
From the extensive literature review, we have divided the iris recognition approaches
roughly into four major categories based on feature extraction strategy: Phase based
approaches ([1],[2],[3]), Zero-crossing representation approaches ([4], [5]), texture analysis
based approaches ([6], [7], [8], [9], [10], [11], [12]) and intensity variation analysis
approaches ([13], [14]).
Phase Based Approach: In this approach, the iris structure is encoded by demodulating it
with Gabor wavelets. Each phasor is quantized in the complex plane to the quadrant in which
it lies for each local element of the iris structure, and this operation is repeated all across the
iris, at different scales of analysis [2].
Zero-Crossing Representation Approach: The zero-crossing representation of the 1D
wavelet transform at different resolution levels of a concentric circle on an iris image is used
to characterize the texture of the iris [4].
Texture Analysis Based Approach: The iris pattern provides abundant texture information
and it is desirable to explore the representation scheme to acquire the local information in an
iris. Texture analysis based approach captures discriminating frequency information of the
local spatial pattern in an iris, to represent the significant local structure of the iris [9].
Intensity Variation Analysis Approach: In this scheme, the local sharp variation is used to
represent the distinctive structure of the iris. The most discriminative variations of an iris are
characterized by constructing a set of 1-D signals by adopting wavelet transform to represent
these signals [14].
In 1987, Flom and Safir first introduced the concept of automated iris recognition system.
Later on, several researchers worked on iris segmentation, normalization, feature extraction,
and recognition process and achieved a great progress in this field. The multi scale Gabor
filters were used to demodulate the texture phase structure of iris ([1], [2], [3]).A 1024-point
complex valued phasor is computed by filtering Iris image with filter banks. It showed the
19
phase structure of iris at different scales. Each phasor was then quantized into one of the four
quadrants in the complex plane, as shown in Fig. 2.1.
Figure 2.4 The phase demodulation process used to encode the iris patterns
The 2048 resulting components of iris code were used to represent an iris and the
matching between a pair iris codes was measured by their hamming distance.
At various resolution levels, zero crossing of the wavelet transform was calculated over
the concentric circles on the iris. The resulting 1-D signals were compared with model
features using different dissimilarity functions [4]. The method represented in [4] was further
enhanced in [5] where different distance measures for matching such as: Hamming Distance
and Euclidean Distance were used. The iris recognition algorithm in [5], exploited the
integro-differential operator to find the inner and outer boundaries of the iris, and the Gabor
filters are used to extract the unique binary vectors constituting iris code. The Hough
transform was applied for iris localization, a laplacian pyramid was used to represent
distinctive spatial characteristic of human iris and the modified normalized correlation was
applied for matching process [20].An iris image was decomposed into four levels using 2-D
Haar wavelet transform. The fourth level high-frequency information was quantized to form
87-bit code and competitive learning neural network was acquired for classification.
20
Authors introduced an iris detection technique based on the combination of integro-
differential operator with a Hough transform [21]. Firstly it consist of an edge calculation
method, in which the position of the eye is approximated in the global image and secondly
integro-differential operators, to search for pupil boundary, iris boundary and iris centre.
From the original image and its Hilbert transform the analytic image is constructed to analyze
iris characteristics. For extraction of local iris features, emergent frequency functions were
used. The extracted features are encoded to generate iris code by thresholding both the
modules of emergent frequency, and the real and imaginary parts of the instantaneous phase.
And finally for the recognition purpose, the Hamming distance was applied to measure the
fraction of disagreeing bits between a pair of templates by a bit by bit comparison.
In [10], the band pass filtering was applied on the input iris image to minimize the
effect of high frequency noise and DC frequency difference. Resultant image was divided
into eight directional sub band outputs using a directional filter bank and the normalized
directional energy was extracted as features. The iris matching was achieved through
Euclidean distance between the input and output feature vectors.
Another technique, namely, Independent Component Analysis (ICA) was deployed to
produce the optimal basis vectors for extracting feature vectors [6].These basis vectors were
used to isolate both in the space and frequency domains and the ICA expansion coefficients
were used as feature vectors and each iris feature vector was encoded as iris code.
In [16] Scale Invariant Feature Transform (SIFT) approach is proposed. Contrary to
the traditional Iris recognition systems, the SIFT approach does not rely on the
transformation of the iris patterns to polar coordinates or on highly accurate segmentation,
thereby ,allowing less constrained image acquisition conditions. Characteristic sift feature
points are extracted in scale space and matching is performed based on the texture
information around the feature points using the SIFT operator. The SIFT algorithm works as
follows:
First of all, scale space local extrema detection is performed. Then, in accurate key
point localization, two thresholds are used, one to exclude low contrast points and other to
exclude edge points. Then, in orientation assignment an orientation histogram is formed from
21
the gradient orientations within a 16x16 region around each key point. Then, in key point
descriptor stage, a distinctive descriptor is computed around each key point. The image
gradient magnitudes and orientations are sampled within a 16x16 regions around each
keypoint. These samples are then accumulated into orientation histogram of summarizing the
contents over 4x4 sub regions. Each orientation histogram has 8 bins covering the 360 degree
of orientations. The descriptor is then formed from a vector containing the values of all the
orientation histogram entries; therefore having a 4x4x8 =128 points feature vector for each
key point. Finally keypoint matching is performed between two images by comparing each
local extrema based on the associated descriptors. For experimental evaluations, BioSec
multimodal database is used, which includes 3,200 iris images from 200 individuals. The
SIFT operator along with popular approach based on transformation to polar coordinates and
1D Log Gabor filters achieve a performance improvement of 24% in EER.
In [17], a segmentation method is proposed to handle degraded images acquired in
less constrained conditions. The following contributions are offered: 1) the sclera is
considered the most easily distinguishable part in the eye in degraded images, 2) a new type
of feature is proposed to measure the proportion of sclera in each direction and is
fundamental in segmenting the iris and 3) for real time applications, the entire procedure is
run in deterministically linear time in respect to the size of the image. The segmentation
method is divided into two parts: 1) detecting noise free iris regions and 2) parameterizing
the iris shape. Feature extraction stages uses local features to detect sclera and noise free iris
pixels.
In [19], an iris recognition system is presented based on the Morlet Wavelet
Transform real coefficients. Firstly it locates the iris, then it normalizes it and gets 512x64
dimensional rectangular iris image. This ensures the effective iris area. Secondly, it applies
one dimension Morlet wavelet transform row by row to the iris image in the effective iris
area. This gives a series of wavelet transform real coefficients hamming different scales
along with distribution figure of these coefficients at different scales. Thirdly, it makes the
binary iris codes according to the real coefficients at different scales and figures the iris
22
pattern by iris codes. Finally, it sorts the different iris patterns by pattern matching method.
Experiments were done using CASIA iris database of total 567 images, 7 images each from
81 individuals. Recognition rate of 96.474% is achieved.
In [22], Grey Level Co-occurrence Matrix (GLCM) is proposed for iris texture
extraction. This approach is a non-filter based technique and is invariant to image rotation.
For pre-processing, traditional method of J. Dougman is used. From normalized iris image,
the iris features are extracted through second order statistics of pixel intensities. The co-
occurrence matrix estimates the joint probability distribution function of grey level pairs in
the image. This technique uses the GLCM of an image and provides a simple approach to
capture the spatial relationship between two points in a texture pattern. It is calculated from
the normalized iris image using pixels as primary information. The system has been tested on
CASIA database for total 756 images, where 7 images each from 108 individuals and gave
EER of 9.32%.
In [23], GLCM based Haralick features are used for recognition. Iris localization
includes the detection of the iris boundaries and isolation of the collarette region, which is
regarded as one of the most important areas of the iris complex pattern. In the normalization
step, localized iris is transformed into a rectangular block of fixed dimension. To extract the
features, the normalized image is processed by 2-D Haar wavelet and from the low frequency
data Grey level co-occurrence matrix (GLCM) based Haralick features are computed.
Finally, probabilistic neural network (PNN) is used for matching purposes. For performance
evaluation, UBIRIS database of 241 individuals having 5 images each is used. By using
different sets for training and testing, the overall accuracy of 97.00% is achieved.
An enhancement in iris recognition performance is discussed in [24] with the
introduction of Template Fusion technique. The concept of template fusion is used in pattern
recognition and is applied generally in two ways. In first way, the image is segmented and
then segmented feature objects and the original image are fused to increase the rate of object
recognition. The second involves segmenting the input image into different regions, and then
the fusion weight is determined according to the values of salience and visibility of each
23
region, which reflects its clarity. Decomposing source image into blocks and reconstructing a
final fused image by selecting the image blocks from the source images based on the
comparison of the outputs is also considered. And finally, the matching process is carried out
using Hamming distance as a metric for iris recognition. Experimental work is done on
MMUI database by using 450 images of 45 individuals, achieving matching accuracy of
93.33%.
A new approach in iris recognition is proposed in [25] based on Circular Fuzzy Iris
Segmentation (CFIS) and Gabor Analytic Iris Texture Binary Encoder (GAITBE). This
procedure is designed to guarantee that from similar eye images similar iris segments are
obtained, but the degree of occlusion may vary from image to image. The resulting circular
iris ring which is concentric with the pupil approximates the iris. Hilbert Transform
technique is used to extract the iris features.
In [26], an approach based on statistical feature extraction is presented. In this
method, segmentation is done by using Sobel Edge Detector to get inner and outer
boundaries of iris and pupil. Various sizes of lines like concentric circles are drawn using
statistical features (mean, mode, median, variance and standard deviation). Hamming
Distance is used for matching purpose.
An accurate and fast segmentation method is proposed in [27]. Specular reflections
appear as the brightest points in the iris image. Bilinear interpolation method is used to fill
these reflections. Fig. 2.2 illustrates the method. First, an adaptive threshold is used to
calculate the binary reflection map of original image, as shown in Fig. 2.2b. To interpolate
the reflection point, four envelop points are calculated. Once all the envelop points are
calculated, reflection is filled, as shown in Fig. 2.2c.
24
Figure.2. 5 Specular reflection removal: (a) An original iris image occluded by reflections. (b)
The reflection map and the envelop points, denoted with bright and grey points, respectively.
(c) The results of reflection removal. [27]
After efficient reflection removal, an adaboost-cascade iris - detector is first build to
extract a rough position of the iris centre. Then, edge points of iris boundaries are detected.
To deal with non circular iris boundaries a spline-based edge fitting scheme is presented.
Then eyelids are localized through edge detection followed by a curve fitting. A rank filter is
adopted for noise elimination and to tackle the shape irregularity of eyelids a histogram filter
is used. At last eyelashes and shadows are detected through a learned prediction model. For
eyelashes and shadows detection the model provides an adaptive threshold by analyzing the
intensity distributions of different iris regions.
In [15], an iris coding methodology based on discrete cosine transform (DCT) is
presented. First, the feature vectors are derived from the zero crossing of the differences
between 1D DCT coefficients calculated in rectangular image patches, as shown in Fig. 2.3.
25
Figure 2.6 Overlapping angular patches with their various parameters [15].
With the help of appropriate windowing, averaging across the width of these patches reduces
the noise effects and smoothes the data. Then, 1D DCT is used to code each patch along its
length. The values are selected by experimenting on CASIA and BATH databases. To reduce
the feature vector length, the most discriminating binarized DCT coefficients are extracted
and the remaining ones are discarded. For comparison of two templates, distance between
two feature vectors is measured using the product-of-sum (POS) of individual sub feature
Hamming distances (HD). Two databases, 2156 images of 308 eyes from CASIA and, 2955
images of 150 eyes from Bath are used for testing. On this data set, it achieved 100% Correct
Recognition Rate (CRR) and perfect Receive-Operating Characteristics (CRR) curves.
2.1 Our Approach
In our project, we have chosen Circular Hough transform for segmentation. The basic
idea of this technique is to find curves that can be parameterized like straight lines,
polynomials, circles, etc., in a suitable parameter space. The transformation is able to
overcome artifacts such as shadows and noise. The approach is found to be good particularly
dealing with all sorts of difficulties including severe occlusions [32]. For normalization,
Rubber Sheet Model is used. This model has been employed mostly in iris related research,
proposed by J. Daugman [30]. For feature extraction we have used 1-D log Gabor filters.
26
Gabor filters have been used extensively in a variety of image processing problems, such as
fingerprint enhancement and iris recognition. The log Gabor function has the advantage of
the symmetry on the log frequency axis. The log axis is the optimum method for representing
spatial frequency response of visual cortical neurons [33]. The Log-Gabor filters spread
information equally across the channels as shown in Fig.2.4.
Figure 2.7 Showing a Log-Gabor function
For matching purpose we have worked on Hamming distance. The Hamming distance
Approach is a matching metric employed by Daugman for comparing two bit patterns and it
represents the number of bits that are different in the two patterns. Another matching metric
that can be used to compare two templates is the weighted Euclidean distance which involves
much computation and this metric is especially used when the two templates consist of
integer values. Normalized correlation matching technique also involves significant amount
of computation. And hence Hamming Distance matching classifier is chosen as it is more
reasonable compared with Weighted Euclidean Distance and Normalized correlation
matching classifiers, as it is fast and simple [34].
Support Vector Machine (SVM) is also employed for classification. SVM can model
complex, real-world problems such as text and image classification, hand-writing
recognition, and bioinformatics and biosequenceanalysis.SVM performs well on data sets
that have many attributes, even if there are very few cases on which to train the model. There
is no upper limit on the number of attributes; the only constraints are those imposed by
hardware. Traditional neural nets do not perform well under these circumstances.
27
Chapter 3 Methodolog
y
Iris recognition system works as a biometric identification of individuals, mainly used
for security purposes. The typical stages of iris recognition system are Image acquisition,
Segmentation, Normalization, Feature extraction and Matching, which are shown in Fig. 3.1.
.
Figure 3.8 Typical stages of iris recognition
Each of the Iris recognition steps is discussed in detail in the coming sections.
3.1 Image Acquisition:
Image acquisition means to capture the user’s iris image. For efficient working of the
iris recognition system a high quality iris image of sufficient resolution and sharpness has to
be captured. It is also important that the interior iris pattern have good contrast. Poor quality
images will be generated if the brightness is not uniformly distributed. Images captured with
infrared camera have good contrast and low reflection [30].The Dougman system captures
images with the iris diameter between 100 and 200 pixels from a distance of 15-46 cm using
a 330 mm lens. Similarly the Wilds et al. system images the iris with approximately 256
28
pixels across the diameter from 20 cm using 80 mm lens. To keep the illumination level
relatively low for operator comfort the optical aperture cannot be too small. Both systems
have fairly small depths of field of 1 cm. This is enough to protect against blur due to eye
movements provided that the operator is attempting to maintain a steady gaze. Both the
Daugman and Wildes et al. systems require the operator to self-position his eye region in
front of the camera [98].
Due to the unavailability of the camera we selected the iris database provided by the
Institute Of Automation from the Chinese Academy of Science (CASIA). This database is
the most preferable choice of researchers. CASIA iris image database (version 1.0) consists
of total 108 individuals images. So there are 108 different classes. Each individual has 7
images so making 756 total iris images. These images were captured in two sessions, 3
samples in first session and 4 samples in second session. All the tests are performed using the
images from this database. In Fig. 3.2, 2 samples of same individual are shown taken in two
different sessions.
(a) (b)
Figure 3.9 Two samples of same individual: (a) Sample taken in session 1, (b) Sample taken in
session 2
29
3.2 Pre-processing
Pre-processing stage of iris recognition consists of two steps:
1) Iris Localization (Segmentation).
2) Normalization.
3.2.1 Iris Localization
3.2.1.1 Overview
The image acquisition stage may not yield an image containing only iris, it can
produce an image in which iris is occluded by eyelids and eyelashes. Also, specular
reflection may occur within the iris region corrupting the iris pattern. Therefore, a technique
is required to isolate and exclude these occurring artifacts as well as locating the circular iris
region prior to performing iris pattern matching. Particularly, it is important to localize that
portion that lies between the sclera and iris i.e. the border. Further, if the eyelids are
occluding part of the iris, then only that portion of the image below the upper eyelid and
above the lower eyelid should be included [3]. Wildes in [3] described various iris
observations and concluded that iris localization must be sensitive to a wide range of edge
contrast, robust to irregular borders, and capable of dealing with variable occlusions.
For the success of an iris recognition system, segmentation stage is very critical,
because data that is falsely represented as iris pattern data can corrupt the generated
biometric templates that result in poor recognition performance. The success of segmentation
particularly depends on the quality of images captured. In CASIA iris image database,
images do not have specular reflections due to the Infra-Red light used for illumination.
Images captured in natural light present very low contrast between iris and pupil region,
consequently, segmentation process will be more difficult for images taken in natural light.
The iris localization process takes the acquired image and finds both the boundary
between pupil and the iris, and the boundary between the sclera and the iris. The noise
30
reduction process refers to localizing the iris from noise in the image. The noise consists of
sclera, pupil, eyelids, eyelashes and artifacts .Fig. 3.3 illustrates the iris segmentation step
[24].
Iris Localization Occlusion effect Noise reduction
Figure 3.10 Iris Segmentation Process
Iris segmentation method includes edge detection using the circular Hough transform
[11] and Integro-differential operator [1].The method in [1] assumes the pupillary and limbic
boundaries of the eye as circles and an integro-differential operator is used to detect the iris
boundary by searching the parameter space. When the integro-differential operator attains its
maximum the circular boundary is detected. The iris boundary is explained with three
parameters, centre coordinates of the circle, xₒ and yₒ, and radius r [24]. In Circular Hough
Transform method an edge map is generated by calculating the first derivatives of intensity
values in an eye image and then thresholding the result. From the edge map, votes are cast in
Hough space for the parameters of circles passing through each edge point. These parameters
are the centre coordinates and the radius which are sufficient to describe any circle.
In our methodology we used the Circular Hough Transform to locate the iris in an
individual’s eye image. The details of the procedure are discussed in the following section.
31
3.2.1.2 Circular Hough Transform
The Hough transform is a standard computer vision algorithm that can be used to
determine the parameters of simple geometric objects, such as lines and circles, present in an
image. The circular Hough transform can be employed to deduce the radius and centre
coordinates of the pupil and iris regions. An automatic segmentation algorithm based on the
circular Hough transform is employed by Wildes et al. [11], Tisse et al. [21], and Ma et al.
[8]. Firstly, an edge map is generated by calculating the first derivatives of intensity values in
an eye image and then thresholding the result. From the edge map, votes are cast in Hough
space for the parameters of circles passing through each edge point. These parameters are the
centre coordinates xcand yc, and the radius r, which are able to define any circle according to
the equation 3.1.
X2c+y2
c = r Eq. (3.1)
A maximum point in the Hough space will correspond to the radius and centre
coordinates of the circle best defined by the edge points. Wildes et al. and Kong and Zhang
also make use of the parabolic Hough transform to detect the eyelids, approximating the
upper and lower eyelids with parabolic arcs, which are represented as
(- (x – hj)sinθj+ (y – kj)cosθj )2 = aj (( x – hj) cosθj + ( y – kj) sin θj) Eq. (3.2)
Where aj controls the curvature, (hj, kj) is the peak of the parabola and θ j is the angle of
rotation relative to the x-axis. In order to detect limbus, only vertical edge information is
used. The upper and lower parts, which have the horizontal edge information, are usually
covered by the two Eyelids. The horizontal edge information is used for detecting the upper
and lower eyelids, which are modeled as parabolic arcs. Fig. 3.4 illustrates the segmented
human iris with and without eyelashes using the circular Hough transform.
32
(a) (b)
(c) (d)
Figure 3.11 (a) original image, (b) Iris localization, (c) segmenting by CHT, (d) Removal of
eyelashes
3.2.1.3 Implementation
By using Circular Hough Transform to detect the iris and pupil boundaries, the first
step is to apply the Canny edge detector to produce an edge map of the image. For the outer
boundary between iris and sclera the first derivatives were biased vertically. For inner
iris/pupil boundary, vertical and horizontal gradients were weighted equally.
Depending on the database used, the range values of radius were set manually. For
the CASIA database, values of the iris radius range from 90 to 150 pixels, while the pupil
radius ranges from 28 to 75 pixels. In order to make the circle detection process more
efficient and accurate, the Hough transform for the iris/sclera boundary was performed first,
then the Hough transform for the iris/pupil boundary was performed within the iris region,
instead of the whole eye region, since the pupil is always within the iris region. After this
process was completed, six parameters were stored, the radius, and x and y centre
coordinates for both the circles.
33
Eyelids were isolated by first fitting a line to the upper and lower eyelid using the
linear Hough transform. A second horizontal line is then drawn, which intersects with the
first line at the iris edge that is closest to the pupil. This process is done for both the top and
bottom eyelids. The second horizontal line allows maximum isolation of eyelid regions.
Canny edge detection is used to create an edge map, and only horizontal gradient information
is taken. The linear Hough transform is implemented using the MATLAB Radon transform,
which is a form of the Hough transform. If the maximum in Hough space is lower than a set
threshold, then no line is fitted, since this corresponds to non-occluding eyelids. Also, the
lines are restricted to lie exterior to the pupil region, and interior to the iris region. A linear
Hough transform has the advantage over its parabolic version, in that there are less
parameters to deduce, making the process less computationally demanding. A simple
thresholding technique was used for isolating eyelashes in the CASIA database, since
analysis reveals that eyelashes are quite dark when compared with the rest of the eye image.
For the eyelid, eyelash, and reflection detection process, the coordinates of any of these noise
areas are marked using the MATLAB NaN type, so that intensity values at these points are
not misrepresented as iris region data.
The CASIA database gave good segmentation results, since images in this database
are taken particularly for iris related search and obviously there is a distinguishable clear
boundary between iris and sclera. A correctly segmented image is shown in Fig. 3.5.
Figure 3.12 Image 001_1_1 of CASIA segmented accurately
34
The eyelash detection system implemented for the CASIA database also proved to be
successful in isolating most of the eyelashes occurring within the iris region as shown in Fig.
3.6.
Figure 3.13 Image 003_1_1 of CASIA database, Eyelashes removal
3.2.2 Normalization
3.2.2.1 Overview
Normalization refers to preparing a segmented iris image for the feature extraction
process. After successfully segmenting the iris region from an image, the next stage is to
transform the iris region into fixed dimensions in order to allow comparisons. The captured
iris images from different persons may differ in sizes even for two same eye images. This is
due to the stretching of iris caused by pupil dilation from illumination variations. Other
sources of inconsistency include, varying imaging distance, rotation of the camera, head tilt,
and rotation of the eye within the eye socket. The normalization process will produce iris
regions, which have the same constant dimensions, so that two images of the same iris under
different conditions will have characteristic features at the same spatial location.
We used Rubber Sheet Model proposed by J. Dougman [1], [2], [3] to normalize the
segmented iris region. This model is discussed in the following section.
35
3.2.2.1.1 Rubber Sheet Model
The homogenous rubber sheet model remaps each point within the iris region to a pair
of polar coordinates (r,θ) where r is on the interval [0,1] and θ is angle [0,2π]. This model is
illustrated in Fig. 3.7.
Figure 3.14 Dougman’s Rubber Sheet Model
The localized region is mapped from Cartesian coordinates to normalized non-concentric
polar representation of fixed size by using the following procedure.
I (x (r,θ), y (r,θ))→ I (r,θ) Eq. (3.3)
With
x ( r , θ ) = ( 1 – r ) xp (θ) + rxl (θ)
y ( r , θ ) = ( 1 – r ) yp (θ) + ryl (θ)
Where I(x, y) is the iris region image, (x, y) are the original Cartesian coordinates, (r, θ) are
the corresponding normalized polar coordinates, and xp, yp and xl,yl are the coordinates of the
pupil and iris boundaries along the θ direction. The rubber sheet model considers several
pupil dilation and size inconsistencies to produce a normalized representation with constant
dimensions. In this way the localized region is modeled as a flexible rubber sheet anchored at
the iris boundary with the pupil centre as the reference point.
36
3.2.2.2 Implementation
The rubber sheet model was employed by considering the centre of the pupil as the
reference point, and radial vectors are passed through the iris region, as shown in Figure 3.8.
Along each radial line, a number of data points are selected and this is defined as the radial
resolution. The number of radial lines going around the iris region is defined as the angular
resolution. Since the pupil can be non-concentric to the iris, a remapping formula is needed
to rescale points depending on the angle around the circle.
r’=sqrt(α)(β) ± sqrt (α β2 – α – r2I) Eq. (3.4)
With
α = ox2 + oy
2
β = cos (π – arctan(oy/ox ) – θ )
Where displacement of the centre of the pupil relative to the centre of the iris is given by ox
,oy, and r’ is the distance between the edge of the pupil and edge of the iris at an angle, θ
around the region, and rI is the radius of the iris. The remapping formula first gives the radius
of the iris region ‘doughnut’ as a function of the angle θ.
A constant number of points are taken along each radial line, such that a constant
number of radial data points are taken, irrespective of how narrow or wide the radius is at a
particular angle. The normalized pattern was created by backtracking to find the Cartesian
coordinates of data points from the radial and angular position in the normalized image.
From the ‘doughnut’ iris region, normalization produces a 2D array with horizontal
dimensions of angular resolution and vertical dimensions of radial resolution. Another 2D
array was created for marking reflections, eyelashes, and eyelids detected in the segmented
image. In order to prevent non-iris region data from corrupting the normalized representation,
data points which occur along the pupil border or the iris border are discarded.
37
Figure 3.15 Outline of the normalization process with radial resolution of 10 pixels, and angular
resolution of 40 pixels.
The normalization process proved to be successful and some of the results are shown
in Figure 3.9. However, the normalization process was not able to perfectly reconstruct the
same pattern from images with some variations in the pupil dilation, since deformation of the
iris results in small changes of its surface patterns. In this example, the rectangular
representation is constructed from 10,000 data points in the iris region.
38
(a)
(b)
Figure 3.16 a) An image from CASIA Database b) its Normalized iris image using the Cartesian
to polar transformation.
Although the normalization method compensates variations due to scale, translation
and pupil dilation, it is not inherently invariant to the rotation of iris. Rotation of an iris in the
Cartesian coordinates is equivalent to a shift in the polar coordinates. In order to compensate
the rotation of iris textures, a best of n test of agreement technique is proposed by Daugman
in the matching process.
3.3 Feature Extraction
3.3.1 Overview
The most distinctive features are required to be extracted from the localized iris
image for accurate identification of persons. The most significant features are needed to be
encoded to construct the iris pattern to compare between several iris patterns for iris
identification. Features involved in an iris image are divided into the following categories.
39
Spectral Features: special color or tone, gradient, spectral parameter etc.
Geometric Features: Edge, lineament, shape, size etc.
Textural Features: Pattern, spatial frequency, homogeneity etc.
In our methodology, we extracted the textural features from iris image and for this purpose;
we used 1D Log Gabor Filters.
3.3.1.1 Gabor Filters
Gabor filters are able to provide optimum conjoint representation of a signal in space
and spatial frequency. A Gabor filter is constructed by modulating a sine/cosine wave with a
Gaussian. This is able to provide the optimum conjoint localization in both space and
frequency, since a sine wave is perfectly localized in frequency, but not localized in space.
Modulation of the sine with a Gaussian provides localization in space, though with loss of
localization in frequency [28].
Decomposition of a signal is accomplished using a quadrature pair of Gabor filters,
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 and imaginary filters are also known
as the even symmetric and odd symmetric components respectively. The centre frequency of
the filter is specified by the frequency of the sine/cosine wave, and the bandwidth is specified
by the Gaussian width. In general, Gabor filters are traditional choice, where they suffer from
two main limitations: the maximum bandwidth is limited to approximately one octave and
they are not optimal if one is seeking broad spectral information with maximal spatial
localization. An alternative to the Gabor filter is the log-Gabor filter [28].
3.3.1.2 Log Gabor Filters
An alternative to the Gabor function is the log-Gabor function proposed by Field [29].
Field suggested that natural images are better coded by filters that have Gaussian transfer
functions when viewed on the logarithmic frequency scale. (Gabor functions have Gaussian
40
transfer functions when viewed on the linear frequency scale). The frequency response of a
Log-Gabor filter is given as;
G( f)= exp{ − 0.5×log(f / f0)2 / log(σ / f)2} Eq. (3.5)
Where fo represents the centre frequency, and σ gives the bandwidth of the filter. Details of
the log-Gabor filter are examined in [29]. There are two important characteristics to note.
Firstly, log-Gabor functions, by definition, always have no DC component, and secondly, the
transfer function of the log Gabor function has an extended tail at the high frequency end.
Fig. 3.10shows the resulting iris code with and without eyelashes [28].
(a)
(b)
Figure 3.17 Iris coding using circular Hough transform: (a) With eyelashes, (b) Without
eyelashes
3.3.2 Implementation of feature extraction
Feature extraction was implemented by convolving 1D Log-Gabor filters with
the normalized iris pattern. The 2D normalized pattern was broken up into a number of 1D
signals, and then these 1D signals were convolved with 1D Gabor filters. The rows of the 2D
normalized pattern are taken as the 1D signal; each row corresponds to a circular ring on the
iris region. The angular direction is taken rather than the radial one, which corresponds to
41
columns of the normalized pattern, since maximum independence occurs in the angular
direction.
The intensity values at known noise areas in the normalized pattern are set to the
average intensity of surrounding pixels to prevent influence of noise in the output of the
filtering. The output of filtering is then phase quantized to four levels using the Daugman
method [1], with each filter producing two bits of data for each phasor. The output of phase
quantization is chosen to be a grey code, so that when going from one quadrant to another,
only 1 bit changes. This will minimize the number of bits disagreeing, if say two intra-class
patterns are slightly misaligned and thus will provide more accurate recognition.
The encoding process produces a bitwise template containing a number of bits of
information, and a corresponding noise mask which corresponds to corrupt areas within the
iris pattern, and marks bits in the template as corrupt. Since the phase information will be
meaningless at regions where the amplitude is zero, these regions are also marked in the
noise mask. The total number of bits in the template will be the angular resolution times the
radial resolution, times 2, times the number of filters used. Fig. 3.11 shows a template
generated.
Figure 3.18 Bit patterns representing a template
3.4 Matching
The last step in iris recognition system is the matching of individual iris template with
that of iris template from the database. Various matching algorithms exist in literature. Some
of them are discussed in the following.
42
3.4.1 Hamming Distance
The Hamming distance gives a measure of how many bits are the same between two
bit patterns. Using the Hamming distance of two bit patterns, a decision can be made as to
whether the two patterns were generated from different irises or from the same one. An
individual iris region contains features with high degrees of freedom; each iris region will
produce a bit-pattern which is independent to that produced by another iris. On the other
hand, two iris codes produced from the same iris will be highly correlated. Hamming
distance is the matching metric employed in [11]. The advantage of hamming distance is fast
matching speed because the templates are in binary format.
3.4.2 Weighted Euclidean Distance
The weighted Euclidean distance (WED) can be used to compare two templates,
especially if the template is composed of integer values. The weighting Euclidean distance
collects the values between two templates. This metric is employed by Zhu et al. [12]. An iris
template is matched with all templates in the database. The two templates are matched if the
WED is minimum.
3.4.3 Normalized Correlation
Normalized correlation between acquired and database representation for goodness of
match is being employed Wildes et.al [11]. Normalized correlation is advantageous over
standard correlation, since it is able to account for local variations in image intensity that
corrupt the standard correlation calculation. The disadvantage of the normalized correlation
is that it is computationally not much effective.
3.4.4 Support Vector Machine
SVM is an upcoming template matching technique. It is used for pattern matching to
verify a person’s identity based on the iris code. SVM is a relatively new machine learning 43
technique, which is based on the principle of structural risk minimization. An SVM is a
binary classifier that optimally separates the two classes. There are two important aspects in
the development of SVM as classifier. The first aspect is determination of the optimal hyper
plane
Which will optimally separate the two classes and the other aspect is transformation of non-
linearly separable classification problem into linearly separable problem.
In our project, we have selected the Hamming Distance for template matching and
Support Vector Machine as a classifier for classification of different individuals.
We also tried a new classifier “Random Forest” for the classification of iris images.
3.4.5 Implementation of Hamming Distance
For matching, the Hamming distance was chosen as a metric for recognition, since
bit-wise comparisons were necessary. The Hamming distance algorithm employed also
incorporates noise masking, so that only significant bits are used in calculating the Hamming
distance between two iris templates. Now when taking the Hamming distance, only those bits
in the iris pattern that corresponds to ‘0’ bits in noise masks of both iris patterns will be used
in the calculation. The Hamming distance will be calculated using only the bits generated
from the true iris region and this modified Hamming Distance formula is given as follows:
Eq. (3.6)
Where Xj and Yj are the two bit-wise templates to compare, Xnj and Ynj are the corresponding
noise masks for Xj and Yj, and N is the number of bits represented by each template.
Although, in theory, two iris templates generated from the same iris will have a
Hamming distance of 0.0, in practice this will not occur. Normalization is not perfect, and
44
also there will be some noise that goes undetected, so some variation will be present when
comparing two intra-class iris templates.
In order to account for rotational inconsistencies, when the Hamming distance of two
templates is calculated, one template is shifted left and right bit-wise and a number of
Hamming distance values are calculated from successive shifts. This bit-wise shifting in the
horizontal direction corresponds to rotation of the original iris region by an angle given by
the angular resolution used. If an angular resolution of 180 is used, each shift will correspond
to a rotation of 2 degrees in the iris region. This method is suggested by Daugman [1], and
corrects for misalignments in the normalized iris pattern caused by rotational differences
during imaging. From the calculated Hamming Distance values, only the lowest is taken,
since this corresponds to the best match between two templates.
The number of bits moved during each shift is given by two times the number of
filters used, since each filter will generate two bits of information from one pixel of the
normalized region. The actual number of shifts required to normalize rotational
inconsistencies will be determined by the maximum angle difference between two images of
the same eye, and one shift is defined as one shift to the left, followed by one shift to the
right. The shifting process for one shift is illustrated in Figure 3.12.
45
Figure 3.19 An illustration of the shifting process
3.4.6 Implementation of Support Vector Machine
SVM is a relatively new learning machine technique which is based on the principle
of structural risk minimization. An SVM is binary classifier that optimally separates the two
classes. There are two important aspects in the development of SVM as classifier. The first
aspect is determination of the optimal hyperplane which will optimally separate the two
classes and the other aspect is transformation of non-linearly separable classification problem
into linearly separable problem. This section will discuss in brief the two aspects of the SVM
development. Fig. 3.13 shows linearly separable binary classification problem with no
possibility of miss-classification data. Let x and y be a set of input feature vector and the
class label respectively. The pair of input feature vectors and the class label can be
represented as tuples {xi,yi} where i = 1,2,L,N and y = ±1 . In the case of linear separable
problem, there exists a separating hyperplane which defines the boundary between class 1
(labeled as y = 1) and class 2 (labeled as y = -1). The separating hyperplane is:
46
Figure 3.20 SVM with Linear separable data
w .x +b = 0 Eq. (3.7)
Which implies
yi(w.xi + b ) ≥ 1 , i= 1,2……N Eq. (3.8)
Basically, there are numerous possible values of{w,b} that create separating
hyperplane. In SVM only hyperplane that maximizes the margin between two sets is used.
Margin is the distance between the closest data to the hyperlane. Referring to Fig. 3.13, the
margins are defined as d+ and d-. The margin will be maximized in the case d+ =d−
.Moreover, training data in the margins will lie on the hyperplanes H+ and H-. The distance
between hyperplane H+ and H- is
d++ d- = 2/║w║ Eq. (3.9)
47
As H+ and H- are the hyperplane in which the closest training data to the optimal
hyperplane, then there is no training data which fall between H+ and H-. This means the
hyperplane that separates optimally the training data is the hyperplane which
minimizes║w║2so that the distance of Eq. (3.9) is maximized. However, the minimization of
║w║2 is constrained by Eq. (3.8). When the data is non-separable, slack variables,ξi, are
introduced into the inequalities for relaxing them slightly so that some points allow lying
within the margin or even being misclassified completely. The resulting problem is then to
minimize
1/2║w║2+ C (∑i L (ξi)) Eq. (3.10)
Where C is the adjustable penalty term and L is the loss function. The most common used
loss function is linear loss function, L (ξi) = ξi. The optimization of Eq. (3.10) with linear
loss function using Lagrange multipliers approach is to maximize:
Eq. (3.11)
subject to
0 ≤ αi ≤C Eq. (3.12.a)
w=ΣiN αixiyi Eq. (3.12.b)
As matter of fact, αi is zero for every xi except the ones that lie on the margin. The training
data with non-zero αi are called as support vectors. In the case of a non-linear separable
problem, a kernel function is adopted to transform the feature space into higher dimensional
feature space in which the problem become linearly separable. Typical kernel functions
commonly used are listed in Table 1.
48
Table 1- Formulation for kernel function
3.4.7 Random Forest
Random Forests are a combination of tree classifiers such that each tree depends on
the values of a random vector sampled independently and with the same distribution for all
trees in the forest. The generalization error of a forest of tree classifiers depends on the
strength of the
Individual trees within the forest and their inter correlation. Using a random selection of
features in order to split each node yields output error rates that compare equally to
Adaboost, yet they are more robust with respect to noise. While traditional tree algorithms
spend a lot of time choosing how to split at a node.
A Random Forest multi-way classifier Θ(x) consists of a number of trees, with each
tree grown using some form of randomization, where x is an input instance. The leaf nodes of
each tree are labeled by estimates of the posterior distribution over the data class labels. Each
internal node contains a test that best splits the space of data to be classified. A new, unseen
instance is classified by sending it down every tree and aggregating the reached leaf
distributions. The process is described in Fig 3.14. Each tree is grown as follows:
49
Figure 3.21 Hierarchical decomposition of a Random Forests classifier on a data set
If the number of cases in the training set is N, sample N cases at random but with
replacement, from the original data. This sample will be the training set for growing
the tree.
If there are M input variables, a number m<<M is specified such that at each node, m
variables are selected at random out of the M and the best split on these m is used to
split the node. The value of m is held constant during the forest growing.
Each tree is grown to the largest extent possible. Therefore, no pruning is applied.
As regards to the overall error rate of the Random Forests, this is affected by two different
factors:
1. Tree inter-correlation. Highly correlated trees result in high error rate.
2. Robustness (strength) of each individual tree within the forest. Higher strength results in
lower error rates.
Chapter 4 Result
s and Discussion
This chapter explains the performance of iris recognition system. Several tests were
performed to find out the best threshold value of Hamming distance, so that the false match
and false accept rate is minimized, and to ensure that iris recognition can perform accurately
as a biometric recognition. To confirm that system provides an accurate recognition,
50
experiments were also carried out in order to ensure the uniqueness of human iris patterns by
reducing the number of degrees of freedom present in the iris template representation.
In the iris recognition system, there are a number of parameters and optimum values
for these parameters were required in order to provide the best recognition rate. These
parameters are: the radial and angular resolution, r and θ respectively, which give the number
of data points for encoding each template, and the filter parameters for feature encoding. The
filter parameter consists of the number of filters, N, their base wavelength λn, filters
bandwidths given by σ/f, and the multiplicative factor between center wavelengths of
successive filters given by α.
The optimum encoding of iris features was with one 1D Log-Gabor filter with a
bandwidth given by aσ/f of 0.5. The center wavelength of this filter was found to be
dependent on the data set used. A center wavelength of 18 pixels is proved to be optimum for
CASIA database. An optimum template size with radial resolution of 20 pixels, and angular
resolution of 240 pixels was chosen. These parameters generate a biometric template that
contains 9600 bits of information. After the optimum parameters were determined, the
recognition performance was examined next.
4.1 Tests and Results of Hamming Distance
4.1.1 Performance Measures
This proposed system in general makes four possible decisions; the authorized person
is accepted, the authorized person is rejected, the unauthorized person (impostor) is accepted
and the unauthorized person (impostor) is rejected. The accuracy of the proposed system is
then specified based on the rate in which the system makes the decision to reject the
authorized person and to accept the unauthorized person. False Rejection Rates (FRR) is used
to measure the rate of the system to reject the authorized person and False Acceptance Rates
51
(FAR) used to measure the rates of the system to accept the unauthorized person. Both
performances are expressed as:
FRR = NFR / NAAx100% Eq. (4.1)
FAR = NFA / NIAx100% Eq. (4.2)
Where,
NFR is referred to the numbers of false rejections and NFA is referred to the number of false
acceptance, while NAA and NIA are the numbers of the authorized person attempts and the
numbers of impostor person attempts respectively. Furthermore, low FRR and low FAR is the
main objective in order to achieve both high usability and high security of the system.
4.1.2 Tests and Results
The key objective of an iris recognition system is to be able to achieve a distinct
threshold value of Hamming distance in order to make matching process efficient. With clear
separation, a threshold Hamming distance value can be chosen which allows a decision to be
made when comparing two templates. If the Hamming distance between two templates is less
than the threshold value, the templates were generated from the same iris and a match is
found. Otherwise if the Hamming distance is greater than the threshold value the two
templates are considered to have been generated from different irises.
Since it is not possible to use all the 756 images from the CASIA database, we took
105 images of 15 individuals for experiment purpose.
A number of experiments were performed to find out the best case where the False
Acceptance Rate and False Rejection Rate are minimum and the iris recognition is
maximum.
For different threshold values, the following values of FRR and FAR were recorded,
as shown in Table 2.
Table 2-Different Thresholds for HD and their corresponding FRR and FAR values
Threshold FRR (%) FAR (%)52
0.5 0.00 99.5
0.45 0.00 7.6
0.4 0.24 0.005
0.3 37.88 0.00
0.2 99.05 0.00
As seen from Table 2, some threshold values are best for low FAR and some are best for low
FRR. But both the values of FRR and FAR should be optimized in order to make correct
recognition and to obtain highest success rates. The threshold value of 0.4 gives the
acceptable values for both FRR and FAR. So we conclude that threshold value of 0.4 gives
the best possible recognition results.
After selecting the best threshold value for Hamming Distance, we tested the system
on 15 different iris images and the HD values were recorded that are shown in Table 3. These
values show the accuracy of the system as there is only one match for all 15 iris images that
is given by value ‘0’, and all of the rest values are greater than the set threshold,
consequently very less chances of false acceptance.
Table 3-HD values for 15 different individuals
Iris 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
1 00.455
90.470
9 0.4820.467
30.452
90.465
10.449
90.470
60.477
60.469
80.477
50.454
50.481
40.476
7
2 0.4559
0 0.4737
0.4826
0.4708
0.4478
0.4706
0.4644
0.476 0.4564
0.4392
0.4874
0.4804
0.4843
0.4645
3 0.4709
0.4737
0 0.4786
0.4707
0.4484
0.4582
0.4799
0.4868
0.449 0.4697
0.4832
0.4884
0.4806
0.4501
4 0.482 0.4826
0.4786
0 0.4658
0.4641
0.477 0.4705
0.467 0.4615
0.4665
0.4476
0.4149
0.4642
0.4594
5 0.467 0.470 0.470 0.465 0 0.471 0.458 0.477 0.471 0.486 0.476 0.454 0.474 0.482 0.449
53
3 8 7 8 1 2 7 7 2 4 4 6 5
6 0.4529
0.4487
0.4484
0.4641
0.4711
0 0.4642
0.4697
0.4771
0.4674
0.4634
0.4936
0.484 0.4691
0.4262
7 0.4651
0.4706
0.4582
0.477 0.4582
0.4642
0 0.4571
0.4772
0.4651
0.4715
0.4656
0.4693
0.4753
0.4331
80.449
90.464
40.479
90.470
50.477
70.469
70.457
1 00.483
80.474
60.459
10.487
50.468
9 0.4690.478
2
90.470
6 0.4760.486
8 0.4670.471
70.477
1 0.4770.483
8 00.451
60.471
70.477
40.461
70.442
20.473
6
10 0.4776
0.4564
0.449 0.4615
0.4862
0.4674
0.4651
0.4746
0.4516
0 0.4569
0.4737
0.4657
0.4266
0.4392
110.469
80.439
20.469
70.466
50.476
40.463
40.471
50.459
10.471
70.456
9 00.480
10.442
40.483
40.453
4
12 0.4775
0.4875
0.4832
0.4476
0.4544
0.4936
0.4656
0.4875
0.4774
0.4737
0.4801
0 0.4826
0.4746
0.4616
13 0.4545
0.4804
0.4884
0.4149
0.4746
0.484 0.4693
0.4689
0.4617
0.4657
0.4424
0.4826
0 0.4196
0.4503
140.481
40.484
30.480
60.464
20.482
50.469
10.475
30.469
20.442
20.426
60.483
40.474
60.419
6 00.460
3
15 0.4767
0.4654
0.4501
0.4594
0.449 0.4262
0.4331
0.4782
0.4736
0.4392
0.4534
0.4616
0.4503
0.4603
0
To evaluate the accuracy of hamming distance we perform 5 experiments taking 7
individuals in each experiment randomly from CASIA database. For each experiment the
FRR, FAR and recognition accuracy were recorded as shown in Table 4. The average
accuracy of these random 5 experiments comes to be 93.858 %. Figure 1 shows the FRR and
accuracy of each experiment.
Table 4 Experiments and Results of HD
Experiment # FRR (%) FAR (%) Accuracy (%)1 10.20 0.00 89.802 4.10 0.00 95.90
54
3 6.16 0.00 93.844 6.16 0.00 93.845 4.09 0.00 95.91
Figure 22 FRR and Accuracy of Hamming Distance
4.2 Tests and results of SVM
To evaluate the performance of support vector machine as a classifier for iris
recognition, experiments were performed using CASIA database. For SVM simulation the
pre-built software “Libsvm-3.1” was used [35]. Libsvm is simple, easy-to-use, and efficient
software for SVM classification and regression.
As the feature vectors were generated in MATLAB, so in order to classify the data, the
features must be in libsvm format. Therefore, the feature data were converted in libsvm
55
format in MATLAB. Then this converted feature file was used for training and testing using
“Libsvm-3.1”.
We chose 7 individuals for evaluating SVM performance. We performed 5 runs,
taking 7 individuals randomly in each run. In each run, 3 experiments were performed for
different sets of training and testing images. Firstly 3 images of each individual were selected
for training and 4 for testing. Secondly 4 images of each individual were selected for
training and 3 for testing. Thirdly 5 images of each individual were selected for training and
2 for testing. All these experiments and resulted accuracy of classification are shown in Table
5. As we got the same accuracy for all 5 experiments, so their average accuracy is also the
same i.e. 85.71%.
The observation came from Table 5 says that the change in number of training images
does not affect the accuracy of SVM in our case. Each time when we test the accuracy for 7
individuals anywhere from the database, the accuracy comes to be 85.71%. May be this is
due to the images that we have in database. These images are of very high quality and taken
with a high resolution infrared camera. And features obtained from these images are highly
distinctive. This may be the cause that the accuracy is not affected here with the change in
training images.
Table 5-Experiments and Results of SVM
No of Classes No of Training Images No of Testing ImagesAccuracy (%)
73 4 85.71
4 3 85.715 2 85.71
73 4 85.714 3 85.715 2 85.71
56
73 4 85.714 3 85.715 2 85.71
73 4 85.714 3 85.715 2 85.71
7 3 4 85.714 3 85.715 2 85.71
Chapter 5 Conclusio
n
5.1 Summary of Work
In this thesis we have presented an iris recognition system, which was tested using
CASIA database of grayscale eye images in order to verify the performance of iris
recognition technology.
57
Firstly, automatic segmentation was performed through the use of Circular Hough
Transform, which localized the iris region from an eye image and removed eyelid, eyelash
and reflection areas. Thresholding was also employed for isolating eyelashes and reflections.
Next, the segmented iris region was normalized by implementing Daugman’s rubber
sheet model, where the iris is modeled as a flexible rubber sheet, which is unwrapped into a
rectangular block with constant polar dimensions.
Finally, features of the iris were encoded by convolving the normalized iris region
with 1D Log-Gabor filters and phase quantizing the output in order to produce a bit-wise
biometric template. For matching purposes, two approaches were employed, a distance
measure approach (Hamming Distance) and a statistical classification technique (Support
Vector Machine).
The Hamming distance gave a measure of how many bits disagreed between two
templates. A failure of statistical independence between two templates would result in a
match, that is, the two templates were deemed to have been generated from the same iris if
the Hamming distance produced was lower than a set Hamming distance. The Support
Vector Machine made the classification of different iris patterns by first taking the authorized
persons images as input in training, and then by taking the imposters images as input for
testing.
5.2 Summary of Findings
The developed iris recognition system has revealed a number of interesting
conclusions. It can be stated that segmentation is the critical stage of iris recognition, since
areas that are wrongly identified as iris regions will corrupt biometric templates resulting in
very poor recognition. The results show that segmentation can be the most difficult stage of
iris recognition because its success is dependent on the imaging quality of eye images.
Another interesting finding was that the encoding process only required one 1D Log-
Gabor filter to provide accurate recognition, since the open literature mentions the use of 58
multi-scale representation in the encoding process. Also the optimum center wavelength was
found to be dependent on imaging conditions, since different lighting conditions will produce
features of different frequencies.
CASIA provided optimum recognition when encoded using a filter with center
wavelength of 18.0 pixels. For both data sets, a filter bandwidth with σ/f of 0.5, and template
resolution of 20 pixels by 240 pixels was found to provide optimum encoding. A near-perfect
recognition rate was achieved with the CASIA data set, with a threshold value of 0.4, a false
accept rate of 0.005% and false reject rate of 0.24% was possible. For SVM classification the
default parameters gave the accuracy of up to 85.71% for the fixed number of individuals as
explained in detail in chapter 4. While for the same number of individuals Hamming
Distance gave the accuracy of 93.858%. Hamming distance gave the better accuracy than
SVM.
5.3 Future Work
The system presented in this thesis was able to perform accurately, however there are
still a number of issues which need to be addressed. First of all, the automatic segmentation
was not perfect, since it could not successfully segment the iris regions for the entire eye
images in the database used. In order to improve the automatic segmentation algorithm, a
more elaborate eyelid and eyelash detection system could be implemented.
An improvement could also be made in the speed of the system. The most
computation intensive stages include performing the Hough transform, and calculating
Hamming distance values between templates to search for a match. Since the system is
implemented in MATLAB, which is an interpreted language, speed benefits could be made
by implementing computationally intensive parts in C or C++. Speed was not one of the
objectives for developing this system, but this would have to be considered if using the
system for real-time recognition.
59
We can increase the number of experiments as well as no of individuals in future to
evaluate the performance of the system. We can increase the number from 7 to 10, 15 or 20
and also for SVM the number of training and testing images may also be varied.
We had also worked on Random Forest classifier but unfortunately we did not get
satisfactory results as it was very new to us. We got the classification accuracy near 46%. In
future the performance of Random Forest for iris classification can be improved.
Another extension to the system would be to interface it to an iris acquisition camera.
Now rather than having a fixed set of iris images from a database, a frame grabber can be
used to capture a number of images, possibility improving the recognition rate. An
optimization whose feasibility could be examined with use of an acquisition camera would
be the use of both eyes to improve the recognition rate. In this case, two templates would be
created for each individual, one for the left eye and one for the right eye. This configuration
would only accept an individual if both eyes match to corresponding templates stored in the
database. The recognition rates produced for this optimization would need to be balanced
with the increased imaging difficultly, and inconvenience to the user.
CASIA V.1 was our benchmark database. There are many other available databases
exist e.g. CASIA V.3, UBIRIS, LEI etc. In future we will work on these databases also and
their results will be compared.
REFERENCES
[1]J.Daugman, “High confidence visual recognition of persons by a test of statistical
independence”, IEEE Trans.On Pattern Anal. Machine Intell.vol. 15, pp. 1148-1161, 1993.
[2] J. Daugman,“ Statistical richness of visual phase information: update on recognizing
persons on iris patterns”, Internet. J. On comp. vision, vol.45, pp. 25-38, 2001
[3]J.Daugman, “Demodulation by complex valued wavelets for stochastic pattern
recognition”, Internat. J. on Wavelets, Multi-Res.And Info.Processing, vol. 1, no. 1, pp. 1-17, 2003.
60
[4] W. Boles and B. Boashash,“ A human identification technique using images of the iris
and wavelet transform”, IEEE trans. On Signal Processing, vol. 46, pp. 1185-1188, 1998.
[5] C. Sanchez-Avila and R. Sanchez-Reillo, “ Iris-based biometric recognition using dyadic
wavelet transform”, IEEE Aerosp. Electron. Syst. Mag., vol. 17, pp. 3-6, 2002.
[6] K. Bae, S. Noh, and J. Kim, “ Iris feature extraction using independent component
analysis”, Proc. Of 4th Internat. Conf. On Audio and Video-Based Biometric Person Authentication,
pp. 838-844, 2003.
[7] L. Ma, Y. Wang, and T. Tan, “ Iris recognition based on multichannel Gabor filtering”,
Asian Conf. On Comp. Vision, vol. 1, pp. 279-283, 2002.
[8]L.Ma, Y. Wang, and T. Tan, “Iris recognition using circular symmetric filters”, Internat.
Conf. On Pattern Recog. vol.II, pp. 414-417, 2002.
[9] L. Ma, T. Tan, Y. Wang, and D.Zhang,“ Person identification based on iris texture
analysis”, IEEE Trans. On Pattern Anal. Machine Intell, vol. 25, pp. 1519-1533, 2003
[10] C. Park, J. Lee, M. Smith, and K. Park, “ Iris-based personal authentication using a
normalized directional energy feature”, Internat.Conf.On Audio- and video-Based Biometric
Person Authentication, pp.224 -232, 2003.
[11] R. Wildes, “Iris recognition: an emerging biometric technology”, Proc. Of IEEE, vol.
85, pp. 1348-1363, 1997.
[12] Y. Zhu, T.Tan, and Y. Wang, “Biometricpersonal identification on iris
patterns”,Internat.Conf. On pattern Recog. vol. II, pp.805-808, 2000.
[13] L. Ma, “Personal identification based on iris Recognition,”PhD dissertation, Inst. Of
Automation, Chinese Academy of Sciences, Beijing, 2003.
[14] L. Ma, T. Tan, Y. Wang, and D. Zhang, “Efficient iris recognition by characterizing key
local variations”, IEEE Trans. On Image Processing, vol. 13, pp. 739-750, 2004.
[15]Donald M. Monro, SoumyadipRakshit, and Dexin Zhang, “DCT-Based iris
recognition,”IEEE Trans.On Pattern Anal.and Mach. Intl, VOL. 29, No. 4, April 2007.
[16] Fernando Alonso-Fernandez, Pedro Tome-Gonzalez, Virginia Ruiz-Albacete, Javier Ortega-
Garcia, “ Iris recognition based on SIFT features”.
61
[17] Hugo Proenc, “Iris recognition: On the segmentation of degraded images acquired in
the visible wavelength”, IEEE Trans. On Pattern Anal.and Mach Intel., vol. 32, NO.
8,August2010
[19] Lin Zhonghua, and Ma Hongyan, “ Iris recognition method based on the Morlet
Wavelet Transform Real Coefficients,” Proc. Of the 2009 Internat. Symposium On Information
Processing (ISIP’09)Huangshan, P. R. China, August 21-23, 2009, pp. 113-116
[20]R.Wildes, J. Asmuth, G. Green, S. Hsu, R. Kolczynski, J. Matey, and S. McBridge, “A
machine-vision system for iris recognition ,” Mach. Vision App., vol. 9, pp. 1-8, 1996.
[21]Christel-loïc TISSE, Lionel MARTIN, Lionel TORRES , Michel ROBERT, “Person
identification technique using human iris recognition”, Vision Interface Conf., 2002, pp.
294-299.
[22]GauravGupta ,MayankAgarwal,“Iris recognition using Non Filter-based
technique,”Lane Department of Computer Science and Electrical Engineering West Virginia
University, Morgantown, WV 26506.
[23]R.MeenakshiSundaram, Bibhas Chandra Dhara, “Neural network based iris recognition
system using Haralick Features ,”Department of Information Technology Jadavpur University
Kolkata, India.
[24]Aly I. Desoky, Hesham A. Ali, Nahla B. Abdel-Hamid,“Enhancing iris recognition system
performance,”Dept of comp. sciences Faculty of Eng., Mansoura University Mansoura, Egypt.
[25]NicolaiePopescu-Bodorin, “Exploring new directions in iris recognition,”
11thInternat.Symposium on Symbolic and Numeric Algorithms for Scientific Computing, 2009.
[26]KhinSintSintKyaw, “Iris recognition system using statistical features for biometric
identification,”Internat. Conf.On Elect. Comp. Tech., 2009.
62
[27]Zhaofeng He, Tieniu Tan, Zhenan Sun, and XianchaoQiu, “ Toward accurate and fast iris
segmentation for iris biometrics,” IEEE Trans. On Pattern Anal.And Mach. Intel., vol. 31, no. 9,
September 2009.
[28] A.T. Kahlil, F.E.M. Abou-Chadi, “Generation of Iris Codes Using 1D Log-Gabor Filter,”Electronics and Communications Engineering Department MET Higher Institute of Engineering and Technology Mansoura-Egypt.
[29]D. Field, "Relations between the statistics of natural images and the response properties of cortical cells",American. J. Of the Optic. Socit., vol. 4, no. 12, pp. 2379-2394, 1987.336
[30]J. Daugman. “How iris recognition works”, Proc.Of 2002 Internat.Conf. On Image
Processing, vol. 1, 2002
[31]R. Wildes. “Iris recognition: an emerging biometric technology”, Proc. Of the IEEE, vol. 85, no. 9, 1997.
[32]P. Gupta, H. Mehrotra, A. Rattani, A. Chatterjee, A.K. Kaushik, “Iris recognition using
corner detection”,Indian Institute of Technology Kanpur, India, Department of Information
Technology, New Delhi, India.
[33]RamyZewail, Ahmed Seif, NadderHamdy, MagdySaeb,“Iris identification based on Log-Gabor Filtering”, Arab Academy for Science, Technology & Maritime TransportCollege of Engineering,
[34]ShailjaShukla, BhawnaChouhan,“Iris recognition system using canny edgedetection for biometric identification”,Internat. J. Of Engg. Science and Tech. (IJEST) vol. 3
No. 1 Jan 2011
[35] “http://www.csie.ntu.edu.tw/~cjlin/libsvm”.
63