COPYRIGHT AND CITATION CONSIDERATIONS FOR THIS THESIS/ DISSERTATION

o Attribution — You must give appropriate credit, provide a link to the license, and indicate if changes were made. You may do so in any reasonable manner, but not in any way that suggests the licensor endorses you or your use.

o NonCommercial — You may not use the material for commercial purposes.

o ShareAlike — If you remix, transform, or build upon the material, you must distribute your contributions under the same license as the original.

How to cite this thesis

Surname, Initial(s). (2012) Title of the thesis or dissertation. PhD. (Chemistry)/ M.Sc. (Physics)/ M.A. (Philosophy)/M.Com. (Finance) etc. [Unpublished]: University of Johannesburg. Retrieved from: https://ujcontent.uj.ac.za/vital/access/manager/Index?site_name=Research%20Output (Accessed: Date). Reconnaissance and Assessment of Iris Features towards Human Iris Classification

By: GUGULETHU. P MABUZA-HOCQUET

A thesis submitted in fulfilment of the requirements for the degree

of

Doctor Philosophiae

in Electrical and Electronic Engineering

Faculty of Engineering and the Built Environment UNIVERSITYOF JOHANNESBURG

Supervisor: Co-supervisor: Prof. Fulufhelo. V Prof. Tshilidzi

NELWAMONDO MARWALA

June 15, 2018 Declaration of authorship

I, GUGULETHU. P MABUZA-HOCQUET, declare that this thesis titled, “Reconnaissance and Assessment of Iris Features towards Human Iris Classification” and the work presented in it is my own. I confirm that:

• This work has not been previously presented in any identical or similar form to any other South African or foreign examination board.

• This work was done mainly while in candidature for a research degree at the University of Johannesburg.

• This work has been produced without prohibited assistance of third parties, with all sources of help acknowledged.

• I have clearly attributed where I have consulted the published work of others.

• Where I have quoted from the work of others, the source is given. With the exception of such quotations, this thesis is entirely my own work.

Signature:

GUGULETHU PHUMZILE MABUZA-HOCQUET Pretoria, South Africa June 15, 2018

i To my beloved family.

ii Acknowledgements

I thank God the creator who has granted me the ability, strength and perseverance to see me through this life changing experience. I also extend my sincere gratitude to my supervisor Professor Fulufhelo Nelwamondo and co-supervisor Professor Tshilidzi Marwala for the guidance, patience and believing in me throughout this challenging journey. I also thank the Council for Scientific and Industrial Research (CSIR) for the financial support and the ethical clearance to collect the iris database from its employees that has served as the pillar of this research work.

iii Abstract

The use of counterfeit documents, human trafficking, identity theft, terrorist attacks and cyber-attacks are amongst many of the challenging problems and threats that the whole world is exposed to. Today’s era of advancing technologies means that more people and more devices are connected, and can easily communicate to and with each other via the Internet. Although global communication is a necessity, it also entails the risk of exposing personal or highly classified data to any sort of malicious exploitation. The need to increase security has seen the use of biometrics as a necessary alternative. The growing role of biometric methods have resulted to countries such as India, the United Emirates, Japan, etc. to implement biometric systems for applications in national ID cards, border security, immigration control, and law enforcement, retail stores, banks and government facilities.

Amongst the various biometric modalities, the human iris is regarded as the most accurate, and as such has drawn a lot of attention and gained momentum for over a decade due to the uniqueness, reliability and stability of iris features over a person’s lifetime, as well as the high accuracy achieved for authentication, and ease of image acquisition. A typical iris recognition system (IRS) consists of four modules namely iris segmentation, normalisation, feature extraction and template matching. Each module has automated traditional algorithms that have been successfully used solely for the purpose of uniquely identifying and verifying a person within a large database of enrolled individuals. The drawback of the classical iris segmentation algorithm for instance, is that is assumes that the pupil and iris boundaries are concentric circles, that is, they share the same center, which is not generally the case. The normalisation stage uses the rubber sheet model to transform the segmented iris from a Cartesian plane to polar coordinates to cater for

iv image variances. This method changes the geometrical structure and arrangement of the iris patterns and therefore, the extracted features cannot be traced back to the original image.

Within the existing large body of research, accurate iris segmentation from non-ideal eye images acquired in an uncontrolled room or environment is still a persistent challenge. Furthermore, a long standing argument in iris biometrics has been that the human iris has no genetic penetrance or relation, and as such, cannot be used to determine soft biometric attributes such as gender and ethnicity. Currently, research in iris biometrics research has shifted to either nullifying or validating this argument by advancing the utilisation of iris patterns and textures in order to investigate the possible prediction and classification of individuals according to gender and ethnicity. The issue of concern however is that so far, the only available studies in literature to investigate this concept have been conducted with the use of only two ethnic groups around the world, and that is, Asian participants from China and Caucasian participants from Europe. This thesis is hence motivated by the lack of investigations within this topic that accommodate other ethnicities around the globe in order to fully support or revoke the long standing argument in this field.

The aim of this research is to first address some of the shortcomings accompanying the use of traditional algorithms, especially for segmenting the iris from eye images acquired using ingenuous imaging devices under unfavourable environments. In order to do this, we acquire eye images at different locations with differing environmental conditions and uncooperative participants from two ethnic groups that have never been investigated or covered in literature and that is, black and white participants from the African continent. The proposed method is mostly motivated by the reason that unlike Asian and European individuals, black Africans have very low contrast between the pupil and iris regions. This makes the boundaries that separate the iris and pupil to be very

v smooth and therefore challenging to localise and achieve accurate segmentation using traditional algorithms. This thesis proposes a segmentation method that uses Bresenham’s algorithm to first obtain the parameters of the pupil and iris boundaries from the eye image. The achieved parameters are used as regularising terms for the Chan-Vese algorithm to achieve fast and accurate pupil-iris localisation and segmentation in a single step. The contribution made by the proposed segmentation approach is the robustness and the ability to segment versatile images from our database achiveing 96.5% accuracy; and also the non-ideal public iris databases, achieving 97.6% accuracy when used in the CASIA database; and 96.3% accuracy with the UBIRIS database. In addition to the proposed segmentation approach, this thesis also proposes a fusion of phase congruency and Harris algorithm as a method to detect and extract corner features found within the morphological arrangement of the iris patterns. The contribution made by the proposed fusion of algorithms is that it offers the solution to manage non ideal images by enhancing and preserving the very soft textures mostly found within the black population; producing a compact feature vector with the exact location of corner features that are not only congruent in phase but are also invariant to illumination and rotation. Another contribution is that the obtained feature vector offers the capability to perform successful matching of corner features between a query and a reference iris template. Results of the proposed approach are tested on the three non-ideal databases and obtain an accurate match rate of 99.9% while producing a feature template of 512 bits that requires low storage space.

As the first of its kind to contribute to existing literature in addressing the lack of existing research, this thesis further proposes to use image processing, computer vision, machine learning and artificial intelligence techniques to investigate, extract and utilise the geometrical formation and discriminative characteristic iris texture features and patterns. This is done to develop a robust method that can categorise human irises into classes of

vi gender and ethnicity from African black and white individuals. In order to achieve this goal, the proposed method uses two separate designs of Gabor filter banks to detect the inherent iris textures at different wavelengths and orientations to cover the whole iris image. The image is convolved with the designed Gabor filters to extract the detected textures. Outputs of magnitude and phase from the Gabor filters are used to compute two single quantities of local energy (LE) and mean amplitude (MA) as texture feature vectors. Experimental results achieved from employing statistical methods are able to demonstrate a clear distinction between the two investigated ethnicities. Further tests conducted using a pool of classifiers achieve an overall correct ethnic classification rate of 96.9% and 95% correct gender classification accuracy. The contribution of the proposed design is that it manages to show that both LE and MA features can be used separately or as a joint feature vector to concurrently achieve both ethnicity and gender classification. According to the literature study in this thesis, this is the highest accuracy achieved for gender classification.

As a mechanism to further validate the robustness of the overall experimental methods proposed in this thesis; to determine their possible integration to an existing IRS and to test the discoveries made in this thesis, we employ Bayesian networks to model and test the performance of four classifiers. The obtained experimental results show eminent accuracy for the models of black male and black female classifiers as well as white male and white female classifiers. For classification models between black and white ethnic groups, a correct classification probability of 88% and 82% accuracy is achieved respectively. However, when testing the model for gender classification from the mixed ethnic groups, results show that the proposed concept and the overall contribution of this research holds true, and that is; it is agreeable and easier to first classify an individual according to their ethnicity, so that the search for gender can be specifically conducted within the individual’s ethnic belonging.

vii Contents

Declaration of authorship i

Acknowledgements iii

Abstract iv

1 Introduction 1

1.1 Biometrics and biometric systems ...... 1 1.2 The iris as an organ and biometric trait ...... 4 1.2.1 The principle of an iris recognition system ...... 6 1.3 Problem description ...... 8 1.4 Research goal and objectives ...... 10 1.4.1 Research questions ...... 11 1.5 Research contributions ...... 12 1.6 Delimitations, limitations and assumptions ...... 13 1.7 Thesis layout ...... 14

2 Literature Review 16

2.1 Iris biometrics history ...... 16 2.2 Stages of an iris recognition system ...... 17 2.2.1 Eye image acquisition and datasets ...... 18 2.2.2 Iris segmentation stage ...... 20 2.2.3 Iris normalisation module ...... 24 2.2.4 Iris feature extraction module ...... 27 2.2.5 Iris template matching ...... 30 2.3 A survey on iris segmentation methods ...... 32 2.3.1 Segmentation techniques based on the integro differential operator (IDO) ...... 33

viii 2.3.2 Segmentation techniques based on the (HT) ...... 36 2.3.3 Segmentation techniques based on hybrid methods . 39 2.4 Problematic encounters ...... 42 2.5 Summary of conclusions ...... 43

3 Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images 46

3.1 Introduction ...... 47 3.2 Background on active contour models ...... 48 3.3 Related work on active contour models ...... 50 3.3.1 The model of active contours without edges . . . . . 55 3.3.2 Related work on iris segmentation using the Chan- Vese algorithm ...... 58 3.3.3 Summary on related work ...... 61 3.4 Proposed experimental approach ...... 62 3.4.1 Eye image acquisition ...... 64 3.4.2 Reflection noise removal approach ...... 65 3.4.3 Iris segmentation ...... 68 3.4.4 Results analysis ...... 74 3.4.5 Brief summary on proposed iris segmentation method 79 3.4.6 Feature detection and extraction with phase congruency 81 3.4.7 A brief review on phase congruency ...... 82 3.4.8 Phase congruency in iris recognition ...... 84 3.5 Proposed experimental approach ...... 86 3.5.1 Fusing phase congruency and Harris algorithm . . . 86 3.5.2 Concluding summary ...... 96

4 Ethnicity distinction and classification from iris images 98

4.1 Introduction ...... 98 4.2 Related work on ethnicity prediction and classification . . . 101 4.3 Summary of conclusions on related work ...... 107 4.4 Proposed experimental approach on ethnic classification . . 108

ix 4.4.1 Gabor filters for texture extraction ...... 109 4.4.2 Proposed design for iris texture extraction ...... 111 4.4.3 Experimental results and analysis ...... 114 4.4.4 Ethnic classification from extracted Gabor features . . 119 4.5 Concluding summary ...... 124

5 Gender prediction and classification from iris images 125

5.1 Introduction ...... 125 5.2 Literature on gender prediction and classification from iris images ...... 127 5.3 Proposed experimental approach on gender classification . . 129 5.3.1 Results and analysis ...... 130 5.4 Concluding summary ...... 134

6 Assessment of experimental results using Bayesian networks 135

6.1 An overview on Bayesian networks ...... 135 6.2 Problem formulation ...... 136 6.2.1 Model design ...... 137 6.3 Achieved experimental results and analysis ...... 138 6.4 Concluding summary ...... 146

7 Conclusions on proposed methods 148

7.1 Summary of findings and contributions ...... 148 7.2 Encountered challenges during study ...... 153 7.3 Recommendations for future work ...... 154

x List of Figures

1.1 Human iris anatomy showing cross sectional view and front view ...... 4 1.2 Typical iris recognition system ...... 7

2.1 Pupil and iris boundary localisation with Daugman’s approach 22 2.2 Failed segmentation using traditional segmentation method 22 2.3 Wildes et et al. Hough Transform ...... 23 2.4 Daugman’s rubber sheet model ...... 25 2.5 Normalised iris image ...... 26 2.6 Daugman’s phase demodulation process ...... 28 2.7 Average square shrinking approach ...... 34 2.8 Center tracing from image domain to Hough domain . . . . 38

3.1 Iris center detection: (a) Input Image. (b) Pre-processed image. (c) Binary image. (d) Detected iris center ...... 60 3.2 (a) Pupil boundary detection. (b) Iris boundary detection . . 61 3.3 Flow diagram of the proposed methodology...... 64 3.4 VistaEY2 imaging device ...... 64 3.5 Noise removal:(a) Input grey scale Image. (b) Histogram of grey image. (c) Dilated binary image. (d) Reflections removed from independent database...... 66 3.6 (a) Input grey scale image. (b) Histogram of grey image. (c) Dilated binary image. (d) Reflections removed from CASIA database...... 67 3.7 (a) Input grey scale Image. (b) Histogram of grey image. (c) Dilated binary image. (d) Reflections removed from UBIRIS database...... 68

xi 3.8 Detection of iris and pupil diameters...... 69 3.9 Pupil localisation: (a) Input image. (b) Pupil initial contour. (c) Localised pupil boundary. (d) Pupil mask...... 70 3.10 Iris localisation: (a) Input grey scale image. (b) Iris initial contour. (c) Localised iris boundary. (d) Iris mask...... 71 3.11 Example of a segmented iris...... 71 3.12 (a) Input grey scale image. (b) Pupil initial contour. (c) Localised pupil boundary. (d) Pupil mask...... 72 3.13 (a) Input image. (b) Iris initial contour. (c) Localised iris boundary. (d) Iris mask...... 72 3.14 Example of the CASIA segmented iris...... 73 3.15 (a) Input grey scale image. (b) Pupil initial contour. (c) Localised pupil boundary. (d) Pupil mask...... 73 3.16 (a) Input image. (b) Iris initial contour. (c) Localised iris boundary. (d) Iris mask...... 74 3.17 Example of the UBIRIS segmented iris...... 74 3.18 (a) Input image. (b) Pupil initial contour. (c) Localised pupil boundary. (d) Pupil mask...... 75 3.19 (a) Input image. (b) Iris initial contour. (c) Localised iris boundary. (d) Iris mask...... 76 3.20 Example of segmented iris on noisy image...... 76 3.21 (a) Input image. (b) Pupil initial contour. (c) Localised pupil boundary. (d) Pupil mask...... 77 3.22 (a) Input image. (b) Iris initial contour. (c) Localised iris boundary. (d) Iris mask...... 77 3.23 Example of segmented iris on CASIA noisy image...... 78 3.24 (a) Input image. (b) Pupil initial contour. (c) Localised pupil boundary. (d) Pupil mask...... 78 3.25 (a) Input image. (b) Iris initial contour. (c) Localised iris boundary. (d) Iris mask...... 79 3.26 Example of segmented iris on UBIRIS noisy image...... 79

xii 3.27 (a) Original input image. (b) Processed image with noise removed...... 81 3.28 Samples of self-acquired images...... 87 3.29 Iris enhancement: (a) Input segmented iris. (b) Enhanced iris. (c) Connected iris features...... 88 3.30 (a) Phase congruency image. (b) Histogram of phase congruency image...... 89 3.31 (a) Enhanced PC image. (b) Histogram of enhanced PC image. 89 3.32 (a) Non-maximum suppresion image. (b) Hysteresis threshold image. (c) Histogram of hysteresis image...... 90 3.33 (a) Skeletonised image. (b) Complement of skeletonised image. (c) Phase congruency feature vector image...... 91 3.34 (a) Harris corner features on PC image. (b) Harris corner features on skeletonised image. (c) Histogram of Harris features...... 92 3.35 Corner points match between reference and query image . . 93 3.36 (a) Input segmented iris. (b) Enhanced iris. (c) Connected iris features...... 93 3.37 (a) Phase congruency image. (b) Histogram of phase congruency image...... 94 3.38 (a) Enhanced PC image. (b) Histogram of enhanced PC image. 94 3.39 (a) Non-maximum suppresion image. (b) Hysteresis threshold image. (c) Histogram of hysteresis image...... 94 3.40 (a) Skeletonised image. (b) Complement of skeletonised image. (c) Phase congruency feature vector image...... 95 3.41 (a) Harris corner features on PC image. (b) Harris corner features on skeletonised image. (c) Histogram of Harris features...... 95 3.42 Failed corner points match between UBIRIS reference and self-acquired query image ...... 96

4.1 Gabor envelope at different wavelenghts and orientations. . 112

xiii 4.2 Iris image convolved with Gabor array...... 112 4.3 Gabor magnitude and phase for one iris image...... 113 4.4 Ethnicity distinction using Gabor mean amplitude features. 115 4.5 Ethnicity distinction on larger group...... 115 4.6 Mean amplitude and local enegy features showing no distinction...... 116 4.7 Mean amplitude and local enegy features showing no distinction...... 116 4.8 Iris image convolved with improved design of Gabor array. 117 4.9 Mean amplitude and local energy features with improved design...... 117 4.10 Mean amplitude and local energy features with improved design...... 118 4.11 Ethnic distinction with improved Gabor design...... 118 4.12 Ethnic distinction between all black and white participants. . 119

5.1 Gender distinction between black males and black females. . 130 5.2 Gender distinction between white males and white females. 131 5.3 All males vs females confusion matrix...... 133 5.4 All males vs females confusion matrix with no validation. . 133

6.1 Graphical structure of model design...... 137 6.2 Original training data and synthetic data from model distribution for black males vs black females...... 139 6.3 Black male and black female classifier performance...... 140 6.4 Confusion matrix of black males and black female classifiers. 140 6.5 Original training data and synthetic data from model distribution for white females vs white males...... 141 6.6 White female and white male classifier performance. . . . . 141 6.7 Confusion matrix of white females and white males classifiers.142 6.8 Original training data and synthetic data from model distribution for blacks vs whites...... 143

xiv 6.9 Classifier performance outputs for black and white ethnic groups...... 143 6.10 Confusion matrix for black and white ethnic groups...... 144 6.11 Original training data and synthetic data from model distribution for all males vs all females...... 145 6.12 Classifier performance outputs from all male and female participants...... 145 6.13 Confusion matrix for all males and females in the database. . 146

xv List of Tables

2.1 Singular and hybrid segmentation techniques proposed in literature ...... 33

3.1 Performance of proposed iris segmentation method . . . . . 80

4.1 Experimental results ...... 107 4.2 Comprehensive comparison ...... 108 4.3 Ethnic classification between black males and white females 121 4.4 Ethnic classification between black females and white males 122 4.5 Ethnic classification between all black and white participants 123

5.1 Male vs female classification ...... 132

xvi List of Algorithms

3.1 Chan-Vese principal algorithm ...... 58 3.2 Proposed iris segmentation, feature extraction and matching algorithm ...... 87

4.1 Gabor texture feature extraction and computation ...... 113

xvii List of Abbreviations

2D 2 Dimensional BNT Bayisian Network CASIA Chinese Academy of Sciences Institue of Automation CBIR Content Based Image Retrival CCD Charge-coupled Device CCR Correct Classification Ratio CCTV Closed-Circuit Television CGF Circular Gabor Filter CHT Circular Hough Transform CLAHE Contrast Limited Adaptive Histogram Equalisation CSIR Council for Scientific and Industrial Research CV Chan-Vese DAG Directed Acyclic Graph DLS Direct Least Square EER Equal Error Rate EM Expectation Maximisation FPR False Positive Ratio GAC Geodesic Active Contours GE Gabo Energy GER Gabo Energy Ratio GVF Gradient Vector Flow HD Hamming Distance HSI Hue Saturation Intensity HT Hough Transform ICE Iris Challenge Evaluation ID Identity Document

xviii IDO Integro Differential Operator IRS Iris Recognition System ISO/IEC International Organisation for Standardisation ISO/IEC International Electrotechnical Commission LE Local Energy LED Light Emitting Diode LLC Locality constraint Linear Coding LoG Laplacian of Gaussian MA Mean Amplitude MICHE Mobile Iris Challenge Evaluation MMU Multimedia University MSE Mean Square Error NICE Noisy Iris Challenge Evaluation NIR Near Infrared PC Phase Congruency PDE Partial Differential Equation PPC Principal Phase Congruency RGB Red Green Blue ROI Region of Interest SIFT Scale Invariant Feature Transform SMO Sequential Minimum Optimisation SVM Support Vector Machine TPR True Positive Ratio UBIRIS University of Beira Interior UPOL University of Paleckecho and Olomopuc WVU West Virginia University

xix “Yes, I might have not invented all the tools in my garage, but I sure did create an even greater and more intelligent machine using all of them.”

Unknown

xx Publications derived from this thesis

[1] G. Mabuza-Hocquet, F. Nelwamondo, T. Marwala. "Robust iris segmentation through parameterization of the Chan-Vese algorithm". In Advanced Computer and Communication Engineering Technology, Springer, 2016. pp. 183-194. Related content appears in Chapter 3. This paper reports on iris segmentation from non-ideal images by parameterizing the Chan-Vese algorithm.

[2] G. Mabuza-Hocquet & F. Nelwamondo. "Fusion of phase congruency and Harris algorithm for extraction of iris corner points". In Artificial Intelligence, Modelling and Simulation (AIMS), 2015 3rd International

Conference on, IEEE, 2015. pp. 315-320. Related content appears in Chapter 3. This paper is the first of its kind to report on fusing the two algorithms to perform iris feature etxraction and matching.

[3] G. Mabuza-Hocquet, F. Nelwamondo, T. Marwala. "Ethnicity Prediction and Classification from Iris Texture Patterns: A Survey on Recent Advances". In Computational Science and Computational Intelligence (CSCI), 2016 International Conference on, IEEE, 2016. pp. 818-823. Related content appears in Chapter 4. This paper details all relevant and recent literature on ethnic prediction and classfication from iris textures.

[4] G. Mabuza-Hocquet, F. Nelwamondo, T. Marwala. "Ethnicity Distinctiveness Through Iris Texture Features Using Gabor Filters". In Asian Conference on Intelligent Information and Database Systems, Springer, 2017. pp. 551-560. Related content appears in Chapter 4. This

xxi paper is the first of its kind to report on ethnic distinction between ethnic groups from the Southern African continent.

[5] G. Mabuza-Hocquet, F. Nelwamondo, T. Marwala. "Ethnicity Distinctiveness Through Iris Texture Features Using Gabor Filters". Vietnam Journal of Computer Science (VJCS), 2017. To appear. Related content appears in Chapter 4. This paper extends on ethnic distinction and presents new added results from a new group of subjects than the ones tetsted above.

[6] G. Mabuza-Hocquet, F. Nelwamondo, T. Marwala. "Gender Distinctiveness and Classification based on Ethincity from iris texture features and Bayesian networks." Current Science Journal, 2017. Under submission. Related content appears in Chapter 4, 5 and 6. This paper reports on gender prediction and classfication based on ethnicity from iris texture features.

xxii Chapter 1

Introduction

1.1 Biometrics and biometric systems

Selecting and deciding upon a suitable and relevant user authentication method is an important task in designing a security system. Over the years, security system designers have employed and relied on traditional approaches to user authentication such as access cards, ID cards and tokens for access control to premises, username and password for accessing devices, software and securing sensitive information.

Today’s advancing technologies have seen more devices connected to the Internet. Authentication methods such as passwords for instance, are normally left to the end users’ responsibility hence exposing them to vulnerabilities of being easily forgotten, not updated and being easily predictable. On the other hand authentication methods such as ID cards, access cards and tokens can easily be stolen, lost or duplicated.

Although end users may be constantly guided and advised to strengthen the security of pass phrases, hackers are also constantly deriving advanced methods to crack passwords with intentions to either illegally access sensitive information or to cause malicious damage to data. This reality has led to the development of biometric solutions as a means to safe guard and protect valuable systems.

1 Chapter 1. Introduction

Biometrics are simply defined as the science of recognising a person based on acquired and measured physical or behavioral characteristics [1,2]. The term is derived from the Greek word bio meaning "life" and metrics meaning "to measure" [3]. A person’s fingerprint and palm print for example, qualify as physical characteristics, while keystroke and signature are considered as behavioral characteristics. A biometric system is either an automated or semi-automated system that uses the measured attributes to uniquely recognise an individual.

Generally, a biometric system has two modes of operations under which individual recognition is established; the identification system and verification system. For identification, the system seeks to identify an unknown presented biometric data or an unknown person by asking "who presented this biometric data?", or "who is this person?". In this case, the presented biometric data will be checked against all the available biometric data within the database. Here, there are four possible outcomes [4]:

(i) A true positive occurs when the system correctly matches and reports that an unknown presented biometric data is correctly matched to a trait in the database.

(ii) A false positive occurs when the system says that an unknown presented biometric data matches a particular person in the database and the match is in fact, incorrect.

(iii) A true negative occurs when the system correctly reports that the presented biometric data does not match any of the biometric profiles stored within the database.

(iv) A false negative occurs when the system says that the presented biometric data does not match any of the biometric profiles within the database, whereas the profile does actually belong to someone within the database.

2 Chapter 1. Introduction

This makes the identification system to be described as a 1:N system of matching, where N is all the biometric data contained within the database.

For verification, the system aims to verify an identity claimed by a specific individual, by seeking to answer the question "is this person who he/she claims to be?". In this case, the system will check the presented biometric data against an existing biometric profile that is linked to the claimed individual within the database. Similarly, there are also four possible outcomes for the verification system:

(i) A true accept takes place when the system accepts, or verifies, a claimed identity, and that claim is in fact true.

(ii) A false accept occurs when the system accepts a claimed identity, but the claim is in fact not true.

(iii) A true reject occurs when the system rejects a claimed identity and that claim is indeed false.

(iv) A false reject happens when the system denies a claimed identity, but the claim is actually true.

In this manner, the verification system is described as a 1:1 matching system [1–3].

The use of biometric methods for individual recognition has shown an incline over the past few decades due to the levels of identity security they provide. Various forms of biometric modalities include speech, palm prints, gait, vascular pattern recognition, retina, fingerprint, face, hand geometry and iris being the most recently widely modalities for access control [5]. From all the biometric modalities that have been investigated in literature, only a few have explored the possibility of determining attributes such as the gender and ethnicity of an individual.

3 Chapter 1. Introduction

The rest of this chapter is arranged as follows: Section 1.2 presents an overview of the human iris both as an organ and a biometric trait and how it is used for individual recognition; Section 1.3 describes the problems and challenges in iris recognition that this thesis aims to address; Section 1.4 presents the goals and objectives of the research work in this thesis; Section 1.5 presents the contributions made by this thesis; Section 1.6 presents the delimitations, limitations and assumptions made in this thesis; and Section 1.7 presents the thesis layout.

1.2 The iris as an organ and biometric trait

Amongst the various modalities, the iris is regarded as the most reliable and accurate biometric modality [2, 3, 6–8]. The iris is the thin coloured ring which lies between the cornea and the aqueous humour but in front of the lens on the human eye. Within the iris lies various physical biometric characteristics such as the pigment related crypts, radial furrows, contraction furrows, pigments spots or freckles. The iris encloses a circular black aperture called the pupil. It is surrounded by the white sclera and basically the most protected and only internal organ of the eye that can be seen from outside [9, 10]. Figure 1.1 represents the anatomy of the human iris showing some of the described components, with left and right images depicting the cross sectional view and front view respectively.

FIGURE 1.1: Human iris anatomy showing cross sectional view and front view

4 Chapter 1. Introduction

The formation of the iris begins during the third month of gestation, with its structure completed by the eighth month [7, 11, 12]. The colour of the iris is determined by the pigment and density of a soft tissue called the stroma, that continues to develop up to the first year after birth. Stromal pigmentation commonly ranges between brown, blue, hazel and green; with brown and blue irises depicting high and low quantities of pigmentation respectively [13].

As a biometric trait, the human iris has drawn a lot of attention and gained momentum for over a decade due to the physiognomy of its features. The uniqueness, reliability, stability of iris features as well as the high accuracy achieved for authentication, and ease of image acquisition, are within the advantages that meet the seven desirable characteristics that any biometric trait or attribute must have, in order to be considered suitable for a biometric system [2], and they are:

(i) Universality or commonality meaning that every individual or almost all the relevant population must possess the biometric characteristic.

(ii) Uniqueness or distinctiveness: the biometric trait must be unique and provide a clear distinction between any two individuals.

(iii) Permanence or robustness: the biometric attribute must remain stable and not change due to ecological changes, injuries or diseases over a person’s lifetime.

(iv) Collectability: the trait must be collectable and quantifiable. Also, capturing or gathering of the trait must be reasonably simple and non-invasive or forceful.

(v) Performance: this characteristic focuses on the speed and accuracy of the biometric system which is often a trade-off. Therefore, advanced technologies are ideal to enhance both requirements to better the performance of the system.

5 Chapter 1. Introduction

(vi) Acceptability: the larger society should willingly be receptive to the trait and be able and comfortable enough to provide the trait even across cultural and religious beliefs which can also pose hindrances.

(vii) Circumvention: the biometric trait should not be easily vulnerable to spoofing. This can be achieved through the utilisation of sophisticated technologies and sensors. For instance in iris recognition, factors such as liveness detection should be a main consideration [2].

On its own accord, the iris contains unique features that are stable over a person’s lifetime [7,8, 12, 14]. Furthermore, the measurable physical feature of the iris does not alter due to any ecological effects, [12]. Daugman [11] further describes iris features as a trabecular meshwork of connected tissue and texture, such that not even the left and right eye of the same person contains the same features. Iris features are known to contain approximately 266 degrees of freedom or statistical distinguishing features out of which 173 can be quantified and computed to be used for personal identification [15, 16].

Essentially due to the location of the iris as an organ, the stable, unique and intricate morphology presented by its patterns, textures and features for each person for each eye and even for monozygotic twins, make it attractive for use as a biometric. Furthermore, the non-intrusive nature accompanying eye image acquisition, as well as the high accuracy achieved during identification and verification technology has earmarked research in iris biometrics to gain more attention and momentum in over a decade.

1.2.1 The principle of an iris recognition system

A typical iris recognition system (IRS) comprises of four traditional stages or processes that prevail after image acquisition as presented in Figure 1.2. Each process within the system uses fully automated algorithms with no

6 Chapter 1. Introduction requirement for human intervention. Researchers [2, 5–7] describe the stages as:

(i) Iris segmentation; to separate the iris from the rest of the eye.

(ii) Normalisation; which maps the segmented iris to size invariant coordinates to correct for image size, illumination variations and pupil dilations.

(iii) Feature extraction is achieved by demodulating the normalised iris to extract phase information using 2D Gabor wavelets. This stage produces a unique feature template of size 1024 bits, referred to as an IrisCode, which is then stored within a database for each individual.

(iv) Lastly, feature comparison or matching, happens when a user attempts to be identified or verified by the system. Here, the Hamming distance (HD) is used to match a reference iris template to a query image. The failure of the test of statistical independence is the key to iris recognition. This means that whenever the IrisCodes for two different eyes are compared, the test will certainly be passed. However, whenever an IrisCode is compared with another version of itself, the test will be uniquely failed [7, 8].

FIGURE 1.2: Typical iris recognition system

7 Chapter 1. Introduction

1.3 Problem description

Traditionally, the basic requirement of every biometric system is for subjects to be enrolled first, before any form of identification or verification can take place. This simply means, a willing individual’s biometric data is acquired by means of various sensors depending on the operational modality, and stored as a template within a central database. It is only if a person is enrolled that a successful match will be achieved by the biometric system, otherwise a "non-match" status will be returned by the system.

In iris biometrics, the previous body of research [6–8, 11, 12, 14, 15] has been focused on using eye images that have been acquired with advanced imaging devices, from cooperative participants and under highly controlled conditions in order to develop classical algorithms that achieve accurate segmentation of the iris. This is done to further extract local iris features and to generate a feature template or IrisCode, solely for the purpose of uniquely identifying and verifying a person within a large database of enrolled individuals. The challenge however is that in real life scenarios where standard imaging devices are used, subject cooperation, controlled acquisition conditions and illumination are not always possible; as a result the use of traditional algorithms tend to be unsuccessful for such images [17–20].

During enrollment, other forms of demographic data such as an individual’s ethnic belonging, age and gender depend on the information willingly provided by the enrolling subjects. Even if falsified information is presented by the enrolling individual, the administrator will have no choice but to trust that information, as there are no biometric nor automated methods in place to confirm or verify the provided information. Although automated, the iris recognition system (IRS) has a limiting restriction in that it has been designed and developed with the

8 Chapter 1. Introduction capability to only perform matching of enrolled individuals; and to only return two types of errors which are false accept and false reject. This means it is compulsory for an individual to be enrolled in the IRS before any form of recognition can occur. In this case the next question then becomes: what if a person has not been enrolled and yet the system still needs to determine some form of their identity?.

Demographic data such as ethnicity, gender and age are all forms of attributes collectively referred to as "Soft Biometrics". The significance and advantage of exploiting these attributes to automatically classify individuals according to gender and ethnicity, is that they:

(i) are readily available from acquired iris images.

(ii) can also be extracted, measured and analysed through mathematical methods.

(iii) provide metadata that can be resourcefully linked to an individual simply through eye image scanning and without enrollment; and

(iv) can be duly integrated to the existing IRS, as a pro-active verification system to the information presented by user during enrollment.

Today [21–26], focus in iris biometrics research has advanced to the utilisation of iris patterns and textures in order to investigate the possible prediction and classification of individuals’ soft biometrics which include gender and ethnicity. The reason behind the advancement of research towards this direction of inquiry is that, the uniqueness of iris patterns amongst different individuals has always been acclaimed, however, this approval also came with a long standing notion that "iris features are neither related to nor dependent on human genetics” [6,7, 11, 14].

Another challenge in this regard is that so far, very little experimental work has been conducted to scrutinise this supposition. Unlike the "familiar" local features used for recognition purposes, there is still a lack

9 Chapter 1. Introduction of precision regarding the types of iris texture features that are particularly suitable to achieve ethnic and gender distinction; also, robust texture feature extraction algorithms that are computationally inexpensive have not yet been fully established. Additionally, all the research work that has been done today, has been focused on investigating the ethnicity of Asian and European participants only. Another reported reason is that the topic of achieving automated systems that detect and classify soft attributes from iris images, is still the most underdeveloped and yet crucial problems of computer vision in the research space [21, 24, 27]

1.4 Research goal and objectives

Motivated by the extant challenges, the goal of this thesis is to develop a method that can determine the distinction and also classify individuals according to ethnicity and gender from iris images acquired under non-constrained and differing environments between two racial groups namely, black and white males and females from the African continent. By performing directed experiments in order to propose solutions to the stipulated challenges, the aim is to:

(i) study and assess the performance of traditional iris segmentation algorithms when applied to non ideal iris images that are not centered around the camera’s focal point; and have been acquired under differing environmental conditions with varying illumination.

(ii) investigate, propose and implement an alternative and robust segmentation method that can address the challenges of dealing with non ideal images; and can accurately process any image quality in real time scenarios.

(iii) Design and implement a model that :

• is non-complex for detecting and extracting iris texture features that can be used to single handedly determine ethnic distinction

10 Chapter 1. Introduction

and also classifies individuals according to their particular ethnicity and gender.

• is computationally efficient to also produce a compact feature vector;

• can be easily integrated to an existing iris recognition system.

(iv) use artificial intelligence techniques to evaluate the probability and accuracy of achieving both ethnic and gender prediction from the extracted features.

In order to achieve the goals and objectives of this thesis, the following research questions have to be answered?

1.4.1 Research questions

(a) Apart from eye images acquired under strictly controlled environments, are traditional algorithms employed in each phase of the IRS vigorous enough to handle other images with inherent burdens such as poor quality, poor focus, reflections and illumination variances?

(b) Which segmentation method can be suitably employed to achieve fast and accurate iris segmentation given any eye image?

(c) Is it possible to successfully match non-ideal images for recognition purposes?

(d) With all the various unique features presented by the iris, which features posses universality and permanence; and can also be extracted using image processing techniques; and effectively utilised to uniquely show ethnic distinction and also achieve ethnic and gender classification of an individual?

(e) Is there a suitable method that can detect and extract reliable texture features even from non ideal iris images?

11 Chapter 1. Introduction

(f) Can the detection of ethnicity and gender use or rely on the same features?

(g) Which classification methods are appropriate to achieve accurate ethnic and gender classification?

(h) Is the proposed model automated and robust for integration to an existing IRS?

1.5 Research contributions

The comprehensive studies and experiments conducted in this thesis are designed to consider real time deployment and make the following contributions to the existing body of research:

(i) A non-complicated and robust method based on active contours; that can simultaneously perform pupil and iris localisation; to achieve faster and accurate iris segmentation from non ideal images having low contrast between the pupil and iris regions as well as very smooth pupil-iris boundaries, (Chapter 3).

(ii) As an alternative to the traditional algorithms, a methodology that detects, extracts and matches corner features within the iris trabercula meshwork and generates a compact feature vector to reduce storage requirements, (Chapter 3).

(iii) A methodology that detects, and extracts specific texture features from the iris to produce a concise feature vector to be effectively integrated to an existing IRS, (Chapter 4).

(iv) A novel approach in iris recognition that determines and demonstrates ethnic distinction between individuals, (Chapter 4).

(v) A novel prospect in iris recognition that it is easier to first determine and classify the ethnicity of an individual, followed by gender classification, (Chapter 4 and 5).

12 Chapter 1. Introduction

(vi) A novel concept in iris recognition demonstrating that gender classification is efficiently achieved from individuals belonging to the same ethnic group, (Chapter 5).

(vii) A novel approach in iris recognition to not only evaluate the accuracy of the used texture features but also the performance and validity of the results achieved from used classifiers for ethnic and gender classification, (Chapter 6).

(viii) A new study in iris recognition that accommodates other ethnic groups other than the ones already investigated in existing literature, (Chapter 3-6).

(ix) An overall methodology that equips a traditional IRS with:

• a fully automated recognition capability, especially for cases where such technologies are already deployed for access control, (Chapter 3-6).

• a proactive biometric data verification mechanism for readily enrolled individuals; also to eliminate system administrators’ dependence on enrolling individual for cases where enrollment is not necessary, such a model can be applied by businesses to analyse consumer trends; in airports for passenger statistics; in government entities to avoid duplicate allocation of grants and also for conducting national census, (Chapter 4-6).

1.6 Delimitations, limitations and assumptions

The focus of this thesis will be centered on the following factors:

(i) Data collection: eye images are self-acquired using the Vista EY2 dual iris and face camera under differing and uncontrolled conditions. The VistaEY2 uses state of the art digital imaging technology and simultaneously captures both eye images that are ISO/IEC 19794-6 compliant at a sensing distance of 38.1 cm to target.

13 Chapter 1. Introduction

(ii) Only non ideal images from publicly available iris databases, that is, CASIA, UBIRIS as well as the self-collected database from consenting participants and age groups of inherently born males and females between two ethnic groups which include African blacks, whites are used in this thesis.

(iii) Collection and testing of algorithms are only performed on eye images assumed to be free of any eye diseases, contact lenses and spectacles.

(iv) Traditional algorithms from some of the IRS stages are used as reference for benchmarking the algorithms proposed in this thesis.

(v) Only the classification of human irises into gender and ethnicity of the two mentioned groups is performed for this thesis.

(vi) All simulations and image processing techniques are performed using Matlab 2016b.

1.7 Thesis layout

The rest of the chapters in this thesis are arranged with an aim to answer the presented research questions. Each chapter will be providing: related work, the challenges and problems facing the presented topic, proposed experimental approach to address the challenges, achieved experimental results and analysis, as well as concluding summaries for each chapter.

• Chapter 2 lays the foundation towards providing some insight to some of the research questions by presenting a literature study on the processes or modules that are involved in a classical iris recognition system. The cornerstone algorithms that have been proposed and utilised in each IRS module are also presented. It also focuses more attention on the most recent work proposed by other researchers regarding the segmentation module as the pillar of an IRS. Challenges and drawbacks faced by proposed methods, which also serve as motivations for this research are also presented.

14 Chapter 1. Introduction

• Chapter 3 provides an answer to the first research question on iris segmentation using active contour models. This is followed by the method proposed to perform iris segmentation, feature extraction and feature matching from non ideal images. Experimental results to deduce answers are also presented.

• Chapter 4 presents the method proposed in this thesis in order to achieve ethnic distinction and classification from iris images; obtained experimental results, accompanied by the analysis are also presented. The goal of this Chapter is to address research questions 2-5 on the topic of ethnicity.

• Chapter 5 is a follow up to Chapter 4, exploring the subject of gender prediction and classification from iris images. This is followed by the presentation of the proposed method, obtained experimental results to achieve gender classification and the analysis and discussions of the results.

• Chapter 6 investigates the robustness of the work done in Chapter 4 and Chapter 5. This is done by modeling the obtained results through using artificial intelligence techniques in order to fulfill research questions 6.

• Chapter 7 concludes the work done in this thesis by discussing recommendations and future work pertaining the experimental results obtained from the investigated Chapters.

15 Chapter 2

Literature Review

This chapter presents the background on literature that is relevant and crucial to the problems and solutions that are being investigated in this research work. Section 2.1 discusses the historical evolution of the iris as a biometric trait and its role leading to the development of the first iris recognition system (IRS); Section 2.2 presents the fundamental stages or modules of an iris recognition system (IRS) and the classical algorithms employed in each module; Section 2.3 is a survey on the recent research work that has been conducted by other researchers, especially focusing on the segmentation module, with an aim of achieving more robust methods to improve the overall performance of an IRS; Section 2.4 discusses the problems pertaining the segmentation module that are still widely encountered in this research field and Section 2.5 is a concluding summary that links the motivation of this work and the contribution made by this research work towards the segmentation module.

2.1 Iris biometrics history

Historical uses of the iris as a biometric date back to 1936 where the first concept of using iris patterns for identification was proposed by an ophthalmologist named Frank Burch [2, 3]. In his work as an opthalmologist, Frank Burch discovered the prospective use of the iris to distinguish between the same person’s left and right eye; as well as one individual from another. The concept that "no two irises are the same" was officially proposed by ophthalmologists Dr. Leanord Flom and Aran Safir

16 Chapter 2. Literature Review

[28]. The official proposal was initially a conceptual design that was not yet implemented but later patented in 1987. It was only later that Dr. J. Daugman [11] studied the findings of Flom and Safir to develop an automated iris recognition algorithm, which was patented in 1994. In his research work, Daugman further discovered that the developed system can be extended for use in other biometric security applications such as biometric identity cards and passports, e-commerce, law enforcement and banking systems.

Due to advancements in technologies and global demands for the need to effectively secure information, the algorithms proposed by Daugman have become widely licensed and are the cornerstone of most iris recognition products [7]. This motivation rapidly led to the collection and establishment of public iris databases such as the Chinese Academy of Sciences Institute of Automation (CASIA), University of Paleckecho and Olomopuc (UPOL) and University of Beira Interior (UBIRIS) for the purposes of continued research that has drawn a lot of attention and gained momentum for over a decade.

2.2 Stages of an iris recognition system

A summarised concept defining the stages of an iris recognition system was presented in Chapter 1. This section presents a fundamental review of:

(i) the main modules and the respective cornerstone classical algorithms that are involved in each stage of a typical IRS, as implemented by two pioneers in iris recognition research, Daugman [7, 11, 12, 29] and Wildes et al.[8, 30, 31]

(ii) the drawbacks that accompany the presented traditional algorithms employed in each module.

(iii) a principal understanding of the IRS automation, and

17 Chapter 2. Literature Review

(iv) a summary of the similarities and differences of the segmentation algorithms employed by the two pioneers.

The stages or modules of a typical iris recognition system include eye image acquisition, segmentation, normalisation, feature extraction and template matching [6–9, 14]. The stages that prevail after image acquisition have traditional algorithms that are used as reference methods for further developments and enhancements of factors such as performance speed, accuracy and computational complexities for easier deployment in real time applications. However, the implementation of traditional algorithms also rests upon the availability of eye images as biometric data that meets the ISO/EC standards of interchangeable format; to allow for interoperability and to circumvent vendor lock-in [32].

2.2.1 Eye image acquisition and datasets

The acquisition of eye images is the first and basic step towards achieving the objectives of an IRS. However, capturing high quality images with good resolution and sharpness while using safe and non-invasive devices, are some of the few challenges of an automated IRS. With the iris being a small object within the human eye, the capturing process is not easy [33]. Imaging devices and acquisition conditions or environments play a vital role, since the acquired images have to meet a standard suitable enough for continued processing.

Daugman [29] uses a Light emitting diode (LED) based point light source combined with a standard video camera to acquire eye images with the iris diameter between 100 and 200 pixels from a distance of 15-46 cm using a 330-mm lens. The acquisition process from willing subjects is strictly controlled and constrained, such that the resulting images are well focused on the camera lens. The positioning of the light source below the operator aims to avoid reflections in the imaged iris. However, Daugman’s design still acquires images with the light source inherently adopted by the iris

18 Chapter 2. Literature Review region where the point of the light source is seen by the eye. The observable light reflection within the iris region becomes a dominating artefact that must be omitted during matching phase to achieve accurate template matching.

Wildes et al.[8, 30] use a diffuse source and matched circular polarisers combined with a low-light level video camera to acquire images with the iris approximately 256 pixels across the diameter, from a distance of 20 cm using an 80-mm lens. Although Wildes et al.’s setup design is more complex than Daugman’s, its advantage is that using the matched polarisers eliminates the specular reflections within the iris region that tend to be introduced by the light source. However, ultimately both systems make use of a light source that is visible to the human eye and can cause harm that may lead to user rejection due to the amount of illumination entering the eye.

For research purposes, the publicly available CASIA-Iris-V3 iris database captured with a self designed circular near-infrared (NIR) LED light source, has a total of 22,051 eye images. Resulting images are 8-bit grey level images acquired from 700 Asian subjects captured under a strictly controlled environment to capture high quality images.

The UPOL database has been captured with TOPCON TRC50IA optical device connected with SONY DXC-950P 3CCD camera. The database has 384 images that are 24-bit true colour (RGB) from European subjects. The challenge with images in this database is that the "segmentation" work has already been "done" through cropping, so only the white sclera and iris are visible.

Contrary to the noise free databases, the UBIRIS database [34], also uses European subjects captured with a Nikon E5700 camera acquired under uncontrolled environments. This database has two versions; UBIRIS v1

19 Chapter 2. Literature Review and v2. The first version consists of 1877 noisy images acquired from 241 non cooperative subjects. The second version (v2) contains over 11 000 images with more practical noise factors captured at "a distance and on the move" [34]. The purpose of this database is to provide a platform for the development of robust algorithms that can withstand the hindrances accompanying its nature.

Wildes [8] reports that NIR illumination is a better light source since it is neither harmless nor intrusive to human eyes and has the advantage of revealing details within the iris structure even from dark or highly pigmented iris images. Ideally, the cooperation of users is also a necessity in order to reduce complexities and the quantity of pre-processing techniques needed [9]. However, the demand for cooperation may affect users’ feelings which may result to the rejection of the technology [2]. Finally, eye images used as biometric data for an automated IRS must meet the ISO/EC standards with 70% usable iris region of interchangeable format [32].

2.2.2 Iris segmentation stage

Given an eye image, the first step is to separate the iris from the rest of eye, a process referred to as segmentation. Automated iris segmentation is the first most crucial, challenging and computationally heavy stage of the recognition system [7, 35]. This is mainly because all other subsequent stages and further processing are highly dependent on the success and accuracy of the segmentation phase. Falsely represented iris data can corrupt the generation of feature templates and thus increase poor recognition rates [36].

With the iris region being small, hindering factors such as illumination variations, eyelids and eyelashes are amongst the few reasons for unsuccessful segmentation. Therefore a robust segmentation method is required to accurately capture the iris region of interest [37]. The

20 Chapter 2. Literature Review pioneering work of Daugman has contributed immensely to the field of iris biometrics such that not only are his methods always used as a point of reference in research, but almost all commercially deployed iris recognition technologies are based on his work [4].

Given an input image I(x, y) at location (x, y), Daugman’s approach starts by approximating the boundary of the pupil (pupillary boundary ) and the boundary of the iris (limbic boundary) as circles with three parameters of a certain radius (r), and center coordinates x0 and y0. The approximated circles are fitted based on an increase in gradient while searching through the image parameter space using the integro-differential operator (IDO):

I ∂ I(x, y) max Gσ(r) ∗ ds (2.1) r,x0,y0 ∂r 2πr r,x0,y0 where the term G(σ) is the Gaussian smoothing function of scale (σ), convolved (*) with the partial derivative of the candidate parameters (r, x0, y0) of the eye image I(x, y) in spatial coordinates, and ds representing the circular arc of integration. Within the parameter space, the operator searches for a circular path where a maximum change in pixel intensities of the three candidate parameters occurs. This is done through the iterative variation of the three parameters coupled with progressive fine tuning of the smoothing function until precise location of the boundaries is achieved. The boundaries of the upper and lower eyelid are also detected by applying the same method however the integration arc is altered from circular to arcuate paths. The result is the detection and localisation of the pupil and iris boundaries, with the segmented iris region shown by Figure 2.1, [38].

The drawbacks of this segmenation method are that it:

(i) rigidly assumes that the iris and pupil have perfect circular boundaries. However, this assumption is not homogenous for all iris images.

21 Chapter 2. Literature Review

FIGURE 2.1: Pupil and iris boundary localisation with Daugman’s approach

(ii) is computationally expensive due to the large parameter search space.

(iii) strictly requires that the input eye image be centred on the camera’s focal view [39], otherwise the segmentation process fails as shown in Figure 2.2, [38].

FIGURE 2.2: Failed segmentation using traditional segmentation method

Within any given image, the parameters of geometric objects such as lines and circles can be determined by the use of the Hough Transform (HT). This algorithm is a standard computer vision technique that can also determine the parameters of annular regions; like that of the pupillary and

22 Chapter 2. Literature Review limbic boundaries.

The work of Wildes [8] uses the Hough Transform (HT) to fit the pupillary and limbic boundaries in two steps. First, the input image is changed to a binary edge map. This is done by calculating the first derivatives of the image intensity values I(x, y) and then thresholding the resulting magnitude to generate a mapping of edges of the input image with ones and zeros, that is; |∇G(x, y) ∗ I(x, y)|, where ∇ ≡ (∂/∂x, ∂/∂y) while:

2 2 1 − (x−x0) +(y−y0) G(x, y) = e 2σ2 (2.2) 2πσ2

G(x, y), is a two dimensional Gaussian function with center (x0, y0) and σ as the standard deviation to smooth the image so that the spatial scale of edges under consideration are chosen. Secondly, the edge points cast votes into the Hough space for circle parameters passing through each edge point. The circle parameters are those that will fit the pupillary and limbic boundaries defined by: 2 2 2 xc + yc − r = 0 (2.3) where xc, yc define the center coordinates, with the maximum point of the Hough space represented by r as the radius of the circle parameters. Therefore, in order to fit the limbic boundary, the derivatives are biased to be selective of vertical edges, while for detecting eyelids, derivatives are biased to be selective of the horizontal direction, as shown in Figure 2.3 [8, 36].

FIGURE 2.3: Wildes et et al. Hough Transform

The reasoning behind the selection of the directions is that eyelids are generally aligned horizontally. Therefore using vertical gradients to fit the

23 Chapter 2. Literature Review limbic boundary, reduces the detection and influence of eyelids with the HT. The artistic factor about employing the edge map and Hough transform is that, for the successful location of the boundaries, not all edge pixels defining the circles are needed. This makes Wildes’ method to be more effective and accurate, since there are less edge points needed to cast votes in the Hough space.

The drawbacks of this segmentation method are that:

(i) the selection of the threshold for generating the edge map might not be similar for all images. This means some critical edges might be missed on some images.

(ii) the computational intensity required by the HT might not be practical for real time applications.

2.2.3 Iris normalisation module

After the iris has been successfully segmented, the succeeding challenge becomes that not only are the segmented iris images inconsistent in size, but also suffer from variations such as illumination, head tilt, pupil dilation and capturing distance from the camera [7,9, 11]. In order to correct for such variations, unwrapping of the segmented iris is performed, a process referred to as normalisation. The normalisation module uses the rubber sheet model also proposed by Daugman [7, 11]. The objective is to fix the dimensions of the segmented iris in order to facilitate a uniform comparison of iris characteristic features between different iris images. Therefore regardless of image size and pupil dilation or constriction, the model assigns a pair of doubly dimensionless coordinates (r, θ) to each point within the segmented image, shown by the model in Figure 2.4. The normalisation process, remaps the raw segmented iris image from the Cartesian plane I(x, y) to a pair of corresponding polar coordinates I(r, θ)

24 Chapter 2. Literature Review

FIGURE 2.4: Daugman’s rubber sheet model by the equation: I(x(r, θ), y(r, θ)) → I(r, θ) (2.4) where r is the radial radius in the interval [0, 1] and θ is the angular quantity in the range [0, 2π]. I(x, y) describes the Cartesian coordinate of the segmented iris and I(r, θ) is the corresponding polar coordinate. The unit of the inner boundary which depicts the border from the pupil to the iris; known as pupillary boundary, is represented by (xp, yp) in the Cartesian coordinate. While the unit of the outer boundary depicting the border between the iris and white sclera, known as the limbic boundary, is represented by (xl, yl). The transformation is then defined by the equation:

x(r, θ) = (1 − r)xp(θ) + rxl(θ) (2.5)

y(r, θ) = (1 − r)yp(θ) + ryl(θ) (2.6)

The resulting normalised iris image becomes a rectangular block with the radial and angular coordinates on the vertical and horizontal axes respectively. Within the rectangular block, the pupil boundary may be on the top of the image with the limbic boundary at the bottom. The 0◦ mark is on the left side of the rectangular block and the 2π mark rests on the right side as shown in Figure 2.4 and Figure 2.5. The change in pupil size causes an elastic or linear stretch of the iris region that appears like a deformation is accounted for by the dimensionless coordinate system. This method does not compensate for rotational inconsistencies; as a result this

25 Chapter 2. Literature Review is accounted for during the succeeding matching phase, where the feature templates are shifted by θ until two iris templates align [36]. It is from the normalised iris that distinguishing features are extracted.

FIGURE 2.5: Normalised iris image

The drawbacks with this method are that:

(i) due to the separation between the 0 and 2π marks being so random, a minor head tilt can affect the angular coordinate.

(ii) it also assumes concentricity between the pupil and iris boundaries which produces a linear stretch of the normalised iris even when the pupil is dilated or constricted [4].

(iii) based on the objective for feature extraction, the transformation introduces changes in the geometrical structure and arrangement of the iris patterns [39].

In order to compensate for image rotation and scaling, Wildes et al.[8, 30] employ image registration. With this approach, a newly acquired image for instance, a query image Ia(x, y), is geometrically warped into alignment with an image that is already in the database Id(x, y), (a reference image) using the mapping function: u(x, y), v(x, y). With the mapping, the intensity values of the query image are made to closely correspond to those points in the reference image, to minimize the function: Z Z 2 (Id(x, y) − Ia(x − u, y − v)) dxdy (2.7) x y

26 Chapter 2. Literature Review while being restricted to only capture a similarity transformation of image with coordinates (x, y) to (x0, y0), described as :

      x0 x x         =   − sR(φ)   (2.8) y0 y y where s is a scaling factor and R(φ) is a matrix represented φ as rotation.

Practically, given a query image (Ia) and a reference image (Id), the recovering of the warping parameters s and φ is achieved via an iterative minimisation procedure [8].

The work of Wildes et al.[8, 30] admits that their system and that of Daugman [7, 11, 29] consumes and needs more processing time. It is further reported that in order for both systems to establish a high correspondence between two images, the eye image acquisition process has to be strictly controlled [8].

2.2.4 Iris feature extraction module

In order to encode a unique feature template to serve as person identifiers, features have to be extracted from the normalised iris. With Daugman’s approach, the normalised iris is demodulated using quadrature 2D Gabor filters in order to compute phase coefficients that will extract coherent and incoherent textures such as those found within the iris [11]. A Gabor filter is a Gaussian kernel modulated by a sine or cosine wave [40]. A 2D Gabor filter over the an image domain (x, y) is represented as:

2 2 2 2 G(x, y) = e−π[(x−x0) /α +(y−y0) /β ]e−2πi[u0(x−x0)+v0(y−y0)] (2.9)

where (x0, y0) specify position in the image, (α, β) specify the effective width and length, and (u0, v0) specify modulation, which has spatial frequency: q 2 2 ω0 = u0 + v0 (2.10)

27 Chapter 2. Literature Review

For iris recognition, the used 2D Gabor filters are defined in the dimensionless polar coordinate by [11]:

2 2 2 2 G(r, θ) = e−iω(θ−θ0)e−(r−r0) /α e−(θ−θ0) /β (2.11)

The Gabor filters result in a complex response of the image projection, where the real part, known as even symmetric; and imaginary part, known as odd symmetric, are respectively specified by the cosine and sine wave modulated by the Gaussian kernel as shown in Figure 2.6, [7].

FIGURE 2.6: Daugman’s phase demodulation process

Amplitude information is not very discriminating, therefore only phase information is used to generate an individual’s IrisCode. The phase information is quantised into four levels for each possible quadrant in the complex plane. The four levels are represented using two bits of data, 1 and 0 in each quadrant. This means each pixel in the normalised iris pattern corresponds to two bits of data in the iris template. Only the phase information allows encoding of discriminating information from the iris, while redundant information such as illumination represented by the amplitude component is discarded. Advantages of using phase information are that phase angles are assigned regardless of how poor the image contrast is. Furthermore, phase encoding prevents different poorly

28 Chapter 2. Literature Review focused irises from being confused with each other [7].

A total of 2,048 bits per individual are calculated for the feature template or IrisCode, and an equal number of masking bits are generated to mask out corrupted regions within the iris. As a result, a compact 256-byte template is created, allowing for storage and comparison of irises.

The system of Wildes et al.[8, 30] encodes iris features by decomposing the iris region using Laplacian of Gaussian (LoG) filters. Applying LoG filters to an image will place an emphasis on regions where a rapid change in pixel intensity occurs. It makes more sense for Wildes et al.[8, 30] to continue with this approach since LoG filters are also often used as edge detectors, following their segmentation approach. For decomposition of the iris region, filters are represented as:

  1 2 2 2 ∇G = − 1 − ρ e−ρ /2σ (2.12) πσ4 2σ2 where σ is the standard deviation of the Gaussian and ρ represents the radial distance of a point from the center of the filter. The output image of the filters is constructed with four different resolution levels to produce its representation as a Laplacian pyramid. Using the Laplacian pyramid to represent the resulting filtered image offers the opportunity to compresses data without any loss of information. This generates a compact iris template on the order of the number of bytes of the original input image.

The similarity in the work of Daugman [7, 11, 29] and Wildes et al.[8, 30] is that both approaches profit from the multi scale structure of the iris to encode iris features that generate a template using bandpass filtering methods.

29 Chapter 2. Literature Review

2.2.5 Iris template matching

Daugman [7, 11] states that "the failure of the test of statistical independence is the key to iris recognition". This means whenever the IrisCodes of two different eyes are compared, the test is guaranteed to be passed. However, whenever the IrisCode of the same eye is compared to another version of itself, the test will be uniquely failed. Here, the Hamming distance (HD); which measures the dissimilarity between two bit patterns, is employed to evaluate whether two patterns are either generated from two different irises or from the same iris.

The test of statistical independence is executed by the Boolean Exclusive-OR operator (XOR) applied to the 2,048 bit phase vectors that encode any two iris patterns (code A; code B). The XOR (⊗) operator serves to spot the disagreement between any corresponding pair of bits. In order to prevent non-iris effects from influencing iris comparisons, the two codes are further masked or (AND’ ed) by their corresponding mask bit vectors (mask A, mask B). The AND (∩) operator ensures that both compared bits are uncorrupted by noise such as eyelashes, eyelids and specular reflections. The norms (|| ||) of both the bit phase vectors and mask bit vectors are then computed to finally get the HD by:

k(codeA ⊗ codeB) ∩ maskA ∩ maskBk HD = (2.13) k(maskA ∩ maskB)k

Since, any given bit in the iris phase code is equally likely to be 1 or 0, the resulting HD is a fractional measure of dissimilarity. Therefore, statistically any two different irises are “guaranteed” to pass the test of independence. However, any two images that fail this test by producing an HD ≤ 0.32 must be from the same iris [7, 11].

Wildes [8] refers to this stage as “ the goodness of match”. The method employs normalised correlation between a query image and a reference

30 Chapter 2. Literature Review image to quantify a match, represented as :

Pn Pm i=1 j=1(p1[i, j] − µ1)(p2[i, j] − µ2) σ = (2.14) nmσ1σ2

where p1 and p2 respectively represent the query and reference images of size nxm, µ1 and σ1 are the respective mean and standard deviation of p1, and µ2 and σ2 are the respective mean and standard deviation of p2. Inasmuch as standard correlation and normalised correlation capture the same type of information. Normalised correlation however, manages to also account for local disparities in image intensity that corrupt standard correlation.

A brief summary of the similarities and differences of the algorithms employed by [7, 11, 12, 29] and Wildes et al.[8, 30, 31] to localise the pupil and iris boundaries in order to perform iris region segmentation is presented below. Similarities of both methods are that:

(i) Only eye images acquired under strict and controlled conditions are used.

(ii) They localise the limbic and pupil boundaries with mathematical methods that only employ perfect circular contours.

(iii) The methods are either gradient or intensity based.

(iv) The first derivatives of image intensity to indicate the location of edges that correspond to the borders of the pupil and iris are used.

(v) The image intensity derivative information is iteratively fine-tuned to achieve the expected configuration of model components.

The notable difference between both methods is in the approach used to search their parameter spaces to fit the contour models to the image information. The method of Wildes et al. starts by detecting edges from an input image in order to account for eyelids and eyelashes, thus adding a preprocessing step before estimating the pupilllary and limbic boundaries

31 Chapter 2. Literature Review to perform segmentation; whereas the method of Daugman directly estimates the pupilllary and limbic boundaries and accounts for eyelashes and eyelids via the same segmentation method.

2.3 A survey on iris segmentation methods

Now that the fundamentals of a typical IRS have been discussed, most of the research that has been conducted over the years has been and is still based on the concepts introduced by Daugman and Wildes. Due to its undoubtable importance, a valuable fraction of research in iris recognition has been dedicated to the segmentation module. This is purely because accurate segmentation is paramount to all subsequent stages and overall performance of any IRS.

Public challenges and events such as the Iris Challenge Evaluation (ICE) conducted and managed by the National Institute of Standards and Technology [41], the Noisy Iris Challenge Evaluation (NICE) [42] and the Mobile Iris Challenge Evaluation (MICHE) [43] have been hosted towards solving this formidable module.

Table 3.1 presents some of the several segmentation methods that use different techniques based on either single or hybrid strategies proposed in literature. There is often a compromise or trade-off where "perfect segmentation", in all influencing scenarios, is concerned. This refers to, factors varying from devices used to acquire eye images, lighting conditions under which images are acquired, the cooperation of subjects or the lack thereof, which often results in off angle images, as well as poor quality of the visible iris region due to eyelid occlusions, are all unideal hindrances, yet realistic and practical issues that are continuously addressed by researchers.

32 Chapter 2. Literature Review

TABLE 2.1: Singular and hybrid segmentation techniques proposed in literature

Authors Segmentation technique

Ma et al.[44] Grey level information, Canny & HT

He et al.[45] Pushing and pulling model

Puhan et al.[46] Fourier spectral density

Luengo-Oroz et al.[47] Mathematical morphology

Li et al.[48] Level set evolution

Daugman [49] Active contour and Fourier series expansion

As one of the main contributions of this thesis, segmentation of non-ideal iris images had to be conducted, as eye images were self-acquired. This section reviews some of the methods recently proposed by other researchers on iris segmentation from non-ideal iris images with an aim of overcoming the presented drawbacks or improving the existing traditional algorithms.

2.3.1 Segmentation techniques based on the integro differential operator (IDO)

The work done by Ren et al.[50] sets out to improve Daugman’s iris localisation algorithm. Their work admits that Daugman’s approach of approximating the pupil and iris boundaries by searching a large parameter space is challenging if the center coordinates and the pupil radius are unknown. Therefore, their approach proposes to improve Daugman’s method by performing pupil and iris boundary localisation in two folds using a coarse to fine strategy. The coarse strategy refers to the rough approximation of parameters and; the fine strategy refers to using the approximated parameters to achieve segmentation precision.

33 Chapter 2. Literature Review

As the first step, the area of the circle formula is used for pupil coarse localisation. This is followed by finding the precise center and radius coordinates of the pupil by employing Daugman’s (IDO). The localisation of the limbic boundary is also done in two steps: (a) The noise introduced by the presence of the upper and lower eyelid covering the iris region is considered. Therefore the outer boundary is coarsely localised using the grey difference on the arcs of the right and left canthus, by choosing the arc to be −45◦ < θ < +45◦ for the right side canthus and 135◦ < θ < 255◦ for the left canthus. A canthus is the outer or inner corner of the eye, where the upper and lower lids meet. (b) By so doing, the center of the limbic boundary is redefined around the pupil’s center to get its precise location. This approach is reported to reduce the search parameter and decrease computational complexities.

The work of Shamsi et al.[51] also aims to enhance Daugman’s approach by using an average square shrinking approach. Their work also states that it is challenging to employ Daugman’s operator to locate circles from all eye images due to varying image sizes, illumination and contrast. Instead of searching for circles through the whole image parameter space, they propose to restrict the search space by applying average square shrinking to the source image, as shown in Figure 2.7.

FIGURE 2.7: Average square shrinking approach

The detection of the iris center is manoeuvred through by using the darkest place within an eye image, the pupil. This is done by using a

34 Chapter 2. Literature Review square of 10x10 dimensions around the center of the source image as the potential iris center (x, y) and a range of 55 to 65 pixels as the potential radius r. From the source image, each square is converted into one pixel in the shrunken image. The square size and the shrinking stages are respectively related to a smoothing factor and a number of shrinking stages which are manually adjusted. This means pixel values inside the square will be averaged in the shrunken image. Therefore, in the last shrunken image, the darkest pixel (x0, y0) is deemed as the pupil center in the range of the smoothing factor. The potential parameters are then applied to the IDO, where the higher value of operator corresponds to the precise center and radius of the iris.

Another research work that is based on Daugman’s method, is the work done by Proença [52]. The work proposes an iris segmentation algorithm that offers reliability and performance speed for non-ideal UBIRIS iris images. Similar to the drawback pointed out by the work of Ren et al.[50] and Shamsi et al.[51], that the challenge with the traditional IDO is the computational time required to localise the center of the pupil, iris and the respective boundaries. The model proposed by Proença [52] is done in three modules which include: the removal of reflections; detection of pupil and iris; and lastly employing the live wire technique to detect eyelids.

In order to remove reflections from images acquired under visible light, the complement of the input image is computed. Complementing a binary image means zeros become ones and ones become zeros. This process is followed by hole filling, where within an image, the holes will be the set of dark pixels surrounded by brighter pixels. After the holes have been filled, the image is complemented again resulting to the original representation.

The second module of pupil and iris detection is performed using the circular Gabor filter (CGF). In order to get a rough or coarse estimate of the pupil center, the coarse to fine strategy is used. This is done by coarsely

35 Chapter 2. Literature Review marking the center of the iris which also includes the pupil region. The CGF is then used to identify the pupil center. For the coarse localisation of the iris, the image is down scaled by 20% from its original size to be further filtered by the CGF. The down scaling is done to speed up the process of filtering while the filtering is done in order for the iris to protrude, where the most highlighted pixels within the filtered image are considered as the pupil region. The boundary of the iris is then detected by the adaptive IDO, with the consideration that the iris center also lies in a set of pixels around the pupil center.

In the last module of eyelid detection, the canthus and the limbic boundary are detected using the Hue, Saturation, Intensity (HSI) model. The boundaries of the eyelids are then localised using the live wire method. The live wire method is an interactive segmentation technique whereby through using simple mouse clicks, a user can quickly and accurately select regions of interest to be extracted [53]. Experimental results show that the proposed method minimises the time required to segment the iris when compared to the traditional IDO. Other segmentation methods that have been newly proposed by various researchers based on the IDO include [54–58].

2.3.2 Segmentation techniques based on the Hough Transform (HT)

Following the use of the HT from Wildes et al.[8, 30], the work done by Nkole et al.[59] proposes to enhance iris segmentation through the utilisation of eight neighbourhood operators to localise the center of the pupil and hence the iris. In order to detect the pupil boundary, a global threshold is applied to the input grey scale thereby converting it to a binary image. This is done to segment the pupil because it has much smaller grey scale values than the other parts of the eye image. Due to this, the thresholded binarised image results to the pupil being black within a white background. The binarised image is then complemented so that the

36 Chapter 2. Literature Review dark colour of the pupil is white and visible as the area of interest. By applying averaging filters, the blurred pupil boundary is smoothed for ease of detection.

The center of the pupil is computed first by scanning horizontally through the complemented image from left to right using the 8-neighbourhood. Here, the first and last 8-neighbourhood whose value is 1, is marked. Through this process, a vertical bisector of the pupil circular boundary is formed at 0◦ and 180◦. Another procedure is followed with a vertical scan to form a horizontal bisector of the pupil circle at 270◦ and 90◦.Through connecting the midpoints of the two opposite bisectors, the center of the pupil is assumed. In this case, the radius of the pupil is also assumed to be half of the length of the opposite connected pixel. With the computed parameters, the circular shape of the pupil is accurately detected.

For localising the limbic boundary a different approach is taken, since the pixel intensities of the iris and white sclera are much closer. The image noise (eyelashes) is first removed by applying minimal blurring on the original image. The image is further smoothed using a 5 x 5 averaging filter. Finally, the previously computed pupil parameters are fed to the HT to detect the boundary of the iris.

Another research work, is one done by Sanap and Chaskar [60], which also proposes to use the HT in order to improve iris segmentation using the NIR images from CASIA database. The proposed method starts by pre-processing the input by removing reflections followed with global luminance thresholding to binarise the image. For pupil detection, the luminance threshold is set such that a black pupil is produced. From the binarised image holes are filled to better perform edge mapping of the pupil using the Sobel edge detector.

The generated edge map is a circular figure showing the pupil edges. Since

37 Chapter 2. Literature Review the radius is unknown, the HT is applied to the edge map and every point on the edge map is considered as a possible center where circles with radius r are traced as illustrated by Figure 2.8.

FIGURE 2.8: Center tracing from image domain to Hough domain

The precise value of the radius r, is one that forms maximum overlapping. Therefore, with the known value of the radius, the center of the circle can be traced by: F x = a + rcosθ (2.15)

y = b + rsinθ (2.16)

With the circle equation expressed as :

(x − a)2 + (x − b)2 = r2 (2.17) where (a, b) represent the center of the pupil with radius r. Using the known pupil radius as a measuring reference, a square portion of the image where the iris lies is cropped. The median filter is then applied to smoothen the cropped region containing the iris. The cropped iris region is further thresholded and the black pupil replaced with a grey value to compute the average value of pixels over the square portion containing the iris. This results in an iris edge map. Based on the computed average pixel value, the HT is applied to the edge map to get the exact radius in a similar fashion as the pupil, the center coordinates of the iris are obtained leading to iris localisation.

38 Chapter 2. Literature Review

Other recent research work that follow the same approach to perform iris segmentation includes the work by [61–65].

2.3.3 Segmentation techniques based on hybrid methods

As a measure to address the persistent iris segmentation challenges, researchers have also looked into employing hybrid methods. Hybrid methods are those that combine various algorithms particularly to increase factors such as performance speed, accuracy and pupil-iris boundary localisation from non-ideal images while taking into account that both the pupillary and limbic boundaries are non-concentric hence should not be modeled as perfect circles.

The work done by Sahmoud and Abuhaiba [66] uses the CASIA v4 and noisy UBIRIS to propose an algorithm to reduce the percentage error from images that still have certain types of noise including iris obstructions and specular reflection. This is done by adding a pre-processing task by employing K-means clustering to exclude the non iris regions; which cause many errors, and decrease the searching time of the later used Circular Hough Transform (CHT).

The proposed algorithm starts by using K-means clustering algorithm to divide the image into three regions which are; iris region, skin region and sclera region. This serves to determine and mask the expected region of the iris. From the grey level coded output image of the clustering algorithm, vertical Canny edge detection is applied produce a binary edge-map with detected edges. The usual errors resulting from the horizontal edges due to eyelashes and eyelids are decreased by this process. Morpholgical operations are then applied to remove any noise from the binary image. By combining the two described processes the searching time within the image parameter is further reduced.

39 Chapter 2. Literature Review

The produced binary edge image is then applied to the CHT in order to estimate the center and radius of the iris. After applying the CHT on the binary edge image, a maximum group of the center and radius parameters is selected from the CHT accumulator to find Cartesian parameters that successfully localise the iris. The results of the proposed method demonstrate that segmentation time and accuracy are highly improved enabling it to be used in real-time applications.

Uhl and Wild’s work [67] uses the CASIA v4 database to propose a drastically fast two stage generic iris segmentation approach under hard constraints. Their work starts by detecting and removing reflections from the pupillary region by computing a mask in three modules which include adaptive thresholding, region size filtering and morphological dilation. The process is followed by inpainting the original image with the computed mask, meaning all selected regions are reconstructed from their boundary using the Navier-Stokes method. From the inpainted image, the edge phase and its magnitude are estimated with the use of Sobel kernels.

The second stage is to find an estimate center point within the image’s pupillary and limbic boundary using a weighted iterative approach of the HT. This method finds the center of the most distinctive concentric circles in the image using the gradient magnitude and gradient orientation of the pupillary and limbic boundaries. The estimated center point is said to be close to the centers approximating the pupillary and limbic boundaries. This process is followed by transforming the input image into a polar representation using the estimated center as the origin. From the polar transformed image, a boundary which can either be pupillary or limbic is detected. This is achieved by determining the maximum-energy horizontal line, maximizing the vertical polar gradient for each column, smoothing the resulting curve, remapping candidates to Cartesian coordinate and fitting the edge points with an oriented ellipse. As mentioned that the detected boundary is either the limbic or pupillary boundary, a hypothesis

40 Chapter 2. Literature Review that determines either of the boundaries is applied such that the algorithm continues to find the second boundary using an Ellipsopolar transform.

Results report that the proposed individual modules can be efficiently extended to integrate more complicated techniques for individual tasks. Also that unsuccessful segmentation can be detected and modified in early stages, hence leading to more robustness.

The work of Roy et al.[68] is one example that uses Direct Least Square (DLS) and active contours in variational level set methods to perform iris segmentation. Their work proposes to demonstrate the significance of applying the variational level set-based curve evolution approach that employs a large time step to solve a partial differential equation (PDE) for iris segmentation from non ideal images. The argument posed in the work is that most active contour models are ineffectively designed and adjusted to perform iris segmentation in non ideal images. The aim is to speed up the segmentation process and gain robustness against poor pupil and iris localisation.

Roy et al.[68] use the ICE 2005, West Virginia university (WVU), UBIRIS datasets to conduct their experiment. The proposed approach starts by reducing noise, more specifically, the specular reflection within the pupil. To do this, first the input image is complemented through pixel intensity scaling and dark holes within the pupillary region are filled. The filling of wholes is done using 4-pixel connectivity based on the pixels of the pupil region’s background. The noiseless image is complemented again followed with applying a Gaussian filter to smoothen the image.

This process is followed by suppressing eyelashes through applying a morphological operation known as opening. In mathematical morphology, opening is explained as the dilation of the erosion of a set A by a

41 Chapter 2. Literature Review structuring element B [69], as follows:

AoB = (A B) ⊕ B (2.18) where and ⊕ respectively denote erosion and dilation.

For estimating the pupil boundary, the ellipse fitting method of DLS is employed. Fitzgibbon et al.[70] reports that ellipse fitting techniques can be divided into clustering techniques and least-squares fitting. Least-squares techniques are aimed at “finding the set of parameters that minimize some distance measure between the data points and the ellipse”. So, from the proposed scheme, five parameters are returned which include; the horizontal and vertical coordinates of the center of the pupil, the length of the major and minor axes and the ellipse orientation.

Here, a sudden change in luminance is summed around the pupil circumference, with its boundary detected as an ellipse on the iris image. Similarly, to get an approximation of the iris contour, the elliptical fitting process is applied again. Results report that the proposed method provides a reasonable estimation of iris and pupil boundaries.

2.4 Problematic encounters

The modifications done on the traditional algorithms by other researchers have been presented. Indeed the proposed techniques may have shown some improvements to a certain degree, however they still present some unsettling concerns such as :

(i) Most of the methods share a repetitive nature or are very closely related in the manner in which the overall segmentation process is approached.

(ii) The need to perform a reasonable amount of pre-processing, with the in between, multiple sub model designs, and implementations to

42 Chapter 2. Literature Review

perform a single segmentation task, calls for high computational demands and thus posing a challenge for real time applications.

(iii) Some of the methods still need manual modifications on pupil and iris center detection, meaning full automation has not been reached and as a result, cannot be used in automated iris recognition systems.

All these factors are evidence that plausible algorithms that can satisfy "all” scenarios and still be deployed in real time have to be tailor made for specific needs, for specific applications or else, bear a certain trade-off.

2.5 Summary of conclusions

From the literature review that has been presented, it has been learned that as proposed by pioneers in iris recognition, Daugman and Wildes, even the most recent segmentation methods still employ similar approaches. Most work has been done with a focus and an aim of correcting the drawbacks and assumptions that still burden traditional methods, However burdens such as:

(i) the different devices used to capture eye images for database accumulation;

(ii) the willingness or cooperation of participants to have their eye images captures;

(iii) the capturing process; in which environmental conditions are assumed to be highly constrained to acquire well focused eye images. are some of the inevitable challenges that lead to the prerequisite to involve numerous sub models and subtasks resulting to multiple models needed to achieve a single task of accurate iris segmentation. This manner of approach ultimately deems some of the methods proposed in literature to be more applicable for certain or specific applications but unpractical for real time deployment.

43 Chapter 2. Literature Review

Based on the investigated literature, the next Chapter is motivated by :

1. The lack of available segmentation methods that do not regard the pupil and iris boundaries as perfect and concentric circles. In the literature review, the perfect circle approach has demonstrated to be a challenge; especially when working with eye images that have not been acquired under a strictly controlled real life environment; thereby coming with non-desirable artefacts that lead to unsuccessful segmentation, difficulty in template creation and high matching errors.

2. The restriction introduced by the rubber sheet model to perform iris normalisation. Although this method is excellent for addressing the challenges of iris image variations after segmentation, it is more relevant for person recognition purposes. The challenges with the rubber sheet model is that: (a) the extracted iris features cannot be traced back to the original input image and (b) this model alters the global geometric formation of iris textural patterns that are needed for ethnic and gender classification.

3. The unavalilablity of algorithms that use alternative iris features and methods for recognising and matching an indvidual’s query to its reference image.

4. The lack of alternative algorithms that address the consequence of the large amount of an iris feature vector generated or produced by a single iris image through the use of traditional methods, hence requiring more storage space.

From the work that has been reviewed in this Chapter, the topic of employing level set methods and active contour models for iris segmentation, is amongst one of the contributions of this thesis. This approach is investigated so as to respect the shape and non concentric nature of the pupil and iris; to address and also solve the problem of pupil and limbic boundary localisation from non ideal eye images. This is done

44 Chapter 2. Literature Review because the images used to conduct experiments in this work have been self-acquired, from a population that has never been investigated in literature; under non constrained environments; with different acquisition location; varying illumination; and lack of subject cooperation. Details on this approach are presented in Chapter 3.

45 Chapter 3

Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images

This chapter discusses the proposed experimental approach; paying special attention to iris segmentation from non-ideal eye images towards achieving one of the research objectives discussed in Chapter 1. Section 3.1 introduces the problems and challenges associated with the task of performing iris segmentation from non-ideal eye images; as well as the contributions and motivations behind the proposed segmentation approach. Section 3.2 presents a background on active contour models as one of the tools employed in this thesis to achieve accurate iris segmentation from non-ideal eye images. Section 3.3 reviews the related work on iris segmentation using active contour models and how other researchers have modified the algorithm to deal with eye images acquired from non cooperative participants under differing environmental conditions. Section 3.4 presents the proposed experimental approach on iris segmentation, feature detection, extraction and the matching of extracted iris features. Presented in section 3.4 are also the results and the performance analysis of the proposed iris segmentation approach. Section 3.5 presents a concluding summary of this chapter.

46 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images

3.1 Introduction

The task of segmenting the iris from images captured with a standard iris camera under uncontrolled environments; such as different locations with varying illumination and lack of subject cooperation, is challenging. Eye images captured from non-cooperative participants neither have perfect circular nor elliptical shapes, since they are often obstructed by noises such as:

(i) Eyelashes and eyelids.

(ii) A highly dilated pupil.

(iii) Low contrast difference between the iris and sclera regions.

One of the main contributions and objectives of this thesis is to propose a fast and accurate iris segmentation method from eye images that have been captured under uncontrolled environments. The motivation is based on the challenges that accompany eye image acquisition in a real environment; whereby participants are non cooperative and the acquisition conditions are not favorable. Another motivation is the inability to deploy some of the previously proposed iris segmentation algorithms in real time applications due to the assumption that all investigated eye images have been acquired under strict conditions, thereby producing images that are well centered and focused on the camera lens.

The following section reviews the segmentation of non-ideal iris images based on active contour models and level set methods. These methods are novel and alternative approaches that can deal with realistic and practical scenarios where the available images might also be affected by pupil dilation, deviated gaze, head tilting , reflections, motion blur, non-uniform intensity and low image contrast. Furthermore, the similarity with these methods is that they do not assume concentricity of the pupil-iris boundaries like the state of the art algorithms. Ultimately, a segmentation

47 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images method that can manage eye images captured at a distance without enforced subject cooperation is a practical requirement that could result in an even better control over conditions such as transit points.

3.2 Background on active contour models

Active contour models, also known as snakes, is an image segmentation technique that was proposed and developed by Kass [71]. The technique is referred to as "snakes" because of the manner in which the active curve mimics the behaviour of a snake wrapping itself around any particular object in order to perform segmentation. This technique is used in many computer vision applications as a method that creates an active curve to suitably separate an object of interest (of any shape) from the rest of the image. The active contour moves through the spatial domain of an image to minimise an energy functional through level-set formulation.

Given certain constraints and image forces such as lines and edges within an input image, the model minimises energy through an active guided spline that will lock itself around the desired constraints, lines and edges. The flexible and active spline employed by active contour models makes it possible to perform and achieve accurate iris segmentation, even from unfocused eye images, thus addressing some of the concerns associated with the use of traditional iris segmentation algorithms [72].

The basic model uses a snake represented by a curve [73, 74], v(s) = [x(s), y(s)]s ∈ [0, 1], where s represents the length of the curve and the energy of the snake defined by:

Z 1 Z 1 Esnake = Eint(v(s)) ds + Eext(v(s)) ds (3.1) 0 0 where Eint(v(s)) and Eext(v(s)) represent the internal and external energy of the snake respectively. The internal energy Eint(v(s)) controls the

48 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images deformability of the snake, and is represented as:

1 E (v(s)) = (α(s) |v (s)|2 + β(s) |v (s)|2) (3.2) int 2 s ss

where vs is the first spatial derivative representing elasticity. The term vss is the second spatial derivative representing rigidity or thin-plate behavior. The coefficients α and β are weighting parameters respectively controlling the elasticity and rigidity of the contour.

The external energy function is derived from the energy of the image, especially where image edges are of interest, the energy is defined as:

2 Eext(v(s)) = − |∇(Gσ(x, y) ∗ I(x, y))| (3.3)

where Gσ(x, y) is a Gaussian function, with σ as the standard deviation , ∇ is the gradient operator and ∗ represents convolution while I(x, y) represents the image intensity. Substituting 3.2 and 3.3 in 3.1, expresses the snake as:

Z 2 2 2 Esnake = (1/2(α(s) |vs(s)| + β(s) |vss(s)| ) − (∇Gσ(x, y) ∗ I(x, y)) ) ds s (3.4) For the snake energy to be minimised, the following Euler equation must be satisfied:

αvss − βvssss − ∇Eext = 0 (3.5)

The solution of 3.5 is the final contour minimising Esnake. This function could be considered as the force equilibrium function, expressed as:

Fint + Fext = 0 (3.6)

where Fint is the internal force that discourages stretching and bending.

The external force Fext pulls the snake towards the desired edges of the image. Therefore, when the original contour evolves and deforms into the

final contour (Fint = −Fext), it means that for every point along the curve,

49 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images the internal and external forces are not only equal but also act in opposite directions to each other producing a stable state.

3.3 Related work on active contour models

This section presents the related work done by other researchers who have employed active contour models with modifications to either segment the pupil or the iris from non-ideal images. For example:

The work of Abdullah et al.[73] proposes a fast and accurate pupil segmentation model from non-ideal images having any iris shape. The proposed algorithm is based on the combination of morphological operations and the snake active contour model. The proposed method is implemented on the CASIAv4 database acquired under near infra-red (NIR) illumination.

Their proposed method consists of five modules namely:

(i) the removal of reflections,

(ii) thresholding for binary image generation,

(iii) removal of eyelashes,

(iv) coarse pupil boundary estimation and

(v) accurate pupil segmentation.

For the removal of reflections caused by the NIR illuminators within the pupil region, a procedure called image opening is performed, (procedure explained in equation 2.18 ). Here, the pixels with reflections are filled with those belonging to the pupil region. This procedure is followed by thresholding with Otsu’s method [75]. Instead of manually selecting a threshold value, Otsu’s method [75] supposes that an image consists of a background and foreground region. Therefore this method combines the best threshold of the two regions to compute a minimal spread of the

50 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images regions as a threshold value. Abudallah et al.[73] further mention that taking a quarter of Otsu’s threshold value also serves as a good approximation to separate the dark pupil and white sclera regions. The result of this step, is a binary image.

From the binary image, the presence of eyelashes, which appear as dark as the pupil have to be removed. This is done by using a morphological dilation operation of applying a 8x8 square structuring element to the binary image. This operation manages to eliminate eyelashes from the binary image. In order to only retain the pupil, another structuring element is applied as a disc having a radius that is lower than the smallest pupil within the database. The output generated by this operation is a clean binary image of the pupil. From the localised binary pupil, a coarse or rough estimate of the pupil center (xc, yc) is computed through calculating the center of mass pixels using:

N 1 X (x , y ) = (x , y ) (3.7) c c N i i i=1 where N is the total number of black pixels. The pupil radius on the other hand, is obtained by using the area of the circular pupil region using:

r N r = (3.8) Π

To finally segment the pupil with accuracy, the estimated center and radius are used as parameters for the snake’s initial curve. Furthermore, to ensure that the curve always lies on the outside of the pupil boundary, a constant value is added to the roughly estimated radius. Their proposed method demonstrates robust localisation and segmentation of the pupil. The manner of approach however, makes pupil localisation another segmentation task on itself, which when coupled with the ultimate task to achieve iris segmentation, requires a lot of computational time, hence making it impractical for real-time deployment.

51 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images

Abdullah et al.[74] later continued this work by proposing a segmentation method for non-ideal iris images captured under visible light from the Multimedia university (MMU2) database and near infrared (NIR) light from CASIAV4 database. Their work proposes to perform iris segmentation in two stages. The first stage is for pupil segmentation using two methods for the images taken under the two different light sources. The stated reasoning behind this approach is that for images acquired under NIR, the pupil appears darker than its surroundings, with a high contrast between the pupil and the iris region. For images captured under visible light however, the contrast difference between the pupil and the iris is reported to be low.

Their proposed model uses a gradient vector flow (GVF) active contour model integrated with an external force field. The argument is that when used alone, traditional active contour models tend to be sensitive to the initial curve and hence fail to detect objects that are non-convex. Therefore the introduced external force acts as a force of pressure that pushes the contour to the boundary of the object. Therefore, in order to segment the pupil for NIR images, a thresholding based method is employed; where GVF active model is combined with the mean shift and morphological operations. The mean shift is for clustering the iris image; while morphological operations will approximate the pupil boundary. The proposed active contour is then used to find the precise boundary.

The second stage is for iris segmentation. Here, the eyelashes are first removed using 2-D statistic filtering. A new circular mask with a larger radius is created outside the pupil and used to initialise the active contour to evolve and converge to the iris boarder. With their proposed method, they achieve a 95.1% correct iris and pupil boundary localisation using the NIR images from CASIAv4, and a 93.7% with the visible light images from MMU.

52 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images

Shah and Ross [72] propose an iris segmentation scheme from non-ideal images using geodesic active contours (GAC). In order to do this, the CASIA v3, West Virginia university (WVU) and UBIRIS datasets are used. The proposed model performs iris localisation in two stages of pupil segmentation and iris segmentation. The process starts by using 2D median filters to smoothen the input eye image and to determine the minimum pixel value (M). This is followed by image binarisation using a threshold value of 25+M. The product of the binarisation is both the detection of the pupil with eyelash noise. The pixel intensity shared by the dark pupil and black eyelashes makes both of them to fall below the selected threshold value. The 2D median filter is again applied to the binary image to eliminate the smaller eyelash regions. However, the median filtered image is still accompanied by some forms of eyelash noise. This is then corrected through tracing the largest circular boundary within the image using the equation of a circle. Here, the circle with a circumference having the maximum number of black pixels is deemed as the pupil.

For the second stage of detecting the limbic or iris boundary, the GAC; which is a scheme based on levels sets, is employed. This means, the model uses the relationship between active contours and the computation of geodesics (curves with minimal length). The GAC is modeled by selecting the order of the Fourier series, such that the curve gravitates towards the boundary of an object at a particular time, which corresponds to number of iterations. The evolving curve, which is a level set function, is used as an embedding function with a stopping term. This embedding function is used to evolve the active contour from inside the iris and to stop on the iris boundary based on the designed stopping criteria.

The results of Shah and Ross demonstrate high accuracy levels, especially for iris segmentation. The detection and localisation of the pupil uses its

53 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images own scheme that is totally different from that used for the iris. Inasmuch as the proposed model performs well in terms of accuracy, the issue of being optimal is a major concern due to the involved number of various schemes that have to be separately designed.

Similar to the related work presented above, other research work that involves using active contours to perform iris segmentation is the work from Daugman[49], where the pupil and iris boundaries are detected using an active contour model in which the contour data is based on the discrete Fourier series expansions. Vatsa et al.[76] evolve the contour based on geometric active contours. The work from Roy and Bhattachriya [77] uses the non-ideal images from UBIRIS Version 2, the ICE 2005, and the WVU datasets to segment the pupil boundary by applying level set based curve evolution method using the edge stopping function. The limbic boundary is detected by employing the curve evolution approach that uses the regularised Mumford-Shah segmentation model with an energy minimisation approach. The work of Moosavi et al.[19] uses the greedy snake algorithm to successfully localise both the pupil and iris boundaries. Another research work that uses geodesic active contours to segment the iris from the noisy UBIRIS database, is the work of Kalavathi and Narayani [78]. Their work uses the to find image edges. In order to identify the circular objects within the edge map, the Hough transform (HT) is employed. The GAC is then used to find the iris boundary.

From the various approaches proposed by other researchers, it is evident that active contour models can be modified in various ways. The challenging similarity with most of the modifications that have been used and presented in the above section, is the use of the edge detection approach for iris segmentation. The approaches which include: probabilistic active contour, geodesic active contour and Fourier coefficients based active contour are challenging when employed in iris

54 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images images acquired from darker skinned participants. This is because of the very low contrast and very smooth edges between the pupil and iris boundaries from dark skinned and highly pigmented eye images.

The following section presents yet another approach to active contours. This approach is one that has been adopted and will be later implemented in this thesis to perform iris segmentation.

3.3.1 The model of active contours without edges

The segmentation method proposed in this work, follows another approach to active contour model modification, called "active contours without edges". This approach was developed by Chan and Vese [79, 80], hence referred to as the Chan-Vese algorithm. The algorithm is an example of a geometric active contour model. While, the previously discussed active contour methods are in some way heavily dependent on detecting edges. This model however, uses image region information allowing it to easily segment any topological variations within an image [81].

The model starts with a contour in the image to define an initial segmentation. Using some evolution equation, the contour is evolved. The objective is to evolve the contour such that it discontinues or stops on the boundaries of the foreground region [35].

For the contour to evolve, the Chan Vese algorithm uses the level set method. The level set function is an excellent means of achieving contour evolution. It is the level set function that defines both the edge contour and the segmentation of the image. The challenge therefore is to determine a suitable level set function. Ultimately, the goal is for a given image u0, to minimise a fitting energy functional where the determined level set function achieves successful segmentation of the object within the image.

55 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images

2 For instance, let Ω be a bounded open subset of R . The model of Chan and Vese defines an evolving curve C in Ω as the boundary of an open subset ω of Ω. The region inside and outside the curve is denoted by ω and Ω/ω respectively. Let u0 be a given image formed by two regions of distinct

i o i values u0 and u0 having piecewise-constant intensities. With u0 ≈ u0 and o u0 ≈ u0 falling inside and outside the object boundary (C0) respectively. Considering the fitting term:

Z 2 F1(C) + F2(C) = |u0(x, y) − c1| dx dy inside(C) Z (3.9) 2 + |u0(x, y) − c2| dx dy outside(C) where C is any variable curve, constants c1 and c2 depend on C, are respective averages of the image pixel intensities inside and outside C. In this case, the boundary of the object C0 is the minimiser of the fitting term.

The objective however, is to minimise the energy functional F (c1, c2,C), by adding regularising terms, i.e. the length of the curve C and (or) the area of the region inside C, which when introduced to the energy functional becomes:

F (c1, c2,C) = µ.Length(C) + ν.Area(inside(C)) Z 2 +λ1 |u0(x, y) − c1| dx dy (3.10) inside(C) Z 2 +λ2 |u0(x, y) − c2| dx dy outside(C) where µ ≥ 0, ν ≥ 0, λ1,λ2 > 0 are fixed paramemeters with λ1 = λ2 = 1, ν = 0 and . is the dot product.

The fitting term is based on the Mumford-Shah functional developed by Mumford and Shah in [82]. It is noteworthy that the term Length(C) can be re-written more generally as (Length(C))p for p ≥ 1. Considering a case of arbitrary dimension N > 1, then p = 1 for all N, or p = N/(N − 1). The isoperimetric inequality can be used which makes (Length(C))N/(N−1) to

56 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images be “comparable” with the Area(inside(C)).

Area(inside(C)) ≤ c.(Length(C))N/(N−1) (3.11)

Since the model of Chan and Vese is that of a minimal partition problem formulated by the level set function, the variable of the evolving curve is replaced by the zero level set of a Lipschitz function φ :Ω → R. This then simplifies the Mumford-Shah functional responsible for segmentation. In other words, we are looking for c1, c2, C that will be a solution to the minimisation problem:

inf F (c1, c2,C) (3.12) c1,c2,C

Thus using the formulation of the level set, the solution image u, constants c1 and c2 can be expressed as a function of φ by:

R Ω u0(x, y).H(φ) dx dy c1 = R (3.13) Ω H(φ) dx dy

R Ω u0(x, y).(1 − H(φ)) dx dy c2 = R (3.14) Ω(1 − H(φ)) dx dy   1, x ≥ 0 where H is the Heaviside function H(x) =  0, x = 0 Parameterizing the descent direction by some artificial time t ≥ 0 and deducing the associated Euler-Lagrange equation, leads to the active contour model, [80].

" ∂φ ∇φ = δ (φ) u div − ν − λ (u − c )2 ∂t ε |∇φ| 1 0 1 # 2 +λ2(u0 − c2) = 0 in (0, ∞) × Ω, (3.15)

φ(0, x, y) = φ0(x, y) in Ω,

δε(φ) ∂φ = 0 on ∂Ω |∇φ| ∂~n where φ(0, x, y) = φ0(x, y) defines the initial contour, ~n is the exterior

57 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images normal to the boundary ∂Ω, and ∂φ/∂~n denotes the normal derivative of φ at the boundary. The principal steps of the Chan Vese algorithm are presented in algorithm 3.1.

Algorithm 3.1: Chan-Vese principal algorithm 0 Data: φ = φ0, n = 0 Result: φ0 =segmented iris, PC2 Initialisation; n n compute c1(φ ) and c2(φ ) by 3.13 and 3.14; solve partial differential equation in φ from 3.15, to get φn+1; See if solution is stationery, otherwise, n = n+1 and repeat, [80];

The following section presents some of the work that has been done by other researchers to segment the iris using the model of Chan and Vese.

3.3.2 Related work on iris segmentation using the Chan-Vese algorithm

The work done by Yahya and Nordin [17] proposes to segment the iris from the noisy UBIRIS database using the Chan-Vese active contour model. Their proposed method starts by removing specular reflections from the noisy input image using an inpainting technique that fills in the reflections caught by the iris region at the time of acquisition. Details on the inpainting technique are described in [83]. The process is followed by determining if the eye is not closed as well as the detection of the iris region using adaptive boosting (AdaBoost)-Cascade Detector. AdaBoost is a classification algorithm that combines weak classifiers to produce a strong classifier. In their work, the algorithm “classifies” a square as an indicator of the iris region. Once the iris region has been detected, the iris center is then estimated. The approximated iris center is then used as the initial point for the curve evolution of active contour until the energy function arrives at the minimum, which is the iris boundary.

The work by Hilal et al.[18] uses the CASIAv3 and UBIRIS database to propose a method that can detect the pupillary boundary in its real shape

58 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images regardless of its irregular geometry. The argument in their proposed work is that the outer iris or limbic boundary is very close to a circle, whereas the inner iris or pupil boundary is close to an ellipse. Therefore, based on this argument, they first employ the circular Hough transform (CHT) to detect the limbic boundary and to get the rough estimate of the circle parameters. Using the estimated parameters, the Hough transform is applied again to define the pupil region of interest, which in this case lies within the limbic boundary.

The circle parameters obtained from pupil region of interest (ROI) are then used to initialise the contour of the Chan-Vese algorithm. Here, the iterative evolution of the active contour manages to detect the actual shape of the pupil and hence the pupillary boundary is well defined. Lastly, the linear Hough transform is used to detect the upper and lower eyelids, and an intensity thresholding method for detecting eyelashes, since they are generally darker.

Generally, the use of the HT for the detection of lines and circles within an eye image consumes a lot of time and thus reduces processing speed. The approach proposed by Hilal et al.’s work, not only makes the arguable assumption that the iris is circular and the pupil elliptical, it also applies the HT three times, before achieving the final result.

Another related work is done by Yanto et al.[84]. Their work proposes to perform iris segmentation through formulating the energy function defined by the Chan-Vese algorithm as a Bayesian optimisation problem. Their work uses the CASIA v3 dataset to perform experiments.

In order to detect the iris center, the method proposed by Yanto et al.[84] starts by preprocessing the input image through applying the mean shift filter. This is done to ease the process of finding the iris center. By plotting a gray level histogram and considering that the pupil is darker, the first peak

59 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images in the histogram is identified as corresponding to the pupil area. From the histogram, a threshold value which serves as the pupil center, is computed using the average values of the pixels in the first peak and the pixels that are adjacent to the peak. This results in a binary image, where zero (0) and one (1) respectively correspond to pixel values greater and lesser than the threshold. The produced image is a white circle that shows the estimated area of the pupil. This process is elaborated by Figure 3.1, [84].

(a) (b)

(c) (d)

FIGURE 3.1: Iris center detection: (a) Input Image. (b) Pre- processed image. (c) Binary image. (d) Detected iris center

By determining the center coordinates of the produced binary image or white circle, the iris is successfully localised. This process is followed by the detection of the iris inner and outer boundaries, that is, the pupillary and limbic boundaries. For the pupillary boundary, the center coordinates obtained from the binary image, are used to create a circular mask having a radius of ten pixels. The number of iterations for the evolving curve is set to 100, to avoid the algorithm from terminating prematurely and ensuring that the pupillary boundary is correctly detected. This process is followed by detecting the limbic boundary or outer iris boundary. In order

60 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images to do this, another circular mask which falls outside the pupil boundaries is newly created. The chosen radius is 2.5 larger than that of the pupil having 100 iterations as well. Results of their proposed work are shown in Figure 3.2.

(a) (b)

FIGURE 3.2: (a) Pupil boundary detection. (b) Iris boundary detection

3.3.3 Summary on related work

The work that has been done by other researchers using the different modifications of the snake model has been presented. The Chan-Vese model which is another modification to active contours has been explained with discussions of the work related as implemented by other researchers.

From the presented literature, it has been learned that the motivation behind the use of active contours is that the dynamic evolution of its active spline makes it appropriate and practical to segment non-ideal iris images. The challenges however, are that the traditional snake model depends or uses the edges of the image, so that the external energy derived from the image energy, can pull the snake towards the edges to perform segmentation. This characteristic poses a drawback because for images acquired from darker skinned participants with very smooth boundaries, the snake can easily pass the desired edge.

The unfavourable similarity that has been learned from the related work is the amount of preprocessing tasks and sub model designs that are

61 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images required to perform accurate iris segmentation from non-ideal images.

The model of Chan and Vese uses image geometry and image region information to segment any object of any shape. The most important task is to determine the formulation of the level set function that will meaningfully segment the object of interest within the image [35].

3.4 Proposed experimental approach

In order to reach our end goal of developing a methodology that uses iris textures to classify individuals according to ethnicity and gender, careful consideration must be given to the segmentation stage. This is because, not only is iris segmentation the most crucial and computationally heavy task, but it also determines the success of all subsequent stages. Therefore, too much pre-processing during this stage might degrade the quality of the desired iris texture features at a later stage.

Taking into consideration the challenges that have been drawn from the presented related work, the experimental approach proposed in this thesis aims to address and contribute to the segmentation of the iris from non ideal eye images and those acquired from darker skinned participants. The motivation here is that there is no existing iris database that includes eye images from darker skinned or Black Africans; hence the images that will be used for the rest of this work have been independently acquired to further investigate ethnic and gender classification. Another motivation is the established restrictions and complexities that accompany the use of state-of-the-art algorithms to achieve fast and accurate segmentation from images acquired under non strict acquisition environments. The proposed approach is summarised as:

(i) Acquisition of eye images from individuals with low iris pigmentation, i.e. white participants; and from individuals with high pigmentation, i.e. black individuals.

62 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images

(ii) The use of morphological operations to remove camera reflections from eye images.

(iii) The use of the antiquated algorithm by Bresenham [85], to detect the pupil and iris center coordinates.

(iv) The combination of Bresenham’s algorithm and Chan-Vese [80] algorithm for pupil-iris boundary localisation and segmentation respectively.

(v) The detection and extraction of corner features from iris patterns through the fusion of phase congruency and Harris algorithm.

(vi) The use of the obtained iris corner features to effectively create and perform iris template matching.

(vii) To test the robustness of the proposed model by implementing it on other databases.

This is achieved by approaching each stage of iris recognition using different algorithms from the traditional ones in order to:

(i) achieve accurate and simultaneous localisation of the pupillary and limbic boundary segmentation, even from images with very smooth pupil-iris boundaries.

(ii) achieve fast and non complex iris segmentation given any image quality in a single step,

(iii) preserve the quality and geometry of iris texture patterns during segmentation and

(iv) offer deployment in real-time applications.

Figure 3.3 is a block diagram elaborating the model of the proposed experimental approach.

63 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images

FIGURE 3.3: Flow diagram of the proposed methodology.

The arrangement of the following sections is such that the results obtained from the independently collected database will be demonstrated first; followed by the results from the CASIA database and the UBRIS database. The discussions and the analysis of the attained results will also be done correspondingly.

3.4.1 Eye image acquisition

In order to objectively investigate and complete the goal of this thesis, an independent database of eye images also accommodating the highly pigmented darker skinned individuals was acquired. This is also done to assess the performance of the proposed experimental approach with challenging eye images from such a population. The Vista EY2 dual iris scanner and face camera [86] shown in Figure 3.4 was used to capture images.

FIGURE 3.4: VistaEY2 imaging device

The VistaEY2 has been developed with both an iris scanner and face camera that simultaneously captures both the left and right eye of an individual at

64 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images a maximum automated detection distance of up to 38 cm. The iris camera features multi wavelength infrared illumination. Acquired images are ISO/IEC 19794-5/6 compliant and the biometric device meets international Eye Safety Requirements. The acquired images are bitmap type of size is 640 x 480 pixels. The camera comes with a standard USB 2.0 interface and the software is fully supported by Windows CE, XP, Vista, Windows 7 and Linux.

Images were acquired from willing, consenting and working adults. The collected database consists of individuals from 22 black males and 22 black females; as well as 22 white females and 22 white males to give a total of 352 images. Unfortunately, four of the white male participants’ images could not be used, this resulted to a total of 18 white male participants in the study. This imbalance is accounted for in all the results. The images were acquired from different locations of the participants’ place of work and at different days and times. This means, the images bring challenges of varying illumination, reflection noise, lack of subject cooperation, out-of-focus, pupil dilation and contraction. From the total amount of images, 312 were of usable standard.

3.4.2 Reflection noise removal approach

Since the imaging device uses NIR light, the acquired images come with some light reflection spots inside the pupil region. This type of noise can have a great impact when the iris features that are generally used for individual recognition have to be extracted for further template creation. In this work, the reflection spots are removed in two step: applying morphological operations of image dilation and region filling. The processes are illustrated by Figure 3.5.

65 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images

(a) (b)

(c) (d)

FIGURE 3.5: Noise removal:(a) Input grey scale Image. (b) Histogram of grey image. (c) Dilated binary image. (d) Reflections removed from independent database.

The process starts with an image captured from the acquisition stage as an input image in Figure 3.5 (a). Since it is an indexed bitmap (.bmp) image, it is first converted to grey scale for computational ease and luminance uniformity amongst various databases. The reflection spots or noise inside the pupil region, appear brighter than the rest of the image. By computing the image histogram, it becomes easier to visualize the varying high (brighter) and low (darker) pixel intensities; as well as to determine a threshold that will be used for detecting the spots. From the histogram scale shown in Figure 3.5 (b), values close to zero (0) and one (1) represent lower and higher pixel intensities respectively. The reflection spots are brighter, hence close to 1. Given this factor, a threshold of 0.9 is set to highlight all the parts of the binary image that are very close to 1, where the reflection spots are detected. This process is followed by applying the mathematical morphological process of image dilation to the binary image to achieve the results illustrated by Figure 3.5 (d). The binary dilation of A

66 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images by B, denoted A ⊕ B, is defined as set the equation:

ˆ A ⊕ B = {z (B)z ∩ A 6= ∅}, (3.16)

The operator uses two pieces of data as inputs: A, the image to be dilated and a small set of coordinate points B, referred to as the structuring element or kernel. The structuring element is responsible for determining the particular effect of the dilation on the input image. This operation returns Bˆ as a reflection of the structuring element B, with z specifying the pixel locations. This process is followed by using the pixel location to fill in the stipulated region of interest with values of the surrounding region. The proposed experimental approach produces the desired result seen in Figure 3.5.

In order to test the robustness of this method, we implement the algorithm on the CASIA v3 and UBURIS databases.

(a) (b)

(c) (d)

FIGURE 3.6: (a) Input grey scale image. (b) Histogram of grey image. (c) Dilated binary image. (d) Reflections removed from CASIA database.

67 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images

For the CASIA and UBIRIS databases, the images are of type .jpg. Here, image the histograms are scaled between zero (0) and 255 to represent low or darker and high or brighter pixel intensities respectively. For this particular case the threshold values are set to 254 for the CASIA images and 240 for the UBIRIS images. The obtained results are shown in Figure 3.6 for CASIA and Figure 3.7 for UBIRIS images.

(a) (b)

(c) (d)

FIGURE 3.7: (a) Input grey scale Image. (b) Histogram of grey image. (c) Dilated binary image. (d) Reflections removed from UBIRIS database.

3.4.3 Iris segmentation

The process of iris segmentation is performed in two steps. The first step is to formulate the level set function or to get the regularising terms that will be fed to the Chan-Vese algorithm. The terms are pre-defined by using Bresenham’s algorithm. Bresenham’s algorithm is based on the midpoint circle algorithm and is used here to compute the pupil-iris boundaries. It is an algorithm that determines points that are needed to draw a circle around an object.

68 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images

The approach of mapping the circle is done with the acknowledgment that the pupil and iris boundaries have annular shapes. Therefore through exploiting their respective diameters, the pupil-iris boundaries (Co) can be simultaneously located prior segmentation. This will also help to easily stop the evolving curve on the area designated by the pupil and iris radii computed from their respective diameters.

Starting with an original input eye image u0(x, y), we independently compute the end points u0(x1, y1) and u0(x2, y2) of both the iris and pupil diameters, elaborated by Figure 3.8. This is done to locate their respective centers (h, k) using the midpoint circle formula:

x1 + x2 y1 + y2 u (x)( ), u (y)( ) ⇒ (h, k). (3.17) 0 2 0 2

The distance formula is employed to compute the radii of both the pupil and iris by:

p 2 2 d = (x2 − x1) + (y2 − y1) , (3.18) and r = d/2. (3.19)

FIGURE 3.8: Detection of iris and pupil diameters.

The second step is to add the output of Bresenham’s algorithm as regularising terms to the Chan Vese algorithm, where the segmentation of

69 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images the iris is represented by:

Z MS 2 F (u, C) = µ.Length(C) + λ |u0(x, y) − u(x, y)| dxdy Ω Z (3.20) + |∇u(x, y)|2 dxdy. Ω/C where F MS is the Mumford Shah functional responsible for segmentation, u0 : Ω → R is the given image and µ and λ are positive parameters. The length of the curve is parameterised by the radii equation 3.19 for both pupil and iris.

The whole process is implemented using images from the independently self collected database having images with very smooth pupil-iris boundaries as seen in Figure 3.9 and Figure 3.10. The initial contour has been defined by the parameters obtained from Bresenham’s algorithm; and the pupil and iris boundaries as detected by the diameters. The stopping of the evolving contour is influenced by the respective computed radii.

(a) (b)

(c) (d)

FIGURE 3.9: Pupil localisation: (a) Input image. (b) Pupil initial contour. (c) Localised pupil boundary. (d) Pupil mask.

70 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images

(a) (b)

(c) (d)

FIGURE 3.10: Iris localisation: (a) Input grey scale image. (b) Iris initial contour. (c) Localised iris boundary. (d) Iris mask.

Evidently the radii of the images in the database are not equal, however due to the fact that the Chan-Vese algorithm also uses region information, it manages to use the computed radii as an estimation to stop the evolving on the desired boundary. This is the added advantage of the proposed approach implemented with the Chan-Vese algorithm. The result is the iris as segmented by the Chan-Vese algorithm through the proposed experimental approach, shown in Figure 3.11.

FIGURE 3.11: Example of a segmented iris.

71 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images

The testing of the proposed approach is also implemented on the CASIA v3 and UBURIS databases. The parameters obtained through Bresenham’s algorithm and fed to the Chan-Vese algorithm as regularising terms, manage to successfully define the pupil and iris initial contours. The localisation of the respective boundaries are also shown by Figure 3.12 and Figure 3.13.

FIGURE 3.12: (a) Input grey scale image. (b) Pupil initial contour. (c) Localised pupil boundary. (d) Pupil mask.

FIGURE 3.13: (a) Input image. (b) Iris initial contour. (c) Localised iris boundary. (d) Iris mask.

From human observation, the difference between eye images acquired from Black, Asian and European subjects can be easily perceived. Black subject

72 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images tend to have a very low contrast between the iris and pupil unlike the Asian and European subjects, where the contrast between the pupil and iris can be easily seen. Ultimately, the objective is to achieve a segmented iris shown by Figure 3.14, that can be used for further processing.

FIGURE 3.14: Example of the CASIA segmented iris.

Another merit of this method is that given the different images from the different databases, the parameters used to segment images from our self-collected independent database, also achieved similar brilliant results with both the CASIA and UBIRIS databases. The results otained from the UBIRIS database are also shown by Figure 3.15, which demonstrates the pupil initial contour and its boundary localisation.

FIGURE 3.15: (a) Input grey scale image. (b) Pupil initial contour. (c) Localised pupil boundary. (d) Pupil mask.

73 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images

Figure 3.16, shows a similar process for the iris, with the resulting segmented iris shown by Figure 3.17.

(a) (b)

(c) (d)

FIGURE 3.16: (a) Input image. (b) Iris initial contour. (c) Localised iris boundary. (d) Iris mask.

FIGURE 3.17: Example of the UBIRIS segmented iris.

3.4.4 Results analysis

While running the segmentation experiments, the first encounter was that, since the main focus for extracting features is done from the iris region, the noise from the pupil region can actually be neglected. The reflection spots define a different region within the pupil, the proposed Chan-Vese

74 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images algorithm then uses this noise to start the evolving initial contour, however, only using the pre-defined pupil parameters to stop the contour on the desired boundary.

In this manner, accurate segmentation is still achieved without the risk of suffering the challenges introduced by noise removal. This discovery means the noise removal step can be skipped hence reducing the in between substaks and the overall processing time. Examples of this case are shown using other images within the databases, in Figure 3.18 for the pupil localisation, Figure 3.19 for the iris localisation and Figure 3.20 for the final segmented iris.

(a) (b)

(c) (d)

FIGURE 3.18: (a) Input image. (b) Pupil initial contour. (c) Localised pupil boundary. (d) Pupil mask.

75 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images

(a) (b)

(c) (d)

FIGURE 3.19: (a) Input image. (b) Iris initial contour. (c) Localised iris boundary. (d) Iris mask.

FIGURE 3.20: Example of segmented iris on noisy image.

For the CASIA images, results are shown in Figure 3.21 for the pupil localisation, Figure 3.22 for iris localisation and Figure 3.23 for the segmented iris.

76 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images

(a) (b)

(c) (d)

FIGURE 3.21: (a) Input image. (b) Pupil initial contour. (c) Localised pupil boundary. (d) Pupil mask.

(a) (b)

(c) (d)

FIGURE 3.22: (a) Input image. (b) Iris initial contour. (c) Localised iris boundary. (d) Iris mask.

77 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images

FIGURE 3.23: Example of segmented iris on CASIA noisy image.

While the results obtained for the UBUIRIS images are shown in Figure 3.24 for the pupil localisation, Figure 3.25 for the iris localisation and Figure 3.26 for the segmented iris.

(a) (b)

(c) (d)

FIGURE 3.24: (a) Input image. (b) Pupil initial contour. (c) Localised pupil boundary. (d) Pupil mask.

78 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images

(a) (b)

(c) (d)

FIGURE 3.25: (a) Input image. (b) Iris initial contour. (c) Localised iris boundary. (d) Iris mask.

FIGURE 3.26: Example of segmented iris on UBIRIS noisy image.

3.4.5 Brief summary on proposed iris segmentation method

Experiments for this section were carried out using non-ideal iris images from three different databases to evaluate the performance of the proposed

79 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images reflection noise removal and iris segmentation approaches. The proposed approach demonstrates robustness in that the system:

(i) effectively removes reflection noise from the pupil region,

(ii) produces accurate localisation of pupil and iris boundaries,

(iii) accurately segments the iris and

(iv) is cross functional for different images suffering from the same challenges.

Table 3.1 shows the speed performance of the proposed iris segmentation approach for the different non-ideal iris databases. The segmentation accuracy is a measure of the successfully segmented images over the total number of images per database.

TABLE 3.1: Performance of proposed iris segmentation method

Database No. of images Noise removal (s) Segmentation (s) Segmentation accuracy (%)

Independant 312 34 50 96.5

CASIA 1639 137 150 97.6

UBIRIS 1877 90 70 96.3

While the proposed experimental approach shows promising results, most of the segmentation failures were due to excessive pupil dilation, eyelash and eyelid exclusion deeming the iris region insufficient for further processing. Inasmuch as the noise removal algorithms such as the morphological operations used here are aimed at removing noise, they can also introduce other challenging ailments within the processed image. For example, selecting a threshold to detect the reflection spots may also affect other parts of the input image. This case is especially witnessed when comparing the UBIRIS original input and processed images shown Figure in 3.27.

80 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images

(a) (b)

FIGURE 3.27: (a) Original input image. (b) Processed image with noise removed.

This is due to the similarity of the pixel intensity values between the white sclera and the white reflection spots within the pupil region. This challenge is one that any algorithm either aims to avoid or manoeuvre, since it is typically inherent for non-ideal images with illumination variations.

Therefore, depending on the objective of the subsequent stage, which is normally the segmentation of the iris and local feature extraction for individual recognition, the noise removal algorithm can be more useful. However, for further processing particularly the extraction of texture features for soft biometrics classification, the careful selection of a threshold that can suit all images is a challenging task.

3.4.6 Feature detection and extraction with phase congruency

This section proposes to identify and detect significant corner features found from the arrangement of the iris patterns, as features for extraction using Phase Congruency (PC). A fusion of this operator with the Harris corner detector is also proposed. This combination robustly localises feature points within the iris image that are not only congruent in phase but are also invariant to rotation. Furthermore with the detected feature points, a compact feature vector of 512 bits containing the exact location of features is easily generated and stored as a template to be used for

81 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images matching purposes.

The remainder of this section is arranged to give a brief description on PC; how it has been employed by other researchers in iris recognition. A summary on Harris algorithm is also presented and how it is fused with PC. The over all proposed algorithm and its implementation; followed by the obtained experimental results, analysis and conclusion is presented.

3.4.7 A brief review on phase congruency

A corner and edge detection method that measures the significance of features in computer images is known as Phase Congruency (PC) [87, 88]. Its invariance to image illumination and contrast makes it a robust and reliable feature detection method.

The structural arrangement of rich features such as lines and edges can offer a good description of an image [88]. Morrone et al.[89] developed a model of feature perception called the Local Energy (LE). The model suggests that features of perceptual significance such as lines and edges within an image are perceived at points where the Fourier components are in phase with each other. It is at these points that the LE is maximal. The model is not based on using local intensity gradients to detect features; hence it is possible to construct a dimensionless measure of PC at any point within the image.

The measure of Phase Congruency as developed by Morrone et al.[89], is defined as the ratio of the Local Energy |E(x)| to the overall path length used up by the local Fourier components in reaching the end point of a signal, and is given by:

|E(x)| PC1(x) = P . (3.21) n An(x)

82 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images where |E(x)| is the magnitude of the vector from the signal origin to the end point. An(x) (with phase angle n, φ is the amplitude of each local, complex valued Fourier components at a location (x) in the signal. However, this particular measure of PC is quite sensitive to noise and does not offer good localisation.

In order to compensate for image noise and good localisation response, Kovesi [87], developed a more advanced measure of PC via Log Gabor wavelets. Log Gabor wavelets offer a large coverage of the frequency space but still maintain a zero DC component in the even symmetric filter. Kovesi [87] states that using PC to mark features within an image is significantly advantageous over gradient based methods. This is because PC;

(i) is a dimensionless quantity;

(ii) uses principal moments of the phase congruency information to determine corner and edge information. A principal moment is a dimensionless quantity refering to that moment wherethe significant features within an image are congruently in phase.

(iii) It is invariant to changes in illumination or contrast and thus

(iv) provides an absolute measure of significant feature points.

(v) It offers accurate feature positioning for instance, phase congruency values are high at edge points and at object boundary and can efficiently be used as a feature detector.

Furthermore, the values vary from a minimum of zero to indicate lack of feature significance, to a maximum of 1 to indicate a very significant feature. This makes threshold specification for feature selection to be much easier even before an image is seen [87]. Above all, PC provides rich texture, edge and structural information that is also in accordance to human vision [90]. Basically, PC is a frequency based-model that looks for points in an image

83 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images where there is a high degree of order in the Fourier domain and described by:

P n W (x)bAn(x)∆φn(x) − T c PC2 = P . (3.22) n An(c) + 

where

∆φ(x) = cos(φn(x) − φ(x)) − sin(φn(x) − φ(x)) . (3.23)

The term W (x) is a factor that weighs for frequency spread. This means that a phase congruency point is of significance if it occurs over a wide range of frequencies. T is the estimated noise influence, hence only the energy values that exceed T are considered. The term ∆φ(x) provides a sensitive measure of phase congruency and φ(x) denotes the weighted mean phase angle.

3.4.8 Phase congruency in iris recognition

Yuan and Shi [91] use 2D PC to extract iris features. They convolve a normalized iris image with a bank of 2D Gabor filters with 4 scales and 6 orientations. They compute the amplitude response at the given scales and orientations, the phase angles and phase deviation measure directly from the filter outputs to finally compute the phase congruency. The resulting phase congruency is down sampled with a 4x4 window. They further concatenate all the rows to a long vector to generate an iris image pattern represented by a vector of 1024 bits. Their approach produces encouraging matching results, with shortfalls due to iris segmentation failure.

Osman [92], employs PC to locate and identify different and significant features found within the iris; which he refers to as “iris minutiae”. His work applies the PC algorithm to eight different angular orientations to cover an entire normalized iris image.

84 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images

For each orientation and location in the image, [92] computes the PC values which are saved as a matrix. However, the computed PC values at feature points are less than one, due to the estimated noise and weighting function. Non maxima suppression and hysteresis thresholding is applied to eight PC images at the different orientation angles. Finally, 8 matrix values of ones and zeros are obtained; where the presence and absence of a significant feature is represented by a 1 and zero respectively. He further locates phase angles corresponding to each feature for each PC image, in which the phase angles are assembled in eight templates for each orientation.

This approach locates the iris features and the corresponding phase angle that determines the feature type. The mean, standard deviation and variance of each template is recorded to form a 24 length feature vector to be used as the iris code. Although experimental results demonstrate the detail of iris features that have been achieved through phase congruency, it consumes many angular orientations to be able to cover a single normalized image.

Du et al.[90] also proposed an iris recognition method based on principal phase congruency (PPC). Their method is a fusion of principal component analysis with phase congruency. This fusion is done to synthesize important information. They use inter-class and intra class fuzzy similarity measure for matching purposes. Their results show a vast range of difference between both classes. They also achieve a correct recognition rate of 98.99% demonstrating the feasibility of their method.

Another work that uses phase congruency in iris recognition is by Patil et al.[93]. Their goal is to evaluate the performance of:

(i) an iris recognition system that uses phase congruency and

85 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images

(ii) content based image retrieval (CBIR) system using precision and recall measures.

For similarity measure between two iris features they calculate the Hamming distance. Their results show a low recognition rate as compared to Daugman’s method. They further state that PC is more efficient for eye images with an uneven illumination.

3.5 Proposed experimental approach

For our proposed approach, the segmented iris first normalised to fixed dimensions of width and height of 200x267, without performing the rubber sheet model. This manner of approach provides the proficiency of tracing the discriminative extracted features onto the original image. The PC is computed with four wavelet scales and six filter orientations to cover the whole image.

The Harris corner detector has been widely used for image matching tasks. While referred to as a corner detector, it does not just select corners; but any location that has large gradients in all directions [94]. The drawback is that when used alone, the response from Harris detector varies with image contrast, hence resulting to difficulties in threshold setting [87]. Furthermore, using Gaussian smoothing to reduce noise manages corrupts the location of corners within the image. Therefore, fusing the two operators provides the greater advantage of generating a feature vector that is objectively invariant to illumination and rotation and moreover rendering corner feature significance and positioning.

3.5.1 Fusing phase congruency and Harris algorithm

Since the important information lies within the iris region, and in order not to risk losing or corrupting any important iris texture information, this section continues to use the segmented iris from the original acquired

86 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images images. Examples of some images from the independently acquired database are shown in Figure 3.28.

FIGURE 3.28: Samples of self-acquired images.

The overall approach and algorithm proposed in this thesis in this Chapter is elaborated by algorithm 3.2 as follows: Algorithm 3.2: Proposed iris segmentation, feature extraction and matching algorithm

Data: u0 = input greyscale image, (h, k) by 3.17, pupil and iris radius by 3.18 and 3.19, pc scales =6,filter orientations=4

Result: φ0 =segmented iris, PC2 localise pupil and iris boundaries; segment pupil and iris by parameterizing 3.20 with calculated diameters and radii; enhance segmented iris with CLACHE; get phase congruency measure by 3.22; enhance PC image with CLACHE; set threshold; if min_moment ≤ threshold then

1; else

0;

detect Harris corner features; match detected corner features; Starting with a segmented iris, the first step is to normalise the iris to fixed dimensions of width and height of 200x267, without performing the rubber sheet model. This approach provides the proficiency of tracing the discriminative extracted features onto the original image. This is followed

87 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images by iris image enhancement by filtering and adjusting its contrast using bottom hat filtering and contrast limited adaptive histogram equalisation (CLAHE). The CLAHE operates on small data regions found within the iris rather than the entire image.

The result is an enhanced image with visually appealing features of interest. This is where regions of adjacent or neighbouring pixels sharing the same intensity values are identified and grouped together. The most connected feature labels are then separated and displayed in an RGB colour scheme for feature articulation as shown in Figure 3.29.

(a) (b)

(c)

FIGURE 3.29: Iris enhancement: (a) Input segmented iris. (b) Enhanced iris. (c) Connected iris features.

The PC computed with four wavelet scales and six filter orientations to cover the whole image, is applied to the enhanced iris. Now, since PC uses principal moments to highlight significant features, a histogram showing the principal moments is computed to determine a threshold value that will be needed to separate insignificant features, as elaborated in Figure 3.30.

88 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images

(a) (b)

FIGURE 3.30: (a) Phase congruency image. (b) Histogram of phase congruency image.

From the histogram of the PC image, values close to zero (0) and one (1) respectively represent less significant and more significant features. Therefore to further articulate more significant features, the obtained PC image is also enhanced using the CLAHE method. Hence plotting the histogram of the newly enhanced PC image, gives an estimate indication of the points that do not lie in important edges; that need to be eliminated through non maximum suppression, as shown in Figure 3.31.

(a) (b)

FIGURE 3.31: (a) Enhanced PC image. (b) Histogram of enhanced PC image.

In order to suppress or eliminate the points that do not lie in important edges, nonmaximum suppression is then applied to the enhanced PC image. This process is followed by hystresis thresholding. Hysteresis thresholding is used for detecting edges by two thresholds.

89 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images

From the enhanced PC image, both the low and high thresholds are selected to 0.6. to detect valuable edges of the iris patterns. Thus, by plotting the histogram of the hysteresis thresholded image, a more efficient distribution of relevant features is demonstrated by principal moments close to 1, as shown in Figure 3.32.

(a) (b)

(c)

FIGURE 3.32: (a) Non-maximum suppresion image. (b) Hysteresis threshold image. (c) Histogram of hysteresis image.

Therefore, in order to retain the most relevant features as seen from the histogram, the process is followed by skeletonisation. The process of skeletonisation in this case helps to fix the non maximum suppression and to retain a skeletal remnant of the connectivity of iris corner features. The skeletonised image is complemented for easier visualisation. Finally, the most significant minimum moments, that is, iris corners that are only congruent in phase, are then extracted to generate a matrix of corner features with their corresponding location as a feature vector. The obtained results are shown in Figure 3.33.

90 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images

(a) (b)

(c)

FIGURE 3.33: (a) Skeletonised image. (b) Complement of skeletonised image. (c) Phase congruency feature vector image.

At this stage the original phase congruency image is used as an input to Harris algorithm where Harris corners have to be detected. This fusion is aimed at confirming that the corner features and their location as detected by the phase congruency operator correspond to those detected by Harris algorithm.

Therefore, plotting the histogram shows the location of Harris strongest features on the phase congruency image. Here, the observation is that the phase congruency feature vector is a replica of the location of strongest features histogram as detected by Harris algorithm. This is shown in Figure 3.34

91 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images

(a) (b)

(c)

FIGURE 3.34: (a) Harris corner features on PC image. (b) Harris corner features on skeletonised image. (c) Histogram of Harris features.

The obtained iris corner feature vector and location of features as detected by Harris algorithm is then stored as a reference template consisting of 218 features and 512 bits per individual. For real time applications, storing the reference template in a database with other corresponding information for instance the name, last name and birth date of the person, means that the individual is now enrolled and can be matched by the recognition system.

Therefore, at a later stage, when the enrolled individual claims a certain identity, the presented iris is swiftly injected through the whole process to be matched against the stored reference template, like in Figure 3.35

92 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images

FIGURE 3.35: Corner points match between reference and query image

The Hamming distance between the reference and query image was calculated to be 0.0588 hence the failure of statistical independence, representing a perfect match.

In order to evaluate the performance of the proposed approach, an attempt to match the corner features between an image from the self acquired database against the UBIRIS image is made. This comparison is made between eye images in these databases because their differences are not so easily observable. Again, the UBIRIS image is injected to go through the whole process as shown by Figures 3.36 to 3.41.

(a) (b)

(c)

FIGURE 3.36: (a) Input segmented iris. (b) Enhanced iris. (c) Connected iris features.

93 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images

(a) (b)

FIGURE 3.37: (a) Phase congruency image. (b) Histogram of phase congruency image.

(a) (b)

FIGURE 3.38: (a) Enhanced PC image. (b) Histogram of enhanced PC image.

(a) (b)

(c)

FIGURE 3.39: (a) Non-maximum suppresion image. (b) Hysteresis threshold image. (c) Histogram of hysteresis image.

94 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images

(a) (b)

(c)

FIGURE 3.40: (a) Skeletonised image. (b) Complement of skeletonised image. (c) Phase congruency feature vector image.

(a) (b)

(c)

FIGURE 3.41: (a) Harris corner features on PC image. (b) Harris corner features on skeletonised image. (c) Histogram of Harris features.

95 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images

The results are shown in Figure 3.42, where there are no valid matching points between the UBIRIS reference and self-acquired query images.

FIGURE 3.42: Failed corner points match between UBIRIS reference and self-acquired query image

3.5.2 Concluding summary

This chapter presented the methodology proposed in this thesis to achieve accurate iris segmentation, feature extraction and matching from non-ideal eye images with smooth pupillary and limbic boundaries. It has also been learned that typical standard devices used under poorly supervised image capturing conditions is indeed a challenge that can threaten the performance of an iris recognition system.

The method proposed in this chapter however, demonstrated that with a modest fusion of befitting algorithms, that is, Bresenham’s and Chan-Vese algorithms for segmentation, as well as Phase Congruency and Harris algorithm to extract iris corner features, the challenges can be efficiently controlled. Compared to the approaches used by other researchers, the proposed method offers the following advantages and contributions:

(i) There is no need to perform noise removal in order to achieve accurate iris segmentation. In fact, the reflection noise can also prove as a useful region to initiate the evolving contour of the Chan-Vese algorithm.

96 Chapter 3. Proposed methodology on iris segmentation, feature extraction and matching from non ideal eye images

(ii) The point mentioned above means that accurate iris segmentation can be achieved faster; with less in-between tasks needed and hence low implementation and computational complexities.

(iii) Due to the fewer modules needed to perform segmentation, the risk of degrading the quality of iris features is highly reduced.

(iv) The computation of pupil and iris diameters through Bresenham’s algorithm for boundary localisation consumes less time and can be easily implemented across different databases, as opposed to the use of the traditional Hough transform.

(v) The fusion of phase congruency and Harris algorithm to detect corner features found within the arrangement of iris patterns, produces a feature vector with the exact location of corner features that are not only congruent in phase but are also invariant to illumination and rotation.

(vi) The proposed experimental approach generates a compact feature template of 512 bits per individual, as opposed to 1024 bits used by state-of-the-art algorithms, hence requiring low storage space.

97 Chapter 4

Ethnicity distinction and classification from iris images

The previous chapter presented the experimental approach proposed in this thesis to perform fast, accurate and real time iris segmentation; corner feature detection, extraction and matching from non-ideal images. This chapter continues to utilise the independently acquired database in order to investigate and address the objective of achieving ethnic distinction and classification from the global iris texture patterns. Section 4.1 introduces the concept of soft biometrics; the motivation and the contribution of this chapter. Section 4.2 presents the related work done by other researchers on ethnic prediction and classification from iris images. Section 4.3 concludes and summarises the related work. Section 4.4 presents the experimental approach proposed by this research work and the results obtained on ethnic classification. Section 4.5 is a concluding summary of this chapter.

4.1 Introduction

Unlike the name and surname of an individual, biographic data such as age, ethnicity and gender are attributes collectively referred to as "Soft Biometrics". Such attributes are used in our daily lives with an aim of providing further or more specified information about an individual. For a person enrolling to any biometric system, soft biometric information is normally captured as additional labels. This is because the automated

98 Chapter 4. Ethnicity distinction and classification from iris images detection of soft biometrics is generally hard [95].

Unimodal biometric systems, like the iris recognition system (IRS), palm or the fingerprint recognition system for instance, are systems that use a single trait for recognition. These types of systems are often affected by several practical problems like noisy sensor data, lack of universality, lack of acceptance and lack of subject cooperation leading to high error rates in the identification and verification of individuals. Unimodal systems are also prone to spoofing attacks [96] as one of the challenges.

The prediction and classification of soft biometrics from iris images is a young and developing topic that has become very important in iris biometrics. This is because generally, the categorisation of such information is always presumably based on the physical appearance and behaviour of an individual. With today’s advancing technologies in human genetics and aesthetic medicine, it has become quite easy for one to alter or disguise their inherently born identity under any particular kind of soft biometric. A more global case is witnessed in the use counterfeit identity cards by under aged teenagers to adopt a certain age as a measure of gaining access to certain premises.

According to Ricanek and Barbour [97] ethnicity and gender are soft biometrics emerging fields that explore and ask new questions, such as “Can certain traits be leveraged to determine age or ethnicity?”, ’Can they reveal intent or deception?’

Aside from the local iris features that are generally use for individual recognition, the aim of this chapter is to investigate other iris features that can be leveraged to determine the distinction and classification of ethnicity between two ethnic groups; that is, the black and white population from the South African hemisphere. The investigation will be done through conducting various experiments with an objective to provide a solution to

99 Chapter 4. Ethnicity distinction and classification from iris images the question posed by Ricanek and Barbour [97].

This research chapter is particularly motivated by the long standing argument that iris features are only viable for recognition purposes and have no genetic penetrance or relation to neither ethnicity nor gender. Another motivation is that from all the available literature, investigations and experiments on ethnic classification from iris images have only been conducted using mostly Asian and European participants. Furthermore, ethnic and gender prediction and classification from iris images is still a new research topic and so far, very little work and very few experiments have been conducted to explore this issue. Through recent studies however, several researchers have reported that iris texture features contain information that is inclined to human genetics and is highly discriminative between different eyes of different ethnicities and gender. Also, when bringing this observation to daily occurrences, people who belong to the same ethnic group tend to have or to possess the same iris primitives such as colour and textures. Also, except of course for a few special cases, genetically, parents belonging to the same ethnic group are more likely to birth a child with a typically related iris pigmentation.

This thesis therefore aims to contribute to the body of knowledge by conducting experiments using a population that has never been investigated in literature; in order to approve or disapprove the distinction and classification of ethnicity from iris features. The use of the proposed population for this thesis’ investigation is the first of its kind to contribute towards existing literature studies. The reasons are: there is no iris database available to the research community to cater for this population. Also, the detection and classification of individuals from soft biometrics is still a difficult and challenging task that requires more research [95], especially from other ethnic groups other than the ones already investigated.

100 Chapter 4. Ethnicity distinction and classification from iris images

Section 4.2 reviews the related work and presents the databases, feature extraction methods and the experimental results obtained over the recent years by other researchers in the field. Table 4.2 presents a detailed overview of how research in ethnic prediction using iris images has evolved.

4.2 Related work on ethnicity prediction and

classification

Qui et al.[21] used global texture information of iris images to develop a novel ethnic classification method. Their work argues that " iris texture is related to race". From investigating Asian images from the Chinese Academy of Sciences Institute of Automation (CASIA v2), non-Asian/ European images from the University of Palackecho & Olomopuc (UPOL) and from the University of Beira Interior (UBIRIS), the assumption and motivation of their work is that the characteristics of iris patterns are different.

Their proposed design of 2D Gabor filters, uses four orientations (θ) values

π π 3π of 0, 4 , 2 , 4 and six frequencies combined with 10 space constants. Their proposed design obtains 240 pairs of Gabor channels. In each image point, the outputs of the even and odd symmetric Gabor filters are combined into a single quantity called the Gabor Energy (GE). By calculating the average of the GE values for regions A and B, two statistical features namely GE and Gabor Energy Ratio (GER) are extracted and combined to characterise the global texture information of the ROI.

Using the AdaBoost algorithm for classification, experimental results report a correct ethnic classification rate of 79.44% using the GE features, and 84.95% correct classification rate (CCR) from the GER features. The combination of both the GE and GER features is reported to achieve an increased overall correct classification rate of 85.95%. In addition, they

101 Chapter 4. Ethnicity distinction and classification from iris images state that region A has a rich texture for Asians than region B. For non-Asians however, both regions are reported to closely have the same rich texture. Qui et al.[21] blame classification errors on the noisiness of the images, eyelid and eyelash occlusions on the ROI and the illumination differences between the images from the different databases.

From the reported results, Qui et al. maintain that:

(i) global iris features seem to be genetically dependent;

(ii) the total statistical measurement of iris texture details is related to genetics on a larger scale rather than on the smaller scales employed by traditional iris recognition algorithms;

(iii) global iris features tend to be similar for a specific race, hence appropriate to achieve ethnic classification.

Although they achieve a reasonable CCR, there is a potential bias accompanying the used iris databases for each class. All the used Asian images come from the CASIA database with Asian subjects only. The UPOL and UBIRIS database consists of images from European subjects only. The trained algorithm could be separating images based on illumination differences rather than on iris texture. Furthermore, the concern associated with the prediction model learning to understated the differences between the iris databases is not addressed.

Aiming to improve the previous method, Qui et al.[22] later continued the research using a different approach. The continued research proposes an automated method to classify ethnicity based on "learning the primitive appearance from iris images". The different approach is based on investigating the discriminative types of visual primitives from iris images that make Asians to be different from non-Asians. In order to conduct their experiments Qui et al. learn the vocabulary of Iris Textons. Textons are fundamental representations of micro structures or visually perceived

102 Chapter 4. Ethnicity distinction and classification from iris images textures in natural images [98].

Different from the previous approach, 2D Gabor filters are designed with eight orientations and five scales. The, response vectors from the filtered image are clustered into small sets of prototypes using the K-means clustering algorithm. The K-centers of each cluster are considered as iris textons while the remaining filter response vectors are considered as appearance vectors. The support vector machines (SVM) classifier is used to train the iris texton histograms.

An increased ethnic classification overall accuracy of 91.02% between Asian and non-Asians is reported using the newly proposed method. Qui et al. further state that half of the of the inner iris region often provides more textural information for Asian images than the outer half of the iris region. For non-Asians however, both regions contain almost similar textural information having capillary like patterns being the main patterns.

The work done by Stark et al.[23] uses human observations to categorise iris images based on similarity in texture and overall appearance. The motivation of their work is based on the notion that a database search could be faster if images of similar appearances could be separated into suitable and relevant categories. They further state that such a study can be used as a guide to create an automated image analysis system that has the capability to categorise iris images by appearance.

The main objective of their work is to determine a consensus across evaluating subjects that iris images of similar texture patterns should belong to the same category. This means a strong validation is based on a large number of subjects categorising the same images into the same category. The ND-Iris-0405 database consisting of 100 hand-picked and well-focused iris images from 100 different individuals is used. The

103 Chapter 4. Ethnicity distinction and classification from iris images selected images represent 26 Asian males and 24 Asian females, as well as 26 Caucasian males and 24 females.

Given a certain set of instructions to follow, 21 human subjects are asked to evaluate and categorise the images. Experimental results report that most human subjects categorised the images based on the level of texture detail. Although the categories identified by the subjects are reported to have projected high levels of similarities, strong levels of dissimilarities are also reported. For ethnicity, a 62.1% average difference between Caucasians and Asians was achieved. The element of gender was also considered and results show a split of 42.5% to 57.5% hence giving an average difference of 15% between males and females.

According to records, Stark et al.[23] are among the first to propose a study of using human subjects for iris categorisation. From their study, ethnic categorisation proved to be a better success than gender categorisation. They further acknowledge that the topic of using iris textures to determine soft biometrics of individuals still needs more attention.

The work of Lagree and Bowyer [24] proposes the analysis and possible use of iris texture features to predict ethnicity. Their motivation is based on the assumption that a particular ethnic group might have the same iris texture features which differ from one ethnic group to another. They also state that a biometric system that has the capability to recognise the ethnicity of individuals can provide automatic classification without human intervention. The objective of their work is to determine the accuracy at which ethnicity can be identified based on iris textures.

Using images from the University of Notre Dame consisting of 600 Asian images and another 600 from European images; their work proposes to eliminate the challenge of biased result by randomly dividing the whole

104 Chapter 4. Ethnicity distinction and classification from iris images dataset into 10 folds of 120 images each. For iris segmentation and normalisation, the IrisBee software is used. The normalised image is divided into ten-4-pixel horizontal bands in order to compute a feature vector that describes the texture of an image. Six different filters are applied at every non-masked pixel location of the normalised image. In addition to the six filters, two more filters referred to as S5S5 and R5R5 are created using Laws’ Texture Measures.

For classification, the obtained feature vectors are used as input to the sequential minimum optimisation (SMO)) algorithm. Their proposed method achieves 90.58% ethnic classification accuracy, and an increased accuracy of 96.17% without person disjoint.

In 2011, Zhang et al.[25] conducted an experiment with an objective of classifying Asian and Non-Asian individuals from iris images. Their work is motivated by the notion that there is a certain degree of correlation and consistence in iris texture patterns belonging to the same race. Their proposed method combines supervised codebook optimisation and Locality-constrained Linear Coding (LLC).

In conducting their experiment, the first database referred to as (DB1) consist of images from Asian participants and non-Asian participants from the UPOL database. The second database (DB2) is a combination of Asian images taken from the CASIA Iris-Lamp database and those from DB1. From each of the databases (DB1 and DB2), 500 images from both Asian and non-Asian classes are used for codebook learning.

Their proposed method uses scale invariant feature transform (SIFT) descriptor to extract low level visual features from the normalised iris. For iris image representation, the codebook optimisation combined with LLC is used. Based on the extracted features, the K-means method is used to learn the initial codebook. The label information learned during codebook

105 Chapter 4. Ethnicity distinction and classification from iris images optimisation is used to choose codes for the classification task as well as the mutual codes for quantisation error retention.

From their results using DB1, a CCR of 96.7% with an equal error rate (EER) of 3.4% is achieved. From DB2 a CCR of 96.67% with an EER of 3.35% is achieved. They further compare their obtained results with the Iris Textons method from Qui et al.[22] by using a similar region of interest (ROI). From DB1, Zhang et al.’s method obtains a CCR of 94.28% and an EER of 6.36% which is a significant increase from results obtained by Qui et al. From DB2, 94% CCR and 6.48% EER is achieved compared to the 85.58% CCR and 19.48% EER by Qui et al. in [21, 22]

The objective of the work conducted by Zarei & Mou [26] in 2012, is to predict the ethnicity of individuals from iris textures using artificial neural networks. The motivation for using the human iris for this purpose is its stability over the progressing age of an individual, the effortless and non–intrusive manner in which eye images are acquired. They also state that the prediction of ethnicity through the iris will provide the efficiency of separating large data into subcategories thus fast tracking the search for an individual within a large database.

The University of Notre Dame database from Asian and non-Asian participants is used to conduct the experiment. Segmentation and normalisation of the is performed using the Iris BEE software. The proposed architecture of the network consists of two networks each with three layers; network 1: for person disjoint, meaning an individual’s image used for testing will not be used for training and validation; Network 2: for non-person disjoint, meaning the same individual’s distinct images are used in training/validation and testing. After attempting several network topology trials, they configure the input, hidden and output layers to have 882 neurons, 10 neurons and 1 neuron respectively.

106 Chapter 4. Ethnicity distinction and classification from iris images

TABLE 4.1: Experimental results

Computed Network 1: Person Disjoint Network 2: Non-person Disjoint Min MSE 0.043 at 37 time period 0.011 at 110 time period Caucasian Asian Class 1 Class 2 Class 1 Class2 CCR 113/120 111/120 110/113 124/127 TPR 94.2% 92.5% 97.3% 97.6% FPR 7.5% 8% 3% 2% Overall Accuracy 93.3% 97.5%

The best results are reported from the use of 10 neurons in the hidden layer. The computed minimun mean square error (MSE), correct classification rate (CCR), true positive (TPR) and false positive ratio (FPR) between the two classes as well as the overall accuracy over the two networks are summarised in Table 4.1.

Zarei and Mou’s approach of using neural networks for ethnic prediction through iris images demonstrated an improved accuracy performance rate. From the results, they iterate that using network 2 (non-person disjoint) introduces a potential bias to the overall performance of the model.

4.3 Summary of conclusions on related work

Table 4.2 shows a comparison of the different databases and techniques used by researchers over the recent years to predict and classify ethnicity based on iris textures. The percentages of the overall accuracies are based on person disjoint.

From Table 4.2, the method of Zhang et al.[25], shows to have achieved the highest ethnic classification accuracy. Although they utilise a different algorithm to perform classification, their approach is closely related and consistently compared to that of Iris Textons by Qui et al.[21, 22] being the highest at the time. Considering the use of human subjects to conduct iris categorisation, the understandable lowest performance is observed from Stark et al.’s [23] approach.

107 Chapter 4. Ethnicity distinction and classification from iris images

TABLE 4.2: Comprehensive comparison

Authors Year Database Ethnicity Technique CCR

Qui et al. 2006 CAS,UP,UB Asian, non-Asian Gabor filters 85.95% Qui et al. 2007 CAS,UP,UB Asian, non-Asian Iris Textons 91.02%

Stark et al. 2010 ND-Iris Asian, Caucasian Human subjects 62.1%

Lag,Bow 2011 Notre Dame Asian, Caucasian 9 Filters 90.58%

Zhang et al. 2011 CAS,UP Asian, non-Asian LLC 96.7%

Zarei, Mou 2012 Notre Dame Asian, Caucasian Neural nets 93.3%

The similarity from all the research work done however, is the utilisation or the classification of ethnicity from only Asians and non-Asians or Caucasian participants. This is a first indication of necessary future work needed to accommodate other ethnicities across the globe. It is also important to not be derailed by the quantity of the accuracy achieved by a certain method but to focus more on quality research with an aim of improving the current and existing ethnic classification methods through comprehensive and duplicative work being reported and methods that have been applied.

4.4 Proposed experimental approach on ethnic

classification

It has been well established that the typical iris recognition system can only identify an enrolled individual to declare him/her as an authentic or imposter. However, for unenrolled individuals, a non-match is simply returned as a response. The proposed experimental approach is also motivated by the observation made from the related work that; so far, research and investigations focused on using iris textures to determine ethnicity, have only been conducted using Asian and Caucasian

108 Chapter 4. Ethnicity distinction and classification from iris images participants from Asia and Europe respectively.

This thesis is the first work to explore ethnic differences from iris images between African black and white participants. Although it is reported that ”most people who live in Asia and Africa have dark brown eyes" [99]; however, the texture patterns and primitive appearance within the iris differ vastly; hence acquired images are captured as monochrome without the detail of iris colour.

The aim of this section is to investigate the different iris textures from the black and white ethnic groups with an objective to:

(i) Develop an automated model that effectively and effortlessly extracts the rich iris textures that are inherent but also distinctive and distinguishable to each ethnic group.

(ii) Use the extracted texture features to evaluate the degree of ethnic distinction and classification.

(iii) Use various classifiers to evaluate the validity of the extracted iris texture features.

(iv) Use the same features to also investigate the feasibility of gender prediction and classification between the two ethnic group.

4.4.1 Gabor filters for texture extraction

While the traditional features used for recognition are based on minute local features, the global iris textures are not taken into account. According to Zhou [100], “Texture provides a rich source of information about the natural scene. For computer scientists, it provides a key to understanding basic mechanisms that underlie human visual perception. A fundamental goal of texture research in computer vision is to develop automated computational methods for retrieving visual information and understanding image content based on textural properties in images. The critical issues in realising the goal include

109 Chapter 4. Ethnicity distinction and classification from iris images understanding human texture perception and deriving appropriate quantitative texture descriptions.”. Wang and Mu [101], also support that the extraction of relevant features for soft biometrics detection, is the foundation that will directly affect the capability of a classifier to perform diligently.

The method used here, exploits texture features within the iris, ones that are not used for recognition. This is done by employing multi-channel Gabor filter banks, because Gabor filters employ frequency and orientation representations which make them effective for texture representation and discrimination [102]. In the spatial domain, a 2D Gabor filter is a Gaussian kernel modulated by a sinusoidal plane wave [40] and is represented as:

( " #) 1 1 x2 y2 h(x, y) = exp − 2 + 2 exp{j2πF x} (4.1) 2πσxσy 2 σx σy where σx/σy is the Gaussian aspect ration, the complex exponential has F and θ as the spatial frequency and orientation respectively. Jain and Farrokhnia [103] implement Gabor filters for texture analysis using only the real or even component. For this purpose, the filter impulse response is represented as:

( " #) 1 1 x2 y2 h(x, y) = exp − 2 + 2 cos{2πF z} (4.2) 2πσxσy 2 σx σy

The response of the filter captures both the magnitude and phase in the entire frequency spectrum. Clausi [40] states that there are two special considerations to be made when generating texture features using multi-channel filters. They are:

(a) careful selection of filter characterisation and

(b) performing feature extraction of the raw filter output to improve the features set.

Depending on the specific implementation of Gabor filters, the tuning frequency or tuning period or wavelength establishes the kind of sinusoidal wave that the filter will respond best. Therefore, positioning of

110 Chapter 4. Ethnicity distinction and classification from iris images wavelengths and orientations must be carefully established to properly capture textural information [40]. Clausi [40] and Zheng et al.[104] state that in order to create unique texture signatures, there are different methods that can be applied to the outputs of the Gabor filters. Further, that a careful analysis of the information produced by the Gabor filter outputs provides appropriate texture features. Information that can be used as features from the filter outputs include, the magnitude response, phase information, mean amplitude, local energy and orientation [40, 104].

4.4.2 Proposed design for iris texture extraction

For this work, the iris image is not normalised with the traditional rubbersheet model, instead, it is scaled and normalised to fixed dimensions of 123 x 127. This is done to preserve the original textures and the geometric formation within the iris region.

The proposed model employs two designs of Gabor filter banks. The first design uses three wavelengths (λ) of 3, 5, and 7 pixels per cycle as well as five orientations (θ) of 0, 30, 60, 90, 120 degrees. The second design uses three wavelengths of 9, 11, and 13 with the same orientations.

The resulting Gabor envelope for the first design is shown as an example in Figure 4.1 for the different wavelengths and orientations. The Gabor envelope helps to inform how an image will be texturally represented and rotated at the different chosen designs.

111 Chapter 4. Ethnicity distinction and classification from iris images

FIGURE 4.1: Gabor envelope at different wavelenghts and orientations.

The choice of parameters is motivated by running a few trials of how dense we want the sinusoidal wave (modulated with the input signal) to be. Smaller wavelengths give a denser wave while larger wavelengths result in larger sinusoidal waves. By varying the orientation, we can look for texture oriented in a particular direction. The design of both filter banks consists of 3 wavelengths x 5 orientations to produce a compact Gabor array of size 1x15. The resulting Gabor array is convolved with the segmented and enhanced iris image at each wavelength and orientation as shown in Figure 4.2 for the first design.

FIGURE 4.2: Iris image convolved with Gabor array.

112 Chapter 4. Ethnicity distinction and classification from iris images

The result is the magnitude and phase of the Gabor filter bank for the input image, shown in Figure 4.3. From the magnitude and phase components, the mean amplitude (MA) and local energy (LE) are computed and horizontally concatenated as a texture feature vector of size 1x30 per left iris of an individual, achieved by implementing algorithm 4.1:

Algorithm 4.1: Gabor texture feature extraction and computation 1 2 Data: u0 = Segmented enhanced input image, λ =[3 5 7], λ =[9 11 13], θ = [0 30 60 90 120] Result: Feature vector=Horzcat(LE,MA) Initialisation; GaborArray = gabor(λ,θ);

[mag,phase] = imgaborfilt(u0,GaborArray); q LE= mag2(p) + phase2(p); MA=abs(mag(p) + phase(p));

FIGURE 4.3: Gabor magnitude and phase for one iris image.

Once the features have been extracted from each group, the texture feature vector is further normalised to zero mean and one unit variance, to eliminate/reduce redundancy by equation 4.3:

x − x¯ Z = (4.3) σ

113 Chapter 4. Ethnicity distinction and classification from iris images where x and x¯ represent the original feature vector and the mean of the vector respectively, and σ is the standard deviation. Since the original signal (feature vector) is positive, the absolute value is computed to revert to the signal originality using equation 4.4:

K = abs(Z) (4.4)

The average of the feature vectors which include; 15 local energy (LE) features and 15 mean amplitude (MA) features, is computed for all subjects per ethnic group using equation 4.5.

n 1 X G = ∗ K (4.5) LE n i i=1 where GLE is the average of the LE features and n is the number of the LE features. The same computation is done for MA features. The comparison of the average difference between the two ethnic groups using LE and/or MA features is finally computed using equation 4.6, where MA is with reference to mean amplitude features in equation 4.6, B and W represent the black and white participants respectively. The same computation is carried out for the comparison of LE features.

1 MA = G B − (G B + G W ) (4.6) B MA 2 MA MA

4.4.3 Experimental results and analysis

Our database has 84 participants constituting of 22 black males, 22 black females, 18 white males and 22 white females. From each individual, two images of each left and right eyes were acquired, to give a total of 336 images. With poorly acquired images and make up challenges from female participants our sample size droppped to 78 participants to finally have 17 white males and 17 black females with usable images. A total 301 images were accurately segmented using the approach proposed in Chapter 3.

114 Chapter 4. Ethnicity distinction and classification from iris images

The first design having three wavelengths of (λ) of 3, 5, and 7 pixels per cycle as well as five orientations (θ) of 0, 30, 60, 90, 120 degrees is used on a smaller number of subjects consisting of 6 black males and 5 white females. This is followed with a gradual increase of subjects until all subjects in each group have been tested. The obtained results showing ethnic distinction between the two ethnic groups are shown in Figures 4.4 and 4.5.

FIGURE 4.4: Ethnicity distinction using Gabor mean amplitude features.

FIGURE 4.5: Ethnicity distinction on larger group.

From the achieved results, two observations are made; (i) black males fall on the negative of the z-plane while white females fall on the positive. (ii) At this stage, the distinction between both ethnicities is most clearly witnessed at the lower wavelengths of only the mean amplitude features. However, another possible option here, is that a gender distinction might be taking place instead of ethnic classification, since the used subjects also

115 Chapter 4. Ethnicity distinction and classification from iris images have opposite genders.

To investigate this possibility, the same design is hence implemented for black females vs white males. Here, both the performance of mean amplitude and local energy features are compared. The obtained results are shown in Figures 4.6 and 4.7.

FIGURE 4.6: Mean amplitude and local enegy features showing no distinction.

FIGURE 4.7: Mean amplitude and local enegy features showing no distinction.

From this group, a different pattern is observed. Unlike the first group, the distinction between the two ethnicities is not as clear as in Figures 4.4 and 4.5. In fact, the observation is that with the continuation of higher wavelengths, more distinction is shown for both local energy and mean amplitude features.

116 Chapter 4. Ethnicity distinction and classification from iris images

A few test designs on the Gabor filter bank were conducted while varying the wavelength such that the second design having three wavelengths of 9, 11, and 13 was optimal for all the ethnic groups for both the LE and MA features. Figure 4.8 shows the iris image convolved with the second filter bank design. The second design demonstrates efficient articulation of the texture features than the first proposed design.

FIGURE 4.8: Iris image convolved with improved design of Gabor array.

Resuming with the first group of black males vs white females using the second design; the achieved results are shown for the performance of both the mean amplitude and local energy features in Figures 4.9 and 4.10.

FIGURE 4.9: Mean amplitude and local energy features with improved design.

117 Chapter 4. Ethnicity distinction and classification from iris images

FIGURE 4.10: Mean amplitude and local energy features with improved design.

The second design used on black females vs white males also achieves robust results as shown in Figure 4.11.

FIGURE 4.11: Ethnic distinction with improved Gabor design.

Finally all the MA and LE features for all the groups are combined to evaluate and ascertain the performance of ethnic distinction instead of "possible gender interference". Achieved results are shown in Figure 4.12.

118 Chapter 4. Ethnicity distinction and classification from iris images

FIGURE 4.12: Ethnic distinction between all black and white participants.

The results obtained with the second design of the Gabor filter bank is able to demonstrate:

(i) relevent texture feature extraction from the different orientations and wavelengths within the iris region.

(ii) robustness in texture discrimination and ethnic distinction between the tested subjects.

(iii) effective use of both local energy and mean amplitude features to achieve ethnic distinction.

(iv) competence for use in classification purposes.

Another crucial overall observation from the achieved results is that although the second Gabor filter bank provides a clear distinction between the two ethnic groups, there is also another visible and automated distinction between the respective genders. While the MA and LE features from both ethnicities are cleary unique, the males and females from each ethnic group are also clearly differentiated. This discovery, is given more attention in Chapter 5.

4.4.4 Ethnic classification from extracted Gabor features

This sub section continues to use the extracted texture features to validate the results obtained in the previous section. Here, different classifiers with

119 Chapter 4. Ethnicity distinction and classification from iris images seeding are used to further confirm the robustness of the proposed model. Tabulated results comparing the performance accuracies of the different classifiers used are also presented.

The features used for classification purposes include, the raw LE and MA as individual features computed directly from the filter outputs for the different ethnic groups. This is followed by classifying the LE and MA features after normalisation. Furthermore results combining both LE and MA features as one feature vector are also presented. The comparison of ethnic groups are arranged as per the previous section.

Table 4.3 presents ethnic classification accuracy between black males versus white females with the different classifiers that have been implemented. The validation methods varying from 5 fold cross validation, 20% hold out validation and no validation, together with their respective performance accuracies are also shown.

With a hold out validation of 80 % for training data and 20% testing data, the performance of the raw LE features achieves 85% ethnic classification accuracy; and MA features achieve a 87.5% ethnic classification accuracy from the different classifiers.

Using the normalised data however, shows an overall better performance with a classification accuracy of 96.5% for both features. The accuracy percentage is the most highest with no validation method. The overall accuracy with hold out validation and all features considered, results to 96.9% and 98.5% with no validation.

120 Chapter 4. Ethnicity distinction and classification from iris images

TABLE 4.3: Ethnic classification between black males and white females

Data Type Feature Type Method Accuarcy Classifier Raw Local Energy Cross Validation 80.30% Cosine KNN Raw Local Energy Hold Out 85.00% Cosine KNN Medium Tree Raw Local Energy No Validation 93.90% Complex Tree Raw Mean Amplitute Cross Validation 79.70% Quadratic SVM Fine KNN Cosine KNN Raw Mean Amplitute Hold Out 87.50% Weighted KNN Ensemble subspace KNN Complex Tree Raw Mean Amplitute No Validation 93.90% Medium Tree Normalised Local Energy Cross Validation 85.50% Fine Gaussian SVM Complex Tree Medium Tree Simple Tree Normalised Local Energy Hold Out 96.50% Logistic regression Cubic SVM Fine Gaussian SVM Ensemble Bagged Trees Complex Tree Medium Tree Normalised Local Energy No Validation 97.20% Simple Tree Cubic SVM Linear Discriminant Normalised Mean Amplitude Cross Validation 87.30% Cosine KNN Logistic regression Quadratic SVM Normalised Mean Amplitude Hold Out 96.50% Cubic SVM Fine Gausian SVM Ensemble subspace KNN Complex Tree Normalised Mean Amplitude No Validation 97.20% Medium Tree Simple Tree Raw All Features Cross Validation 90.50% Quadratic SVM Cubic SVM Raw All Features Hold Out 95.90% Cosine KNN Complex Tree Raw All Features No Validation 97.20% Medium Tree Normalised All Features Cross Validation 91.80% Quadratic SVM Ensemble Subspace KNN Ensemble Bagged Trees Weighted KNN Normalised All Features Hold Out 96.90% Cubic KNN Cosine KNN Fine KNN Linear Discriminant Normalised All Features No Validation 98.50% Ensemble Bagged Trees

121 Chapter 4. Ethnicity distinction and classification from iris images

Table 4.4 presents ethnic classification between black females and white males. The results achieved here are also in agreement with the results from Table 4.3, showing that the use of normalised data for both LE and MA features respectively show better performance with 96.3% and 95.3% accuracy. The overall accuracy with hold out validation and all features considered, results to 96.9% and 98.5% with no validation.

TABLE 4.4: Ethnic classification between black females and white males

Data Type Feature Type Method Accuracy Classifier Raw Local Energy Cross Validation 79.4% Ensemble Subspace Discriminant Ensemble Subspace KNN Raw Local Energy Hold Out 83.3% Ensemble Bagged Trees Complex Tree Raw Local Energy No Validation 92.50% Medium Tree Quadratic SVM Raw Mean Amplitute Cross Validation 76.50% Logistic Regression Complex Tree Medium Tree Simple Tree Linear SVM Raw Mean Amplitute Hold Out 85.30% Medium KNN Cosine KNN Cubic KNN Ensemble Subspace Discriminant Ensemble bagged trees Raw Mean Amplitute No Validation 91.20% Cubic SVM Normalised Local Energy Cross Validation 85.30% Quadratic SVM Normalised Local Energy Hold Out 96.30% Quadratic SVM Normalised Local Energy No Validation 97.10% Quadratic SVM Normalised Mean Amplitude Cross Validation 87.60% Quadratic SVM Normalised Mean Amplitude Hold Out 95.30% Cubic SVM Normalised Mean Amplitude No Validation 97.10% Quadratic SVM Raw All Features Cross Validation 91.2% Ensemble Subspace Discriminant Linear SVM Raw All Features Hold Out 93.30% Logistic Regression SVN Ensemble Subspace Discriminant Raw All Features No Validation 97.10% Quadratic SVM Quadratic SVM Normalised All Features Cross Validation 90.60% Cubic SVN Normalised All Features Hold Out 96.90% Ensemble Subspace Discriminant Normalised All Features No Validation 98.50% Ensemble Bagged Trees

122 Chapter 4. Ethnicity distinction and classification from iris images

Table 4.5 combines both black and white ethnicities with hold out validation and all features considered to achieve an overall accuracy of 96.9% and 98.7% with no validation.

TABLE 4.5: Ethnic classification between all black and white participants

Data Type Feature Type Method Accuracy Classifier Raw Local Energy Cross Validation 89.50% Quadratic SVM Raw Local Energy Hold Out 93.00% Quadratic SVM Ensemble Bagged Trees Raw Local Energy No Validation 98.700% Ensemble RUSBoosted Trees Simple Tree Fine Gaussian SVM Raw Mean Amplitute Cross Validation 92.80% Fine KNN Weighted KNN Linear SVM Cosine KNN Raw Mean Amplitute Hold Out 93.30% Weighted KNN Ensemble Bagged Trees Raw Mean Amplitute No Validation 98.70% Ensemble Bagged Trees Normalised Local Energy Cross Validation 90.80% Quadratic SVM Normalised Local Energy Hold Out 73.30% Ensemble RUSBoosted Trees Complex Tree Normalised Local Energy No Validation 97.40% Medium Tree Normalised Mean Amplitude Cross Validation 79.50% Quadratic SVM Logistic Regression Normalised Mean Amplitude Hold Out 80.00% Quadratic SVM Weighted KNN Nomalised Mean Amplitude No Validation 91.00% Quadratic SVM Raw All Features Cross Validation 90.90% Cosine KNN Linear discriminant Raw All Features Hold Out 93.30% Linear SVM Complex Tree Raw All Features No Validation 96.600% Medium Tree Normalised All Features Cross Validation 96.20% Quadratic SVM Complex Tree Medium Tree Normalised All Features Hold Out 96.90% Logistic Regression Quadratic SVM Ensemble Subspace KNN Complex Tree Normalised All Features No Validation 98.70% Medium Tree

123 Chapter 4. Ethnicity distinction and classification from iris images

4.5 Concluding summary

This chapter presented an automated method for extracting and analysing useful and relevant iris texture features to determine the distinction between black and white ethnic groups from the African continent. Being the first of its kind, the main objective was not to address individual identification but rather to design, analyse and use Gabor features that efficiently provide differentiation and classification between two racial groups. The design of two different Gabor filter banks convolved with enhanced iris images has been used to capture the details of global texture features from iris images. The second design having wavelengths of 9 ,11 and 13 proved to be optimal for both the local energy and mean amplitude features to be used effectively either as individual features or combined features for all the ethnic group.

Although the sample size may be lower than the one used in previous studies by other researchers, when compared to the work of Stark et al. who also had a lower sample size, the proposed model demonstrates a 40% increased accuracy. Unlike most researchers who either used good quality images and software toolkits to perform iris segmentation, the iris images used here have been segmented with a model also proposed and presented in this work.

The promising results shown by the proposed model also come with low computational complexities, making it easier for real time deployment and integration to an existing IRS. The use of different classifiers also demonstrated the robustness of the proposed model design which not only showed potential for ethnic classification, but also for gender classification. Furthermore, the discoveries articulated by the researchers confirm that iris texture is indeed related to race. It also proves true that, ethnic distinction lies on the coarse scale texture features rather than on minute local features used for iris code generation in iris recognition.

124 Chapter 5

Gender prediction and classification from iris images

5.1 Introduction

An automated method that extracts iris texture features to achieve ethnic classification from iris images was proposed and presented in Chapter 4. From the results, feasible classification of gender was detected. Similar to ethnic classification from iris images, the prediction and classification of gender from iris images is also a new, challenging and an important topic in iris biometrics and in computer vision generally. Teaching or training a computer to use an image for purposes of categorising an individual as male or female is a difficult yet possible task.

Usually, for any biometric system, system administrators are responsible for enrolling individuals. The drawback however, is that to gather or confirm information such as gender; additional documentation such as a birth certificates has to be provided by the enrolling individual. This means the system administrator heavily relies on the individual to provide such data accurately. Even when the provided soft biometric data might be falsified in any way, the administrator will have no mechanism to deny or verify the provided information; since there is no proper biometric verification system in place to do this task.

125 Chapter 5. Gender prediction and classification from iris images

A biometric system such as the iris recognition system (IRS) for example, has not been developed to provide information about an individual’ gender. The development of a biometric system that can be able to automatically provide such demographic data, would be useful in situations where the individual has not been enrolled, yet their identity still needs to be determined [27].

In literature, the use of face biometrics to predict gender has gained much attention [105], simply because humans naturally recognise each other from the face. It is also technologically easy to capture a face from CCTV footage for applications in forensics and criminal investigations. Other biometric modalities that have ventured into predicting gender include the use of hand, gait, voice, text and the iris having the least amount of relevant and published studies so far. This is the affirmation of the challenge in this research field.

As a contribution to the existing body of knowledge, the objective of this chapter is to :

(i) continue using the extracted iris texture features to further explore and validate the discoveries made in the previous chapter

(ii) use several classification algorithms in order categorise all the participants within the database as either male or female.

The outcome of this chapter will address and provide answers to research questions 2 to 5, on the subject of gender. The foundation that has already been laid in Chapter 4, depicts that in as much as there are diverse ethnicities in the world; gender (used in the context of this thesis to categorise biological sex), generally falls between male and female. The proposed viewpoint is that, ”a biometric system that can achieve automatic ethnic classification as the first step, will tremendously narrow the search for gender to only that particular ethnic group”.

126 Chapter 5. Gender prediction and classification from iris images

Exploring the proposed analogy, section 5.1 of this chapter exploits the learned lessons and uses the same iris texture features to predict and classify individuals according to gender.

The arrangement of the rest of this chapter is as follows: Section 5.2 presents the available and relevant literature on the work that has been conducted by researchers with an aim to predict and /or classify gender from iris features; Section 5.3 details the experimental approach proposed in this thesis, the achieved results and analysis; Section 5.4 is the concluding summary.

5.2 Literature on gender prediction and classification

from iris images

In 2007, Thomas et al.[27] used machine learning techniques with an aim of developing a model that can learn to predict gender based on iris texture features and iris geometrical features. They argue that there has been sufficient research done on using biometric measures to verify an identity, however, there is still very little work done on using biometric measures to determine specific attributes such as gender from iris images. The main contribution of their work is to derive a meaningful feature vector that predicts gender from iris images.

Their proposed work, uses images that are self-acquired with the Iridian LG EOU200 system capturing NIR illumination images from 300 individuals. Poorly acquired images are discarded. For segmentation, Thomas et al. use the ND_IRIS software. The software uses the method of Wildes detailed in Chapter 2, to perform segmentation. They then use Daugman’s rubber sheet model to normalise the iris using 1-D Gabor filter, to represent the normalised image of size 20 x 480 as rows and columns respectively.

127 Chapter 5. Gender prediction and classification from iris images

Geometric measurements such as the radius and center of the pupil and iris, are extracted and stored among other extracted geometric features. From a total of 57, 137 iris images evenly divided between males and females, only 16,469 images were used. They employ a ten-fold cross validation using the C4.5 decision tree algorithm as the base classifier; the bagging method and random subspaces to increase the prediction accuracy.

The experimental results report a 75% accuracy gender prediction with bagging method. Their work also investigated the use of the same features for ethnic classification between Asians and Caucasians to an accuracy of 80% ethnic prediction. From the experimental results, Thomas et al. acknowledge and further state that ”there is still more research and work needed on determining more global or local features that are useful for gender classification and also for devising a more complicated feature vector”.

The work done by Lagree and Bowyer [106], uses iris textures to predict both gender and ethnicity from iris images. Their method analyses and uses the same features; experimental approach and algorithm discussed in Chapter 4 to address the two separate issues of gender and ethnic prediction.

For results on gender prediction, they report that using the originally proposed sequential minimum optimisation (SMO) approach achieves very low accuracy compared to the same algorithm’s performance on ethnicity. However, after making a few alterations of the proposed SMO classifier with 2, 5 and 10 fold cross validation, a 62% accuracy is achieved.

The work done by Bansal et al.[107] proposes a method to predict gender by extracting statistical features and wavelet texture features from iris images. Their proposed work uses images that have been self-acquired using the ISCAN-2 Dual iris scanner. The imaging device uses NIR

128 Chapter 5. Gender prediction and classification from iris images illumination to acquire a set of images per individual. The acquired images are of size 480 x 480 bmp. The collected database has 150 subjects consisting of 50 females and 100 males.

Their proposed method includes image acquisition, image processing, feature extraction and classification. For iris segmentation the circular Hough Transform is used, and iris normalisation with Daugman’s rubber sheet model is considered as the image processing stage. The normalised iris of size 100 x 500 is first enhanced, here the 100 rows are considered as radial resolution or the real part of the 2D normalised iris. The 150 columns which are imaginary parts, are considered as angular resolution.

Statistical features and texture features using 2D discrete wavelet transform (DWT) are extracted from the normalised iris. Pyramidal decomposition is used to create a feature vector. Both the statistical and texture features are then combined to generate a length of 2619 features per iris image.

The support vector machine (SVM) is used as a classifier with three different kernels and ten (10) fold cross validation. With the first polynomial kernel, a 81.27% overall correct classification rate (CCR) is achieved. With the second Gaussian kernel a mean ccuracy of 83.06% is achieved. The last radial basis kernel (RBF) an overall 80.97% CCR is achieved. In closing, their work acknowledges that by exploring more classifiers and better features, the system accuracy can be duly improved.

5.3 Proposed experimental approach on gender

classification

The work done here, follows the same algorithm presented and implemented in Chapter 4. The difference is that instead of finding the

129 Chapter 5. Gender prediction and classification from iris images average difference between ethnicities by equation 4.6, only the resulting feature vectors computed from equation 4.5 are used. The results of the model proposed in the previous chapter demonstrated the capability to classify an individual’s ethnicity, with the categorisation of gender being the unknown. Therefore, by exploiting the readily known ethnic information, it is justifiable that participants from the same ethnic group having the two genders are considered.

5.3.1 Results and analysis

This section presents the results obtained by using both the normalised local energy (LE) and mean amplitude (MA) texture features. Evaluations are made between black males and black females, as well as between white male and white females. As per the analogy, the aim is to investigate from a specific ethnic group, the different genders.

Starting with black males vs black females, Figure 5.1 presents the graphical representation of the distinction within this ethnic group using the LE and MA features.

FIGURE 5.1: Gender distinction between black males and black females.

From Figure 5.1, it is uniquely noticeable that the proposed model has not only managed to classify ethnicity from the previous chapter, the distinction of gender also proves feasible. The use of the proposed texture

130 Chapter 5. Gender prediction and classification from iris images features still show robustness even when used to determine the attribute of gender.

Figure 5.2 presents the results achieved for white males against white females.

FIGURE 5.2: Gender distinction between white males and white females.

From the achieved results, the distinction between the genders is again clearly witnessed with the use of the proposed model and choice of texture features. To continue with validating the accuracy of the proposed model, a pool of different Matlab classifiers are employed. Similar to the previous chapter, achieved results are tabulated in Table 5.1.

The results shown in Table 5.1, continue to demonstrate the effectiveness of using LE and MA as relevant and distinguishing texture features. The proposed experimental approach also proves to be robust for gender prediction for the tested participants. The combined correct gender classification accuracy with all normalised features using both cross validation and hold out validation is 94.35%. Figure 5.3 and Figure 5.4 are exemplary results showing the confusion matrix of the proposed model’s performance when all texture features are used to predict gender with a hold out validation of 20% and with no validation respectively .

131 Chapter 5. Gender prediction and classification from iris images

TABLE 5.1: Male vs female classification

Data Type Feature Type Method Accuracy Classifier

Raw Local Energy Cross Validation 81.20% Quadratic SVM

Raw Local Energy Hold Out 80.00% Cosine KNN

Raw Local Energy No Validation 91.00% Ensemble RUSBoosted Trees

Raw Mean Amplitute Cross Validation 82.40% Ensemble Subspace Discriminant

Logic Regression

Fine KNN

Raw Mean Amplitute Hold Out 83.30% Weighted KNN

Ensemble Bagged Trees

Ensemble subspace KNN

Raw Mean Amplitute No Validation 96.7.00% Ensemble Subspace KNN

Normalised Local Energy Cross Validation 86.50% Logistic Regression

Ensemble Subspace KNN

Quadratic Discriminant

Quadratic SVM Normalised Local Energy Hold Out 93.30% Fine Gaussian SVM

Fine KNN

Medium KNN

Ensemble Bagged Trees Normalised Local Energy No Validation 98.7 .00% Ensemble subspace KNN

Normalised Mean Amplitude Cross Validation 86.50% Cubic SVM

Normalised Mean Amplitude Hold Out 90.00% Fine KNN

Nomalised Mean Amplitude No Validation 97.40% Ensemble subspace KNN

Raw All Features Cross Validation 84.50% Fine KNN

Logistic Regression Raw All Features Hold Out 83.30% Ensemble Subspace Discriminant

Raw All Features No Validation 96.20% RUSBoosted Tree

Normalised All Features Cross Validation 93.70% Logistic Regression

Normalised All Features Hold Out 95.00% Logistic Regression

Complex Tree Normalised All Features No Validation 98.70% Medium Tree

132 Chapter 5. Gender prediction and classification from iris images

FIGURE 5.3: All males vs females confusion matrix.

FIGURE 5.4: All males vs females confusion matrix with no validation.

133 Chapter 5. Gender prediction and classification from iris images

5.4 Concluding summary

This chapter aimed to validate the discoveries that were made in Chapter 4 and also addressed and successfully provided answers to research questions 2-5 pertaining gender prediction and classification. Comparing the proposed approach to that of Thomas et al.’s [27] model based on similar factors such as: the concept of acquiring an independent database; the same lighting; image quality; size and acquisition conditions, the performance of our proposed model shows a 19.35% improvement.

Considering the work of Lagree and Bowyer [106], who also use their extracted features for both ethnic and gender prediction, the model proposed here shows a 32.35% improvement in classifying gender. The common concern which might cause the inconvenience mostly suffered by the work of other researchers is the use of pupil and iris statistical features. Although such statistical measurements might work well for segmentation purposes, the manner in which they vary in size does not relate to gender, but is a response to the amount of light entering the eye. The produced large-sized feature vectors achieved by other researchers may also introduce other computational complexities of reduced accuracy and performance speed of the overall biometric system.

The proposed model not only produces a compact feature vector even when combing both LE and MA features as a single feature vector, but it also offers the liberty of choice between using either of the two texture features as the main choice. The exploitation of ethnic information by this model also helps to readily and easily narrow the process of gender prediction, hence boosting the system’s processing time and speed. This would make it easy for integration to an existing IRS or for applications where enrollment is not a priority.

134 Chapter 6

Assessment of experimental results using Bayesian networks

The purpose of this chapter is to apply pattern recognition and machine learning techniques in order to assess the robustness of the proposed features and verify their ability to validate the achieved ethnic and gender classification results. In this chapter we employ Bayesian networks to model the probability of correct ethnic and gender belonging of an individual, given a set of data. The first section gives an overview of Bayesian networks; Section 6.2 presents the proposed problem formulation and architecture of our model. Section 6.3 presents the achieved results and analysis. Section 6.4 is the concluding summary

6.1 An overview on Bayesian networks

A Bayesian network (BNT) also known as Bayes net, Belief network and occasionally called Causal network, is a directed acyclic graph (DAG) or probabilistic graphical model [108–110]. The reasoning behind Bayes nets is that it assumes a probabilistic connection between states or attributes; such that given an input state, the objective is to find the conditional distribution of the target state [109]. The graphical model can be used to build various models with states and also the description of the

135 Chapter 6. Assessment of experimental results using Bayesian networks probabilistic relation among the states [111]. The probabilistic nature of Bayes nets is because not only are the models built on probabilistic distributions, but they also use the laws of probability to make predictions, detect anomalies, make decisions under uncertain conditions [108]. Although optional, Bayesian networks can also be graphically depicted. This makes it easy to visualise and understand the structure of the model that is being built. The graphical representation of Bayes nets is made up of nodes to depict varying model states and directed links between the states. The nodes and links between the nodes are referred to as the structural specification of the Bayesian network. More detailed reading on Bayesian networks can be found in [108–112]

6.2 Problem formulation

Based on the concept proposed and supported in Chapters 4 and 5; that is, it becomes less demanding to first classify ethnicity, so that any other search; such as gender, can be narrowed or directed to that particular ethnic group. In this thesis we propose to use Bayesian networks in order to model and create an ethnic classifier, which includes black and white individuals, as well as a gender classifier for male and female individuals. The motivation behind using Bayesian networks is that once a model has been compiled, fast responses are obtained. BNs take into consideration all the available data, regardless of incomplete data sets or sample sizes. This means that even a good prediction accuracy can be achieved with small sample sizes.

The question we are trying to answer can be posed as is it possible to achieve a model that uses existing iris texture features and; can be integrated into an existing IRS and be deployed in real time applications to classify an individual according to ethnicity and gender?.

136 Chapter 6. Assessment of experimental results using Bayesian networks

6.2.1 Model design

There is a total of 78 participants constituting of 22 black males, 22 white females, 17 black females and 17 white males. The LE and MA features jointly produce a 30 dimensional feature space per individual. Using the normalised features, we want to create a classifier that will classify the ethnicities and gender of all the participants within our database.

The nature of the problem is supervised learning, with both aspects of gender and ethnicity as discrete nodes. We find an answer to above question by implementing the approach proposed by Kevin Murphy [112]. We propose to address the problem by using a two-class, two component Gaussian mixture constructed within the BNT framework, with the directed acyclic graph (DAG) shown in Figure 6.1. Nodes 1 and 2 represent the discrete values of ethnicity and / or gender with node 3 representing the continuous values of the Gaussian mixture. The detailed and explained algorithm developed in Matlab is found in [109, 112].

FIGURE 6.1: Graphical structure of model design.

We use 80% of the features from both ethnicities and genders for training and 20% for testing the classifiers. We have four classifiers randomly

137 Chapter 6. Assessment of experimental results using Bayesian networks designed to have:

(i) the training data labeled as class 1: to represent the black males; and class 2 to represent black female participants only. The objective here is to achieve the likelihood of male and female classification from the same ethnic group.

(ii) the training data labeled as class 1: to represent the white male participants; and class 2: to represent the white female participants.

(iii) the training data labeled as class 1: to represent the black participants; and class 2: to represent the white participants; here, the objective is to achieve an ethnic classifier.

(iv) the training data labeled as class 1: to represent all male participants; and class 2: to represent the female participants within the database; here the objective is to achieve male and female classification within a mixed ethnic group.

6.3 Achieved experimental results and analysis

The Bayes net junction tree algorithm is used as an inference engine; and the expectation maximisation (EM) algorithm to train and fit the model. Once the model has been trained, generative modeling is used to understand what the model is doing, that is, to evaluate and visualise a representation of modeled data compared to the trained data. A comparable representation means that a new Bayes net has been efficiently trained.

The plots of the original data used to train our prosed models; and the synthetic data learned from the each model distribution are shown below. Figure 6.2 shows the original data and the synthetic data drawn from the model distribution for black males versus black females. The two plots show that some parts may be better represented by increasing the number of iterations used to train the Bayes net.

138 Chapter 6. Assessment of experimental results using Bayesian networks

FIGURE 6.2: Original training data and synthetic data from model distribution for black males vs black females.

Given the trained model, this is followed by using the testing data as evidence to evaluate the model for its classification performance and for computing the marginal of each class or node. The calculation of the marginal for each class produces the likelihood of the tested features belonging to the two respective classes being evaluated.

Figure 6.3 shows the performance of the modeled classifiers from the testing data containing black males and females respectively; with gender being the difference. The performance of both classifiers demonstrates eminent results with the values near one (1) depicted by the single solid line; showing appropriate regions of the data for both classifiers with 100% accurate classification probability elaborated by the confusion matrix in Figure 6.4; where the labels BM and BF represent black males and black females respectively. From the results, the concept of determining or predicting gender from the same ethnic group is further substantiated.

139 Chapter 6. Assessment of experimental results using Bayesian networks

FIGURE 6.3: Black male and black female classifier performance.

FIGURE 6.4: Confusion matrix of black males and black female classifiers.

140 Chapter 6. Assessment of experimental results using Bayesian networks

The results achieved from the white participants are also demonstrated. Figure 6.5 is a plot of the original data used to train the model; and the synthetic data learned from the model distribution. Figure 6.6 continues to show the performance of the modeled classifiers from the testing data containing white females and males respectively. The results from both classifiers continue to demonstrate success with the confusion matrix in Figure 6.7 showing 100% accurate gender classification probability; where labels WF and WM represent white females and white males respectively.

FIGURE 6.5: Original training data and synthetic data from model distribution for white females vs white males.

FIGURE 6.6: White female and white male classifier performance.

141 Chapter 6. Assessment of experimental results using Bayesian networks

FIGURE 6.7: Confusion matrix of white females and white males classifiers.

We use the same design model to classify between all the black and white ethnic groups within the database. Figure 6.8 shows the original data and the synthetic data drawn from the model distribution for black participants versus white participants. Figure 6.9 shows the performance of the classifier and its ability to perform ethnic classification between all the black participants and white participants. The achieved results elaborated by the confusion matrix in Figure 6.10 show 88% accurate ethnic classification probability for the blacks and 82% accurate classification probability for the white ethnic group.

142 Chapter 6. Assessment of experimental results using Bayesian networks

FIGURE 6.8: Original training data and synthetic data from model distribution for blacks vs whites.

FIGURE 6.9: Classifier performance outputs for black and white ethnic groups.

143 Chapter 6. Assessment of experimental results using Bayesian networks

FIGURE 6.10: Confusion matrix for black and white ethnic groups.

We again continue to evaluate the performance of gender classification from the database; here all the male participants and females participants represent class 1 and class 2 respectively. The results of the original training data with the synthetic data is as shown in Figure 6.11. The performance of both classifiers and their ability to classify males and females from a database with mixed ethnicities is shown in Figure 6.12. The performance of the gender classifiers is shown by the confusion matrix in Figure 6.13. Although the gender classification is tested within a mixed database having different ethnicities, the proposed approach still demonstrates promising results with 78% correct classification probability for females and 70% correct classification probability for males.

144 Chapter 6. Assessment of experimental results using Bayesian networks

FIGURE 6.11: Original training data and synthetic data from model distribution for all males vs all females.

FIGURE 6.12: Classifier performance outputs from all male and female participants.

145 Chapter 6. Assessment of experimental results using Bayesian networks

FIGURE 6.13: Confusion matrix for all males and females in the database.

6.4 Concluding summary

This chapter proposed to apply pattern recognition and machine learning techniques in order to assess the robustness of the proposed features and verify their ability to validate the ethnic and gender classification results achieved from chapters 4 and 5.

A Gaussian mixture constructed within the BNT framework was employed to evaluate the solution to the question of achieving automated ethnic and gender classification from iris features. Four different models

146 Chapter 6. Assessment of experimental results using Bayesian networks with an objective to answer the posed question were trained and tested. Although not as excellent as the results obtained from the separate use of LE and MA, results from the designed models show that the combined use of the extracted LE and MA features as a single texture feature vector for each individual, show credibility as potential features that can be used for automated ethnic and gender classification of individuals. However, it has also been observed that it is much more efficient to perform and achieve better gender prediction within a specified ethnic group than from a database with mixed ethnic groups. Furthermore because the extracted features are already available from any iris image, it becomes easier to integrate the proposed experimental approaches to an existing IRS.

147 Chapter 7

Conclusions on proposed methods

This chapter is the conclusion of this thesis. Section 7.1 summarises the findings and contributions that have been made in this thesis; Section 7.2 briefly lists some of the challenges that have been encountered towards realising the goals of this thesis and to conduct experiments. Section 7.3 presents the recommended future work of this study.

7.1 Summary of findings and contributions

The goal of this thesis was to investigate and evaluate the feasible use of iris features in order to achieve the classification of humans according to ethnicity and gender. The study conducted in this thesis is the first to investigate the calibre of participants (which include blacks and whites from the African continent) to perform experiments and hence the first to contribute to the existing literature in iris recognition in this regard.

In order to achieve the desired goal, eye images were first acquired using the Visat EY2 imaging device. Eye images from consenting but uncooperative black and white participants were captured at different times and locations with varying illumination. This factor poses a threat of acquiring images that are deemed as "non-ideal", because strict image quality control is a basic requirement for most iris identification systems. However, in a real life scenario, it is challenging to acquire images from

148 Chapter 7. Conclusions on proposed methods various participants in a single controlled room to produce high quality images that are well focused on the camera lens. Most state of the art algorithms used in a typical iris recognition system have been developed with the assumption that the pupil and iris are concentric circles that share the same center, when this not generally the case. The traditional segmentation methods in particular, have also been developed with the assumption that all images in the database are acquired under controlled environments, with uniform illumination and cooperative participants. Due to the assumptions, classical algorithms tend to have a high segmentation failure rate when images are non-ideal. It is unfortunately not possible to advance to any other subsequent stage of an iris recognition system, if accurate segmentation has not been achieved.

In order to address this challenge, the contribution made by this thesis was to propose a different approach to still achieve accurate iris segmentation even from non ideal images with very low contrast and very smooth pupil and iris boundaries.

The segmentation approach proposed in Chapter 3; is based on combining Bresenham’s algorithm with Chan-Vese algorithm to simultaneously localise the pupil-iris boundaries and to segment the iris in a single step. This is amongst one of the appealing characteristics of our proposed segmentation approach. Comparing to other methods in literature, another significant factor is its ability to segment noisy images, without degrading the iris textures and other features that are/ may be needed for further processing. Another advantageous finding is that the proposed approach offers flexibility to pre-define specified parameters as regularising that are needed to evolve the active contour of the Chan-Vese algorithm. This makes segmentation to be easier, faster, accurate and robust to any image quality and any object shape. Furthermore, the segmentation is based on the difference in image regions rather than on the difference in image intensities. This makes it feasible to segment any

149 Chapter 7. Conclusions on proposed methods eye image with differing qualities in real time. The disadvantage of the proposed method is that if the regularising terms are not properly defined, the evolving contour simply passes the desired boundary, hence failed a segmentation.

With accurate segmentation achieved, the subsequent stages of a typical IRS would be to perform normalisation, feature extraction and template matching. Traditionally, normalisation is performed to cater for image invariances such as off-angle, head tilt, etc. The normalisation stage uses the rubber sheet model to transform the segmented iris from a Cartesian plane to polar coordinates. The side effect posed by the traditional rubber sheet model is that it changes the geometrical structure and arrangement of the iris patterns; such that one cannot work backwards to place the discriminative features on the original image. The 2D Gabor wavelets used for extracting minute local features to generate an iris code of 1024 bits, do not discriminate the various feature types found within the iris. The goal of any recognition system is to produce a perfect match between a reference image and its query image. Traditionally the Hamming distance, which is a dissimilarity measure calculation, is used to quantify this match. Therefore, achieving a perfect match between a reference template and its query for non-ideal images image becomes challenging. This is due to the external uncontrolled factors under which images are captured in a real life scenario.

The contribution made by this thesis to address such challenges is proposing a method that uses corner and edge detection; and measures the significance of features in any image; known as Phase Congruency. Phase congruency is invariant to image illumination and contrast. This characteristic makes it a robust and reliable feature detection method. The method proposed in this thesis to manage non ideal images; fuses phase congruency with the Harris corner detector. This approach is able to identify and detect significant corner features found from the arrangement

150 Chapter 7. Conclusions on proposed methods of the iris patterns, as features for extraction. This combination robustly localises feature points within the iris image that are not only congruent in phase but are also invariant to rotation. This is a novel and different approach in iris recognition that can be employed as an alternative to using the rubber sheet model. Furthermore with the detected feature points, a compact feature vector of 512 bits containing the exact location of features is easily generated and stored as a template to be used for matching purposes.

Current research in iris biometrics has advanced to either nullifying or validating the long standing argument that iris features have no genetic relation. This particular research topic has gained momentum over the past few years, with many researchers focusing on using iris texture features to classify individuals according to either ethnicity or gender. While researchers in literature remain adamant that people belonging to the same ethnic group have similar iris textures, they also report that it is still an insufficiently investigated subject. It is further reported that the task of determining such soft biometrics attributes from the iris is still a challenging task. A major issue of concern in this area is that, all the research conducted to investigate this topic has been focused on only two ethnic groups around the world, that is, from Asian and European participants. This is affirmation that more research to cover other ethnic groups and to fully support or contradict the standing argument is still a desperate need.

Being the first of its kind, the experimental work proposed in Chapter 4 and Chapter 5 of this thesis sought to contribute to literature by investigating a different population that has never been explored before. This has been done with the acknowledgment and realisation of the major differences that are evident between Asian, European and African people. Chapter 4 proposed the use of two designs of Gabor filter banks to investigate and extract iris texture features. The difference between the

151 Chapter 7. Conclusions on proposed methods features used for recognition and features used for soft biometric detection is that: recognition uses only the local and minute features of the iris; whereas ethnic and gender detection uses the global textures within the iris.

The novelty of the proposed design is its ability to produce two sets of intricate texture features which include local energy( LE) and mean amplitude (MA); that can either be used jointly or separately as feature vectors to determine the ethnic distinction of individuals. The proposed method is implemented such that it first provides an answer to the question: can iris features really demonstrate a distinction and/ or similarity between individuals of similar and /or different ethnic groups?.

The answer to this question was appropriately provided in Chapter 4 through the use of both statistical and classification methods. Experimental findings also showed that it is at the lower wavelengths of the Gabor filters that much variation is seen as pertaining to ethnic distinction. In addition to this, the proposed design of our Gabor filter banks concurrently exhibits the distinction between the genders of individuals. The approach in this thesis is the first to obtain the highest and correct classification rate of both ethnicity and gender classification from using the same features.

Another discovery that was made through the proposed approach is that it is easier to first determine the ethnic belonging of an individual, so that the aspect of gender can be simply narrowed to a specific ethnic group. This newly proposed concept was also shown to hold true by the results achieved in Chapters 4, 5 and 6. In Chapter 5, the topic of gender was investigated using the design proposed in Chapter 4. Results confirmed that a higher correct gender classification rate is achieved when investigated from within the same ethnic group than from a database with mixed ethnicities.

152 Chapter 7. Conclusions on proposed methods

Chapter 6 aimed to validate the eminent results achieved in Chapters 4 and 5; and also to model various classifiers that can predict the probability of an individual’s ethnic and gender belonging. In order to do this an algorithm that uses a Gaussian mixture model constructed within the BNT framework was implemented. According to literature, this thesis is the first of its kind in this particular topic of research to undertake this approach of further exploring and validating the results obtained in Chapters 4 and 5. The results achieved in Chapter 6 continue to confirm the validity of the proposed concept and to demonstrate an effortless and feasible integration to an existing IRS.

7.2 Encountered challenges during study

The different experimental methods proposed in this thesis have demonstrated promising results that are of great need and relevance to existing literature in iris biometrics and other biometric recognition systems. However, in order to realise the goal of this thesis, a different population was needed, as such, the work conducted in this thesis came with the following challenges:

(i) Ethical clearance: since the study is conducted using human participants, ethical clearance from the research council is needed before any data acquisition can commence. This is unfortunately a lengthy and possibly tedious process due to the various stakeholders and decision makers involved.

(ii) Recruiting participants: in order to have a database to conduct our investigations, eye images had to be acquired from consenting adults. This process needs a preparation of relevant documents such as the scope of the study as well as consent forms. Since the consenting participants are working adults, letters to ask permissions from their managers were also needed.

153 Chapter 7. Conclusions on proposed methods

(iii) Eye image acquisition: consenting and willing adults could not be gathered in one room at the same time, due to this, the data acquisition process was arranged to suit the conditions and availability of willing participants, over a certain period of time to acquire enough samples.

(iv) Time constraint issues: because the participants are working adults, appointments to capture eye images had to be arranged. The challenges here arise when a participant is no longer available to have their images captured; and while available, there is a challenge to get full cooperation from the participant, because of other pressing engagements. As a result only two sets of images were acquired from a limited number of individuals, while this is the case, one finds that not all images are of usable standards; and it become even more difficult to get the same participants to sit for re-capturing.

7.3 Recommendations for future work

This goal of this thesis was to address the challenges and problems that have been listed in all the respective chapters, with an aim to propose robust, non complex and cost effective solutions that can enable implementation using any image quality and still be deployable in real time applications. As a follow up to the work done in this thesis and research niche area, the following is recommended as future work.

• Collection of other public iris databases: due to the unavailability of racially diverse iris databases, the collection and development of other iris databases that also accommodate a much younger age group is recommended. The availability of such databases will cater for other ethnic groups around the globe. This can help in:

(i) designing and developing more agile algorithms,

(ii) avoiding biased results, and also

154 Chapter 7. Conclusions on proposed methods

(iii) efficiently address the global problem of identity theft and child trafficking.

• Unlike the well researched and well developed algorithms used for individual identification and verification, the available algorithms proposed in literature to determine ethnic distinction and achieve classification, are focused on two ethnic groups. This is an indication of insufficient research studies in this field; especially for deployment and integration to an existing IRS. In order to cast the net wide, we also aim to make the database publicly available for other researchers in the field of biometrics.

155 References

[1] A. Jain , P. Flynn & A. A. Ross. Handbook of biometrics. Springer Science & Business Media, 2007.

[2] IEEE Society. Biometrics Fundamentals. USA/Canada: IEEE Society, 2010.

[3] NSTC Subcommittee on Biometrics. Biometrics History. 2006. URL: http://www.biometrics.gov/documents/biohistory.

pdf (visited on 11/02/2012).

[4] K. W. Bowyer , K. Hollingsworth & P. J. Flynn. “Image understanding for iris biometrics: A survey”. In: Computer vision and image understanding 110.2 (2008), pp. 281–307.

[5] S. Z. Li & A. Jain. Encyclopedia of biometrics. Springer Publishing Company, Incorporated, 2015.

[6] A. Poursaberi & B. N. Araabi. “Iris recognition for partially occluded images: methodology and sensitivity analysis”. In: EURASIP Journal on Advances in Signal Processing 2007.1 (2006), pp. 1–12.

[7] J. G. Daugman. “How iris recognition works”. In: Circuits and Systems for Video Technology, IEEE Transactions on 14.1 (2004), pp. 21–30.

[8] R. P. Wildes. “Iris recognition: an emerging biometric technology”. In: Proceedings of the IEEE 85.9 (1997), pp. 1348–1363.

[9] A. Uhl & C. Rathgeb. “The State-of-the-Art in Iris Biometric Cryptosystems”. In: State of the art in Biometrics. InTech, 2011.

[10] A. Muron & J. Pospisil. “The human iris structure and its usage”. In: Physica 39 (2000), pp. 87–95.

156 References

[11] J. G. Daugman. “High confidence visual recognition of persons by a test of statistical independence”. In: Pattern Analysis and Machine Intelligence, IEEE Transactions on 15.11 (1993), pp. 1148–1161.

[12] J. G. Daugman. “The importance of being random: statistical principles of iris recognition”. In: Pattern recognition 36.2 (2003), pp. 279–291.

[13] Disabled World. Basic structure of the human eye. 2007. URL: http:// www.disabled-world.com/artman/publish/eye-color.

shtml (visited on 02/02/2016).

[14] H. Proença & L. Alexandre. “Toward noncooperative iris recognition: A classification approach using multiple signatures”. In: IEEE Transactions on Pattern Analysis and Machine Intelligence 29.4 (2007).

[15] P. Khaw. “Iris recognition technology for improved authentication”. In: SANS Institute (2002), pp. 5–8.

[16] A. Panganiban, N. Linsangan and F. Caluyo. “Wavelet-based feature extraction algorithm for an iris recognition system”. In: Journal of information processing systems 7.3 (2011), pp. 425–434.

[17] A. E. Yahya & M. D. Nordin. “Accurate iris segmentation method for non-cooperative iris recognition system”. In: Journal of Computer Science 6.5 (2010), p. 492.

[18] A. Hilal , P. Beauseroy & B. Daya. “Real shape inner iris boundary segmentation using active contour without edges”. In: Audio, Language and Image Processing (ICALIP), 2012 International Conference

on. IEEE. 2012, pp. 14–19.

[19] S. M. S. Moosavi , S. M. Seyedzade & A. Ayatollahi. “A novel iris recognition system based on active contour”. In: Biomedical Engineering (ICBME), 2010 17th Iranian Conference of. IEEE. 2010, pp. 1–4.

157 References

[20] I. K. Khan & M. J. Yogesh. An Analysis on Iris Segmentation Method for Non Ideal Iris Images like off-axis angle and distance acquired.

[21] X. Qiu , Z. Sun & T. Tan. “Global texture analysis of iris images for ethnic classification”. In: International Conference on Biometrics. Springer. 2006, pp. 411–418.

[22] X. Qiu , Z. Sun & T. Tan. “Coarse iris classification by learned visual dictionary”. In: Advances in Biometrics (2007), pp. 770–779.

[23] L. Stark , K. W. Bowyer & S. Siena. “Human perceptual categorization of iris texture patterns”. In: Biometrics: Theory Applications and Systems (BTAS), 2010 Fourth IEEE International

Conference on. IEEE. 2010, pp. 1–7.

[24] S. Lagree & K.W. Bowyer. “Ethnicity Prediction Based on Iris Texture Features.” In: MAICS. 2011, pp. 225–230.

[25] H. Zhang , Z. Sun , T. Tan & J. Wang. “Ethnic classification based on iris images”. In: Chinese Conference on Biometric Recognition. Springer. 2011, pp. 82–90.

[26] A. Zarei & D. Mou. “Artificial neural network for prediction of ethnicity based on iris texture”. In: Machine Learning and Applications (ICMLA), 2012 11th International Conference on. Vol. 1. IEEE. 2012, pp. 514–519.

[27] V. Thomas , N. V. Chawla , K. W. Bowyer & P. J. Flynn. “Learning to predict gender from iris images”. In: Biometrics: Theory, Applications, and Systems, 2007. BTAS 2007. First IEEE International Conference on. IEEE. 2007, pp. 1–5.

[28] L. Flom & A. Safir. Iris recognition system. US Patent 4,641,349. 1987.

[29] J. Daugman. “High confidence personal identification by rapid video analysis of iris texture”. In: Security Technology, 1992. Crime Countermeasures, Proceedings. Institute of Electrical and Electronics

Engineers 1992 International Carnahan Conference on. IEEE. 1992, pp. 50–60.

158 References

[30] R. P. Wildes , J. C. Asmuth , G. L. Green , S. C. Hsu , R. J. Kolczynsk , J. R. Matey & S. E. McBride. “A machine-vision system for iris recognition”. In: Machine vision and Applications 9.1 (1996), pp. 1–8.

[31] R. P. Wildes , J. C. Asmuth , K. J. Hanna , G. L. Green , S. C. Hsu , R. J. Kolczynsk , J. R. Matey & S. E. McBride. Automated, non-invasive iris recognition system and method. US Patent 5,572,596. 1996.

[32] E. Tabassi. Iris Quality Standerdization. 2014. URL: https : / / www . nist . gov / sites / default / files / documents / 2016 / 12 / 05 / xx _ thursday _ tabassi _ iris _ q _ std . pdf

(visited on 02/02/2017).

[33] Q. Tian , H. Qu , L. Zhang & R. Zong. Personal Identity Recognition Approach Based on Iris Pattern. INTECH Open Access Publisher, 2011.

[34] H. Proença & L. Alexandre. “UBIRIS: A noisy iris image database”. In: Image Analysis and Processing–ICIAP 2005 (2005), pp. 970–977.

[35] R. Crandall. “Image segmentation using the Chan-Vese algorithm”. In: Project report from ECE 532 (2009), pp. 1–23.

[36] L. Masek. “Recognition of human iris patterns for biometric identification”. In: BSc Thesis, School of Computer Science and Software Engineering, University of Western Australia (2003).

[37] R. D. Labati , A. Genovese , V. Piuri & F. Scotti. “Iris segmentation: state of the art and innovative methods”. In: Cross Disciplinary Biometric Systems. Springer, 2012, pp. 151–182.

[38] P. Karn , X. H. He , S. Yang & X. H. Wu. “Iris recognition based on robust principal component analysis”. In: Journal of Electronic Imaging 23.6 (2014), pp. 063002–063002.

[39] G. Mabuza-Hocquet & F. Nelwamondo. “Fusion of Phase Congruency and Harris Algorithm for Extraction of Iris Corner Points”. In: Artificial Intelligence, Modelling and Simulation (AIMS), 2015 3rd International Conference on. IEEE. 2015, pp. 315–320.

159 References

[40] D. A. Clausi & M. E. Jernigan. “Designing Gabor filters for optimal texture separability”. In: Pattern Recognition 33.11 (2000), pp. 1835–1849.

[41] P. Flanagan. Iris Challenge Evaluation. 2005. URL: https : / / www . nist . gov / programs - projects / iris -

challenge-evaluation-ice (visited on 01/02/2016).

[42] H. Proença. Noisy Iris Challenge Evaluation. 2009. URL: http://nice1.di.ubi.pt/ (visited on 01/05/2016).

[43] M. De Marsico. Mobile Iris Challenge Evaluation. 2013. URL: http://biplab.unisa.it/MICHE/index.html (visited on 10/05/2016).

[44] L. Ma ,T. Tan , Y. Wang & D. Zhang. “Efficient iris recognition by characterizing key local variations”. In: IEEE Transactions on image processing 13.6 (2004), pp. 739–750.

[45] Z. He, T. Tan , Z. Sun & X. Qiu. “Toward accurate and fast iris segmentation for iris biometrics”. In: IEEE transactions on pattern analysis and machine intelligence 31.9 (2009), pp. 1670–1684.

[46] N. B. Puhan , N. Sudha & A. S. Kaushalram. “Efficient segmentation technique for noisy frontal view iris images using Fourier spectral density”. In: Signal, Image and Video Processing 5.1 (2011), pp. 105–119.

[47] M. A. Luengo-Oroz , E. Faure & J. Angulos. “Robust iris segmentation on uncalibrated noisy images using mathematical morphology”. In: Image and Vision Computing 28.2 (2010), pp. 278–284.

[48] C. Li , C. Xu , C. Gui & M. D. Fox. “Level set evolution without re-initialization: a new variational formulation”. In: Computer Vision and Pattern Recognition, 2005. CVPR 2005. IEEE Computer Society

Conference on. Vol. 1. IEEE. 2005, pp. 430–436.

160 References

[49] J. Daugman. “New methods in iris recognition”. In: IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics)

37.5 (2007), pp. 1167–1175.

[50] X. Ren , Z. Peng , Q. Zeng , C. Peng , J. Zhang , S. Wu & Y. Zeng. “An improved method for Daugman’s iris localization algorithm”. In: Computers in Biology and Medicine 38.1 (2008), pp. 111–115.

[51] M. Shamsi , P. B. Saad , S. B. Ibrahim & and A. R. Kenari. “Fast algorithm for iris localization using Daugman circular integro differential operator”. In: Soft Computing and Pattern Recognition, 2009. SOCPAR’09. International Conference of. IEEE. 2009, pp. 393–398.

[52] A. Radman , K. Jumari & N. Zainal. “Fast and reliable iris segmentation algorithm”. In: IET Image Processing 7.1 (2013), pp. 42–49.

[53] D. L. Baggio. “GPGPU based image segmentation livewire algorithm implementation”. In: Sao Jose dos Campos (2007).

[54] W. Sankowski , K. Grabowski , M. Napieralska , M. Zubert & A. Napieralski. “Reliable algorithm for iris segmentation in eye image”. In: Image and vision computing 28.2 (2010), pp. 231–237.

[55] M. Djoumessi. “Iris segmentation using Daugman’s integro-differential operator”. In: NSF REU at Utah State University (2010).

[56] A. K. Nsaef , A. Jaafar & and K.N. Jassim. “Enhancement segmentation technique for iris recognition system based on Daugman’s Integro-differential operator”. In: Instrumentation & Measurement, Sensor Network and Automation (IMSNA), 2012

International Symposium on. Vol. 1. IEEE. 2012, pp. 71–75.

[57] A. Z. Zainal , M. Manaf , A.S. Shibghatullah , S. M. Yunos , S. Anawar & Z. Ayop. “Iris Segmentation Analysis using

161 References

Integro-Differential Operator and Hough Transform in Biometric System”. In: JTEC (2012), pp. 1–8.

[58] Q. Wang , X. Zhang , M. Li , X. Dong , Q. Zhou & Y. Yin. “Adaboost and multi-orientation 2D Gabor-based noisy iris recognition”. In: Pattern Recognition Letters 33.8 (2012), pp. 978–983.

[59] I. U. Nkole , G. B. Sulong & S. Saparudin. “An enhanced iris segmentation algorithm using circle Hough transform”. In: (2012).

[60] S. Sanap & U. M. Chaskar. “Improved Iris segmentation using circular hough transform”. In: Proceedings of International Conference on Signal and Image Processing (ICISP). Vol. 4. 2013, pp. 153–161.

[61] S. Chawla & A. Oberoi. “A robust algorithm for iris segmentation and normalization using hough transform”. In: Global Journal of Business Management and Information Technology 1.2 (2011), pp. 69–76.

[62] A. Bendale , A. Nigam , S. Prakash & P. Gupta. “Iris segmentation using improved hough transform”. In: International Conference on Intelligent Computing. Springer. 2012, pp. 408–415.

[63] R. Sahak & A. Saparon. “Iris localization using colour segmentation and circular Hough transform”. In: Biomedical Engineering and Sciences (IECBES), 2012 IEEE EMBS Conference on. IEEE. 2012, pp. 784–788.

[64] S. Singh & Shikha Singh. “Iris Segmentation Along with Noise Detection using Hough Transform”. In: International Journal of Engineering and Technical Research (IJETR) (2013), pp. 2321–0869.

[65] C. Houston. Iris Segmentation and Recognition Using Circular Hough Transform and Wavelet Features.

[66] S. A. Sahmoud & I. S. Abuhaiba. “Efficient iris segmentation method in unconstrained environments”. In: Pattern Recognition 46.12 (2013), pp. 3174–3185.

162 References

[67] A. Uhl & P. Wild. “Weighted adaptive hough and ellipsopolar transforms for real-time iris segmentation”. In: Biometrics (ICB), 2012 5th IAPR International Conference on. IEEE. 2012, pp. 283–290.

[68] K. Roy , P. Bhattacharya & C. Y. Suen. “Iris segmentation using variational level set method”. In: Optics and Lasers in Engineering 49.4 (2011), pp. 578–588.

[69] R. C. Gonzalez & R. E Woods. “Image processing”. In: Digital image processing 2 (2007).

[70] A. Fitzgibbon,M. Pilu & R. B. Fisher. “Direct least square fitting of ellipses”. In: IEEE Transactions on pattern analysis and machine intelligence 21.5 (1999), pp. 476–480.

[71] M. Kass , A. Witkin & D. Terzopoulos. “Snakes: Active contour models”. In: International journal of computer vision 1.4 (1988), pp. 321–331.

[72] S. Shah & A. Ross. “Iris segmentation using geodesic active contours”. In: IEEE Transactions on Information Forensics and Security 4.4 (2009), pp. 824–836.

[73] M. A. Abdullah , S. S. Dlay & W. L. Woo. “Fast and accurate pupil isolation based on morphology and active contour”. In: International Journal of Information and Electronics Engineering 4.6 (2014), p. 418.

[74] M. A. Abdullah , S. S. Dlay , W. L. Woo & J. A. Chambers. “Robust iris segmentation method based on a new active contour force with a noncircular normalization”. In: IEEE Transactions on Systems, Man, and Cybernetics: Systems (2016).

[75] N. Otsu. “A threshold selection method from gray-level histograms”. In: IEEE transactions on systems, man, and cybernetics 9.1 (1979), pp. 62–66.

[76] M. Vatsa, R. Singh and A. Noore. “Improving iris recognition performance using segmentation, quality enhancement, match

163 References

score fusion, and indexing”. In: Systems, Man, and Cybernetics, Part B: Cybernetics, IEEE Transactions on 38.4 (2008), pp. 1021–1035.

[77] K. Roy & P. Bhattacharya. “Nonideal iris recognition using level set approach and coalitional game theory”. In: International Conference on Computer Vision Systems. Springer. 2009, pp. 394–402.

[78] P. Kalavathi & J. B. Narayani. “IRIS Segmentation Using Geodesic Active Contour Method”. In: Middle-East Journal of Scientific Research 24.S2 (2016), pp. 330–334.

[79] T. F. Chan & L. A. Vese. “An active contour model without edges”. In: International Conference on Scale-Space Theories in Computer Vision. Springer. 1999, pp. 141–151.

[80] T. F. Chan & L. A. Vese. “Active contours without edges”. In: IEEE Transactions on image processing 10.2 (2001), pp. 266–277.

[81] P. Getreuerl. “Chan-vese segmentation”. In: Image Processing On Line 2 (2012), pp. 214–224.

[82] D. Mumford & J. Shah. “Optimal approximations by piecewise smooth functions and associated variational problems”. In: Communications on pure and applied mathematics 42.5 (1989), pp. 577–685.

[83] A. E. Yahya & M. D. Nordin. “Improving iris segmentation by specular reflections removable”. In: Information Technology (ITSim), 2010 International Symposium in. Vol. 1. IEEE. 2010, pp. 1–3.

[84] G. Yanto , M. H. Jaward & N. Kamrani. “Bayesian Chan-Vese segmentation for iris segmentation”. In: Visual Communications and Image Processing (VCIP), 2013. IEEE. 2013, pp. 1–6.

[85] J. Bresenham. “A linear algorithm for incremental digital display of circular arcs”. In: Communications of the ACM 20.2 (1977), pp. 100–106.

164 References

[86] Cardlogixcorporation. VistaEY2 Dual iris Scanner and Face camera.

URL: http : / / www . cardlogix . com / products / biometrics / vista - ey2- biometric- dual- iris- scanner- face- camera.asp

(visited on 01/02/2013).

[87] P. Kovesi. “Phase congruency detects corners and edges”. In: The australian pattern recognition society conference: DICTA 2003. 2003.

[88] M. C. Morrone & D. C. Burr. “Feature detection in human vision: A phase-dependent energy model”. In: Proceedings of the Royal Society of London B: Biological Sciences 235.1280 (1988), pp. 221–245.

[89] M. C. Morrone , J. Ross , D. C. Burr & and R. Owens. “Mach bands are phase dependent”. In: Nature 324.6094 (1986), pp. 250–253.

[90] P. Du , X. Shi , N. Wang & R. Deng. “Iris recognition based on principal phase congruency”. In: Control and Decision Conference (CCDC), 2012 24th Chinese. IEEE. 2012, pp. 1159–1162.

[91] X. Yuan & P. Shi. “Iris feature extraction using 2D phase congruency”. In: Information Technology and Applications, 2005. ICITA 2005. Third International Conference on. Vol. 2. IEEE. 2005, pp. 437–441.

[92] Z. Osman. “Iris recognition using phase congruency”. In: Computer Modelling and Simulation (UKSim), 2011 UkSim 13th International

Conference on. IEEE. 2011, pp. 341–344.

[93] P. S. Patil , S. R. Kolhe , R. V. Patil & P. M. Patil. “Performance Evaluation in Iris Recognition and CBIR System based on phase congruency”. In: International Journal of Computer Applications 47.14 (2012).

[94] Z. Zhangl , R. Deriche , O. Faugeras & Q. T. Luong. “A robust technique for matching two uncalibrated images through the recovery of the unknown epipolar geometry”. In: Artificial intelligence 78.1-2 (1995), pp. 87–119.

165 References

[95] Hao Zhang. “SOFT BIOMETRICS ARE HARD”. PhD thesis. Colorado State University, 2013.

[96] D. Thakkar. Unimodal Biometrics vs. Mutltimodal Biometrics. 2016.

URL: https://www.bayometric.com/unimodal-vs-multimodal

(visited on 10/18/2016).

[97] K. Ricanek Jr & B. Barbour. “What are soft biometrics and how can they be used?” In: Computer 44.9 (2011), pp. 106–108.

[98] S. C. Zhu , C. E. Guo, Y. Wang & Z. Xu. “What are textons?” In: International Journal of Computer Vision 62.1 (2005), pp. 121–143.

[99] A. Hellem. Brown eyes. 2016. URL: http : / / www . allaboutvision . com / conditions / eye -

color-brown.htm (visited on 10/18/2016).

[100] D. Zhou. “Texture analysis and synthesis using a generic Markov- Gibbs image model”. PhD thesis. ResearchSpace@ Auckland, 2006.

[101] Z. H. Wang & Z. C. Mu. “Gender classification using selected independent-features based on genetic algorithm”. In: Machine Learning and Cybernetics, 2009 International Conference on. Vol. 1. IEEE. 2009, pp. 394–398.

[102] M. Abdel-Mottaleb M. Haghighat , S. Zonouz &. “Identification using encrypted biometrics”. In: International Conference on Computer Analysis of Images and Patterns. Springer. 2013, pp. 440–448.

[103] A.K Jain & F. Farrokhnia. “Unsupervised texture segmentation using Gabor filters”. In: Pattern recognition 24.12 (1991), pp. 1167–1186.

[104] D. Zheng , Y. Zhao & J. Wang. “Features extraction using a Gabor filter family”. In: Proceedings of the sixth Lasted International conference, Signal and Image processing, Hawaii. 2004.

166 References

[105] M. Da Costa-Abreu , M. Fairhurst & M. Erbilek. “Exploring gender prediction from iris biometrics”. In: Biometrics Special Interest Group (BIOSIG), 2015 International Conference of the. IEEE. 2015, pp. 1–11.

[106] S. Lagree & K. W. Bowyer. “Predicting ethnicity and gender from iris texture”. In: Technologies for Homeland Security (HST), 2011 IEEE International Conference on. IEEE. 2011, pp. 440–445.

[107] A. Bansal , R. Agarwal & R. K. Sharma. “SVM based gender classification using iris images”. In: Computational Intelligence and Communication Networks (CICN), 2012 Fourth International Conference

on. IEEE. 2012, pp. 425–429.

[108] Bayes server. Bayesian networks an introduction. 2016. URL: https: //www.bayesserver.com/docs/introduction/bayesian-

networks (visited on 10/18/2017).

[109] Lior Rokach. Pattern classification using ensemble methods. Vol. 75. World Scientific, 2010.

[110] A. Masood. Bayesian networks a brief introduction. 2013. URL: https://www.slideshare.net/adnanmasood/bayesian-

networks-primer-21281913 (visited on 10/18/2017).

[111] Norsys software corp. Introduction to Bayes nets. 1996. URL: https: //www.norsys.com/tutorials/netica/secA/tut_A1.htm

(visited on 10/18/2017).

[112] K. Murphy. “The Bayes net toolbox for matlab”. In: Computing science and statistics 33.2 (2001), pp. 1024–1034.

167