Enhanced face recognition methods

Faculté des Sciences, 4

Tel: +212 (0) 5 37 77 18 34/35/38, Fax: +212 (0) 5 37 77 42 61, http://www.fsr.ac.ma.

UNIVERSIT

Hythem AHMED ABDALLA AHMED

Discipline:

Enhanced

Faculté des Sciences, 4 Avenue Ibn Battouta B.P. 1014 RP, Rabat

Tel: +212 (0) 5 37 77 18 34/35/38, Fax: +212 (0) 5 37 77 42 61, http://www.fsr.ac.ma.

UNIVERSITÉ MOHAMMED V- AGDAL

FACULTÉ DES SCIENES

RABAT

THÈSE DE DOCTORAT

Présentée par:

Hythem AHMED ABDALLA AHMED

Discipline: Sciences de l’Ingénieur

Spécialité: Informatique

Enhanced face recognition methods

Soutenue le 05-06-2013

Devant le jury composé de:

Professeur, Faculté des Sciences,

Professeur, Faculté des Sciences,

Professeur , University of Wisconsin, La Crosse, USA Professeur , Faculté des Sciences,

Professeur , Faculté des Sciences

Professeur, Chargé des statistiques, Office des Nations Unies pour l’Afrique du Nord

Avenue Ibn Battouta B.P. 1014 RP, Rabat- Maroc Tel: +212 (0) 5 37 77 18 34/35/38, Fax: +212 (0) 5 37 77 42 61, http://www.fsr.ac.ma.

AGDAL

No d’ordre: 2644

face recognition methods

, Faculté des Sciences, Rabat

, Faculté des Sciences, Rabat

Professeur , University of Wisconsin, La Crosse, USA , Faculté des Sciences, Rabat

Faculté des Sciences, Rabat Professeur, Chargé des statistiques, Office des

Acknowledgements

This dissertation has been prepared in The Laboratory of Conception and Systems (LCS) of the Faculty of Sciences - Rabat, under the direction of Dr. M. Jedra and Dr. N. Zahid.

I would like to acknowledge the inspiration instruction and guidance to Dr. M. JEDRA, Professor in the Faculty of sciences - Rabat, and responsible of U.F.R: Architecture of Computer Systems and the initial impetus to do my research has given me a deep appreciation.

My deepest gratitude to Dr. N. ZAHID, Professor in the Faculty of sciences - Rabat, for his excellent advices, suggestions and his patience until this research has been real. I would like to thank him for his support.

I would like to sincerely thank Dr. Aziz ETTOUHAMI; Professor and Director of the Conception and Systems Laboratory in the Faculty of sciences – Rabat and the president of the jury of this dissertation, for his participation in the jury of this thesis.

A lot of thanks also go to Dr. M. BENKHALIFA, Professor in the Faculty of Sciences - Rabat, for his participation in the jury of this thesis.

I would like to express my deepest gratitude to Dr. M.A. ELHINDI, Professor, Assistant Vice Chancellor and Chief Information Officer of University of Wisconsin-La Crosse, USA to assume the traveling trouble to take a chair in the jury of this thesis.

My great attitudes for Dr. KHOGALI Ali KHOGALI, Professor Statistician in the Office for North Africa of the United Nations Economic Commission for Africa - Rabat, to accept have a chair in the jury of this thesis.

I have to acknowledge the Moroccan Agency for International Cooperation (AMCI) to have allowed me to have a grant. I am truly appreciate and really indebted to them.

I would like to acknowledge and thank The University of kassala and The College of Sciences for allowing me to conduct my research and providing all the assistance requested. Special thanks go to the staff of the college for their continued supports.

Table of Content

Abstract

General introduction...1

Chapter 1: Background and related work... 4

1.1. Introduction... 4

1.2. Biometrics overviews... 5

1.2.1. Biometrics efficiency measurements... 6

1.2.2. Biometrics technology characteristics... 8

1.2.3. The applications of different biometrics... 9

1.3. Survey on face recognition methods... 14

1.3.1. Human recognition of face...15

1.3.2. Machine recognition of faces... 17

1.4. Statistical approaches for face recognition... 18

1.4.1. Face detection and recognition by PCA... 19

1.4.2. Face recognition by LDA... 21

1.4.3. Feature-based approaches... 22

1.5. Media learning application... 23

1.5.1 Face recognition by neural network... 24

1.5.2. Face recognition by support vector machine SVM... 26

1.6. Other approaches... 26

1.6.1. Transformation based systems... 26

1.6.2. Range data... 28

1.7. The face recognition problems... 29

1.8. Conclusion... 30

Chapter 2: Statistical face recognition methods... 31

2.1. Introduction... 31

2.2. Pattern classification basics... 31

2.3. Nearest mean classifier... 33

2.4.1. Principal component analysis... 34

2.4.2. Linear discriminant analysis... 35

2.5. Face recognition methods... 37

2.5.1. Eigenface method... 37

2.5.2. Linear discriminant analysis... 42

2.5.3. Subspace LDA method... 46

2.5.4. Kernel principal component analysis... 49

2.5.5. Kernel discriminant analysis methods... 51

2.5.6. Fuzzy principal components analysis... 53

2.5.7. Fuzzy linear discriminant analysis... 55

2.6. Conclusion... 56

Chapter 3: Extension of two dimensional PCA and LDA methods in face recognition... 57

3.1. Introduction... 57

3.2. Two dimensional principal component analysis... 57

3.3. Two dimensional linear discriminant analysis... 63

3.4. The 2DPCA+2DLDA method... 68

3.5. Relevance weighted two dimensional linear discriminant analysis... 72

3.6. Weighted scatter-difference-based two dimensional discriminant analysis... 76

3.7. Conclusion... 82

Chapter 4: Kernel two dimensions discriminant analysis methods... 84

4.1. Introduction... 84

4.2. Over-view kernel discriminant analysis methods... 85

4.3. Kernel relevance weighted 2DDA... 87

4.4. Kernel scatter difference two dimensional discriminant Analysis... 95

4.5. Conclusion... 102

General conclusion... 103

Résumé

Les travaux de recherche de cette thèse ont pour objectif principal la recherche de solutions aux problèmes des systèmes de reconnaissance de visage. Ces problèmes sont dûs à de nombreux facteurs qui ont un impact sur la performance de ces systèmes biométriques. Parmi les facteurs déterminants, il y a la variabilité de la luminosité, les expressions faciales, l’angle de capture du visage et les bruits qui entâche les images prises par les caméras. Dans le but de résoudre ces problèmes, cette thèse propose de nouvelles méthodes basées sur l’analyse en composantes principales (PCA) et l’analyse discriminante linéaire (LDA). Elle procède d’abord par l’extension de ces deux techniques au cas 2D (2DPCA+2DLDA), ce qui a permis de réduire en même temps le bruit présent dans les images et le temps de calcul. Les résultats de tests sur des bases libres obtenus avec ces deux méthodes sont très satisfaisants. Cependant, la méthode LDA souffre du problème de singularité appelé Small Sample Size problem (SSS). Pour remédier à ce problème, cette thèse propose deux versions de l’analyse discriminante qui utilisent une fonction de pondération des matrices intra classe et inter classe RW2DLA et WS2DDA. Les tests expérimentaux avec ces deux méthodes ont montré leure bonne efficacité dans la reconnaissance de visage. Enfin, pour améliorer le taux de reconnaissance, cette thèse utilise des fonctions noyaux pour obtenir des versions non linéaires de l’analyse discriminante KRW2DDA et KWS2DDA. Ceci dans le but de travailler dans un espace des caractéristiques où la séparation des classes est efficace. Les résultats des tests sont satisfaisants comparées à ceux des versions linéaires.

Mots clés : Biométrie, reconnaissance de visage, Small Sample Size problem, analyse en

Résumé détaillé

La biométrie de nos jours joue un rôle important dans la sécurité des nations. Elle est devenu nécessaire pour l’identification des personnes et la vérification de leur identité. Elle est basée sur l’identification des traits physiologiques d’une personne (empreintes digitales, visage, rétine, iris, etc) ou sur l’analyse des comportements de l’individu (marche, tracé de signature, frappe sur un clavier, etc.) ou bien sur l’analyse des traces biologiques (ADN, sang, salive, etc.). En particulier, la reconnaissance du visage peut être utilisée pour identifier une personne mais il est plus communément utilisé pour vérifier l'identité. Elle a connu beaucoup de succès dans des applications de sécurité comme la vidéosurveillance, le contrôle d’accès aux zones sensibles, le passeport biométrique, etc.

Dans ce contexte, les travaux de recherche de cette thèse ont pour objectif principale la recherche de solutions aux problèmes des systèmes de reconnaissance de visage. En effet, le taux de reconnaissance de ces systèmes qui est en général très important au laboratoire où les conditions d’éclairage sont contrôlées (jusqu’à 98%), diminue largement dans un environnement réel. Ceci est du à la variabilité de la luminosité et aux effets de bruits et d’ombrage qui entachent les images prises par les caméras. D’autres facteurs sont aussi déterminants dans la reconnaissance à savoir les différentes expressions faciales, l’angle de prise de vue, le port de lunettes, la couleur de la peau, la barbe et le moustache. Dans le but de résoudre ces problèmes, cette thèse propose de nouvelles méthodes basées sur l’analyse en composantes principales (PCA) et l’analyse discriminante linéaire (LDA). Elle procède d’abord par l’extension de ces deux techniques au cas 2D (2DPCA+2DLDA), ce qui a permis de réduire en même temps le bruit présent dans les images et le temps de calcul. Les résultats

des tests sur des bases libres obtenus avec ces deux méthodes sont très satisfaisants. Cependant, la méthode LDA souffre du problème de singularité appelé Small Sample Size problem (SSS). Pour remédier à ce problème, cette thèse propose deux versions de l’analyse discriminante qui utilisent une fonction de pondération des matrices intra classe et inter classe Relevance weighted 2DLDA (RW2DLDA) et Weighted scatter-difference-based 2DDA (WS2DDA). Les tests expérimentaux avec ces deux méthodes ont montré leur bonne efficacité dans la reconnaissance de visage. Enfin, pour améliorer le taux de reconnaissance, cette thèse utilise des fonctions noyaux pour obtenir des versions non linéaires de l’analyse discriminante KRW2DDA et KWS2DDA. Ceci dans le but de travailler dans un espace des caractéristiques où la séparation des classes est plus efficace. Les résultats des tests ont montré des performances meilleures en taux de reconnaissance comparées à ceux des versions linéaires.

Ce mémoire de thèse est structuré en quatre chapitres :

Le premier chapitre est consacré à la présentation de la biométrie d’une façon générale et les mesures de performance adoptées pour les systèmes biométriques. Il relate ensuite les différentes approches pour résoudre les problèmes de reconnaissance de visage tout en précisant leurs avantages et leurs inconvénients.

Le deuxième chapitre montre à travers une bibliographie riche et exhaustive les avantages et les inconvénients des méthodes statistiques. En particulier, il décrit de façon détaillée la méthode d’analyse en composantes principales PCA et la méthode d’analyse discriminante linéaire LDA, ainsi que leurs variantes les plus utilisées dans la littérature Kernel PCA Kernel LDA, Fuzzy PCA et Fuzzy LDA. Ces variantes ont été développées dans le but de surmonter les limitations de PCA et LDA à savoir : la perte d’information en réduisant l’espace des données avec PCA, le problème de singularité dû aux critère d’optimisation de LDA, l’imprécision des données image acquises par les caméras et le temps de calcul relativement long.

Dans le chapitre 3 une extension 2D des méthodes PCA et LDA est proposée afin de réduire le temps de calcul et garder un taux de reconnaissance meilleur. Les résultats obtenus des

tests sur des bases standards ont montré l’efficacité de ces méthodes ( 2DPCA, 2DLDA et 2DPCA+2DLDA) par rapport aux méthodes classiques dans le cas 1D. Ensuite, la résolution du problème de singularité appelé SSS problem est abordée. Deux nouvelles versions bidimensionnelles de RWLDA et WSDA ont été développées pour résoudre ce problème. Elle ont montré leur efficacité en terme de taux de reconnaissance lors des tests menés sur des bases libres de visages.

Finalement, le quatrième chapitre est dévolu à l’utilisation des versions non linéaires de RWLDA et WSDA en utilisant des fonctions noyaux (kernel). Le but de ces transformations est de travailler dans l’espace des caractéristiques de dimension supérieure à celle de l’espace des données réelles. Cela a l’énorme avantage de mieux séparer les classes et par suite améliorer le taux de reconnaissance. Les résultats des tests sont satisfaisants comparées aux versions linéaires.

Mots clés: Biométrie, reconnaissance de visage, Small Sample Size problem, analyse en

Abstract

This research deals with problems in facial recognition systems, where face recognition is mainly determined by many factors including face expression, lighting, noise and the camera’s angle at which the image is taken. We have developed novel algorithms based on basic principle Component analysis (PCA) and linear discriminant analysis (LDA) methods to address these problems. We have extended these methods to Two Dimension (2D) and combined them to propose a new third method and we refer to it as the 2DPCA+2DLDA. This latter method has reduced the effect of noise as measured by recognition rate. However, it has increased the computation cost due to the complexity in its algorithm.

To overcome the Small Sample Size problem (SSS) of LDA we have developed two methods by weighting the between-class scatter matrix and within-class scatter matrix; relevance weighted two dimensional linear discriminant analysis and weighted scatter difference two dimensional discriminant analysis ( RW2DLA and WS2DDA) methods respectively. Moreover, they have shown better experimental results.

Finally, to improve the recognition rate, we have also proposed two non linear methods using kernel functions, kernel relevance weighted two dimensional discriminant analysis and kernel weighted scatter difference two dimensional discriminant analysis (KRW2DDA and KWS2DDA). These methods can analysis data in implicit feature space in which the separation between classes is efficient. The proposed methods are relatively more effective in face recognition as compared with other competing linear methods.

Keywords: Biometrics, face recognition, Small Sample Size problem, principle component

1

General introduction

For society cohesion and individuals security there is a vital need to distinguish between persons. This need of identifying person’s characteristics comes from the importance of distinguishing between different persons who possess different traits, since at any time we depend on the identification of a person to cooperate with him/her in all areas. So unless we identify who is this person we will face the difficulty to transact with him/her. Accordingly, societies are developing methods and techniques for identifying between their members or to recognize a given person who exposed to changes. Given the fact that some of those members are dangerous criminals, we need to identify them in order to monitor them in order to contain their behaviours. Another related application for the importance of person identification is the existence of some areas where not all people are authorized to enter for security reasons.

Many methods of detecting the identities of people were developed during the history of humans being. Ranging from the early years and past history where persons can be identified by looking at their faces or hearing their voices, to now-day with the rise of computer technology age, we reach the stages where persons are sophistically identified by computers using their IDs and passwords. This sophisticated technology can permit either to activated or processed a person data. This method enables the computer in accessing or processing data very fast because such data has already been stored earlier in computer. Thus, this modern technology is providing the societies with effective and efficient method of person identification.

2

With the growing numbers of associated persons due to their demographic trend and their clustering in bigger size with the result that mass places become very crowded, we are facing the challenge to detect many persons in short time. With the application of computer in person identification, this technology has the advantages to memories information recorded at earlier time, yet this technology requires entering the ID or/and the password for everyone and this entering process consumes a considerable time. So we need another processing method which works in real time. The biometrics methods have the solution of this problem, because they work in real time and some of their technologies (eg. Face recognition method) which doesn’t need to enter any character data such as ID, rather it takes a real time picture of human face.

Face recognition has been an active research area over the last 30 years. It has been studied by scientists from different areas of psychophysical sciences and those from different areas of computer sciences. Psychologists and neuroscientists mainly deal with the human perception part of the topic, whereas engineers focus on machine recognition of human faces, to address the computational aspects of face recognition.

This study which is based on computer application aims to obtain a better outcome in face recognition. This aim entails to develop methods that deal with incomplete and uncertainty features information in comparison to the existing methods of face recognition. In effect, the proposed methods attempted to reduce the impact due to the variation that shape the expressions of facial image which include smiling face or crying face and other expressions, as well as lighting effects and the development of head rotation and also the image size and the effects of the angle at which the image is captured by the camera resolution.

we organized this research study into four chapters. The first two chapters present the pertaining topics of the study domain with explicit statements on its importance, the tools which can be used and the specific research area. In the reminder two chapters we present the developed aspects of the study and report the results.

In chapter one we overviewed the biometrics systems including thier efficiency, biometric technology and most application of biometric systems, overview the most famous applications those are suitable for face recognition technology and show the advantages of

3

applying this technology. Also we explain the survey on face recognition methods focusing human recognition of face and machine recognition of faces. Also we study the statistical approaches for face recognition, media learning application and other approaches. Lastly

we identify the research problem.

In chapter two we concentrate on face recognition technology which pertains to our research. Specifically, we have focused the basic of pattern classification, nearest mean classifier and feature dimension reduction within the framework of this technology. In this chapter we have also provided a brief review of the most important algorithms and methods of face recognition.

In chapter three we develop some facial recognition algorithms under linearity which have shown better results in their performances. we have also provided analysis for the results we have obtained.

chapter four deals with non linearity where we have develop methods to improve facial recognition. Again, we have provided and explanation and commenting on the results that we have obtained.

4

CHAPTER 1

Background and related work

1.1.

Introduction

Humans need to distinguish their personal’s identities among themselves to communicate and to interact with each others. Along the passage of time, many methods for people identification were developed in the literature, within the broad applications of biometric technologies in the access control, law enforcement, security and surveillance systems. The biometrics methods automatically verify and identify individuals using their physiological or behavioral characteristics [1]. Those biometric technologies include the face recognition, the recognition of voice and signature, the identification of finger print, hand geometry, Iris, retina and DNA Sequence Matching [2].

The necessity for personal identification in the fields of private and security systems made face recognition one of the common technology among other biometric technologies. The importance of face recognition rises from the fact that a face recognition system does not require the cooperation of the individual while the other systems need such cooperation.

5

Face recognition algorithms try to solve the problems of both verification and identification [3]. When verification is on demand, the face recognition system is given a face image and it is given a claimed identity.

The system is expected to either reject or accept the claim. On the other hand, in the identification problem, the system is trained by some images of known individuals and then given a test image to decide which individual the test image belongs to.

The problem of face recognition can be considered as mainly a classification problem by given the still images or video of a scene, for identifying one or more persons in the scene by using a stored database of faces, where training the face recognition system with images from the known individuals and classifying the newly coming test images into one of the classes is the main aspect of the face recognition systems [4].

In this chapter we classify these methods into two major groups. The first group includes the old methods while the second group contains the new methods. Due to the framework of the theses, the biometric identification methods which are considered within the new group of identification methods will be overviewed in more details.

1.2.

Biometrics overviews

Biometrics technologies refer to the automatic identification of a living person based on the authentication process which determines the credentials that validate a user‘s identity, such as passwords, PINs, digital certificates, smart cards, and biometrics. Typically, authentication in an IDM system supports end-to-end authentication across multiple systems by supporting a SSO or reducing sign-on to all applications (Nikols & Gebel, 2006).

Now a day, there are many types of biometric technologies developed (e.g. face recognition, fingerprint recognition, hand geometry, hand scan, iris recognition, vein recognition, voice and signature) [1][2][3][4][5][6][7][8][9][10]. This technological based method of identification is preferred over old methods which are involving passwords and personal identification numbers (PINs), for various reasons. For instance, the person to be identified is not required to be physically present at the point-of identification. Furthermore,

6

the identification based on biometric techniques obviates the need to remember a password or carry a token.

Historically, with the increased use of computers the managing access to information technology (IT) requires direct interaction with a limited number of users. Increasingly, managing access involves handling not only increased numbers of internal and external users but also partners beyond the information technology environment at any institution (Razavi & Iverson, 2008).

The overhead created by this increased need for access to IT can be substantial as institutions transition from traditional paper-based systems to paperless digital IT in automating and streamlining administrative services, and business operations. The transformation to an all-digital format raises challenges related to protecting IT. According to Amer and Hamilton (2008), electronic fraud and identity theft are among the biggest risks to the effectiveness of IT. With the increased of IT it is necessary to restrict access to sensitive/personal data. By replacing PINs, biometric techniques can potentially prevent unauthorized access to or fraudulent use of, cellular phones, smart cards, desktop personal computers, workstations, and computer networks. PINs and passwords may be forgotten, and token based methods of identification like passports and driver's licenses may be forget, stolen, or lost. Thus, biometric based systems of identification received a considerable interest. Various types of biometric systems are being used for real-time identification. The most popular are based on face recognition, iris recognition and fingerprint matching. However, there are other biometric systems that utilize retinal scan, voice, signatures and hand geometry.

A biometric system is essentially a pattern recognition system which makes a personal identification by determining the authenticity of a specific physiological or behavioral characteristic possessed by the user. An important issue in designing a practical system is to determine how an individual is identified. Depending on this context, a biometric system can be either a verification (authentication) system or an identification system.

1.2.1. Biometrics efficiency measurements and accuracy indicators

Biometric technological applications never produce a 100% match due to many factors including: variations in environmental conditions, differences in biometric sensors, as well as

7

temporary or permanent bodily changes respectively. Hence, the level of correspondence which is considered a “match” must be defined by using a threshold. Theoretically, there are four possible outcomes of a biometric test:

-Accepting a genuine person who has been recognized. - Rejecting an impostor who fails the test.

- Falsely rejecting a genuine person. - Falsely accepting an impostor

Biometrics technologies will therefore never be 100% error-free [11].

On the other hand, the possible outcomes define the following two main measurements in use for measuring biometrics errors:

- False Acceptance Rate (FAR): The (FAR) is a rate of falsely accept comparison and consider it is true while it is not true.

- False Rejection Rate (FRR): The (FRR) is a rate of falsely reject comparison and consider it is false while it is true. The intersection of the overlapping distributions corresponding to (FAR) and (FRR) respectively defines the equal error rate (EER) (Figure 1.1). The evaluation performance of a given system can be measured by using the EER that provides threshold. - Recognition rate (RR) in this study we have used as measure of accuracy the recognition rate defined as the ratio of the number of correctly recognized faces to the total number of faces actually exist in the picture. Therefore, RR = 1- FAR.

8 1.2.2 Biometrics technology characteristics

The following are the major characteristics in estimating the efficiency of the biometrics technology [6][7][8][9][10][12][13]:

-Universal: The ability of a biometric technology to cover all description details and it does not need an additional factor to fully identify persons.

-Uniqueness: By uniqueness we refer to the capability of the biometric technology to uniquely identify persons without any probability of duplicating the features of more than one person.

- Permanence: The Permanence is the constancy of the features of the biometric, with the change of the time or any other factor.

- Collectability: The Collectability is the ease of collecting and entering the biometric technology.

9

- Performance: The performance of the biometric is the comparison through the computer programs.

- Acceptability: The acceptability is the ability of cooperating with different persons without any type of hurt or critical situations.

- Circumvention: Prevent the ability of fraudulence and forge.

1.2.3 The applications of biometrics

We will examine many types of biometric technological applications (e. g DNA, voice recognition, signature verification, handwriting, gait recognition, fingerprint, hand geometry, iris recognition, retina recognition, ears recognition and face recognition). However, we restrict the review only to the related biometrics technologies pertaining to our study:

Iris recognition

The human iris (Figure 1.2) is an annular part between the pupil and the white sclera, emerging as a highly reliable biometric trait for personal identification. Although the area of the iris is small it has enormous pattern variability which makes it almost unique for every person and hence leads to high reliability [14].

The idea of using iris patterns for personal identification was originally proposed in 1936 by ophthalmologist Frank Burch. In 1987 two other ophthalmologists, Aran Safir and Leonard Flom, patented this idea, and in 1989 they asked John Daugman to try to create actual algorithms for iris recognition [15]. Technologies are the basis for all current iris recognition systems and products [15][ 16].

Iris scans analyze the features that exist in the colored tissues surrounding the pupil which has more than 200 points that can be used for comparison, including rings, furrows and freckles [17][18]. The scans use a regular video camera style and can be done from further away than a retinal scan. It will work through fine glasses and in fact has the ability to create an accurate enough measurement that can be used for identification purposes, and not just verification.

10

Figure 1.2: Different irises

The physiological properties of irises are the major advantages that can be used as a method of authentication. The morphogenesis of the iris occurs during the seventh month of gestation and results in the uniqueness of the iris even between multi-birth children. These patterns remain stable throughout life and are protected by the body’s own mechanisms. Moreover, the randomness in irises makes them very difficult to forge and hence imitate the actual person.

In addition to the physiological benefits, iris-scanning technology is not very intrusive as there is no direct contact between the subject and the camera technology.

Scalability and speed of the technology are also major advantage. The technology is designed to be used with large-scale applications such as with ATMs. Given the iris records as stored in the database in such manner that users do not spend a lot of time being authenticated and the ability of the system to scan and compare the iris within a matter of minutes is a major benefit.

As with any technology there are challenges with iris recognition. The iris is a very small organ to scan from a distance. It is a moving target and can be obscured by objects such as the eyelid and eyelashes. Subjects who are blind or have cataracts can also pose a challenge to iris recognition, as there is difficulty in reading the iris.

The camera used in the process needs to have the correct amount of illumination. Without this, it is very difficult to capture an accurate image of the iris. Along with illumination comes the problem with reflective surfaces within the range of the camera as well as any unusual

11

lighting that may occur. All of these impact the ability of the camera to capture an accurate image.

Indeed, enrolling a non-cooperative subject would prove to be very difficult. Inadequate training of users at the initial enrolment period will cause problems both at the initial enrolment time and subsequent authentications. Frustrated users will not help make the system any easier to use and will not be accepted by users as a convenient authentication method. Communication with users plays a major part in introducing such a system successfully.

Normal day-to-day problems such as system failures, power failures, network problems, and software problems can all contribute to rendering a biometric system to be unusable. Once users get accustomed to such a system it is unlikely that they will remember to bring their other forms of identification with them to the office. System administrators also have the additional pressure of ensuring the System that stores the iris record’s database is properly secured to prevent tampering with the data stored. Although these are not major hindrances to the actual iris recognition system, it is important to take these things into consideration and have a backup plan.

Fingerprint method

Fingerprint identification, known as dactyloscopy [19], is the process of comparing two instances of friction ridge skin impressions, to determine whether these impressions could have come from the same individual. The flexibility of friction ridge skin means that no two fingers are ever exactly alike in every detail; even two impressions recorded immediately after each other from the same hand may be slightly different. Fingerprint identification, also referred to as individualization, involves an expert, or a computer system operating under threshold scoring rules, determining whether two friction ridge impressions are likely to have originated from the same finger.

Figure1.3: An image of a fingerprint created by the friction ridge structure.

An intentional recording of friction ridges is usually made with black printer's

across a contrasting white background, typically a white card. Friction ridges can also be recorded digitally.

The main advantages of this technology are

economical biometric user authentication technique, easy to use

for the biometric template, reducing in the size of the database memory required and standardized. The disadvantages, on the other hand are as f

people, because it is still related to criminal identification, make mistakes with the dryness or dirty of the finger’s skin, as well as with the age (is not appropriate with children, because the size of their fingerprint changes quick

Face recognition

The face of the human is an important part of who that person is and how people identify him. Face recognition is an area of biometrics technologies that focuses in how we can recognize and identify the identity of a person using computer as a tool

input data [20][21].

Facial recognition systems are built on computer programs that analyze the characteristics of a person's face images which is the input generated through a digital camera [20][21][22][23][24] [25]. The programs take a

12

An image of a fingerprint created by the friction ridge structure.

An intentional recording of friction ridges is usually made with black printer's

across a contrasting white background, typically a white card. Friction ridges can also be

The main advantages of this technology are as follows: very high accuracy, user authentication technique, easy to use, small storage space required for the biometric template, reducing in the size of the database memory required and standardized. The disadvantages, on the other hand are as follows: very

people, because it is still related to criminal identification, make mistakes with the dryness or dirty of the finger’s skin, as well as with the age (is not appropriate with children, because the size of their fingerprint changes quickly).

The face of the human is an important part of who that person is and how people identify him. Face recognition is an area of biometrics technologies that focuses in how we can recognize and identify the identity of a person using computer as a tool

Facial recognition systems are built on computer programs that analyze the characteristics of a person's face images which is the input generated through a digital camera

]. The programs take a facial image, measure characteristics (F An image of a fingerprint created by the friction ridge structure.

An intentional recording of friction ridges is usually made with black printer's ink rolled across a contrasting white background, typically a white card. Friction ridges can also be

ws: very high accuracy, most , small storage space required for the biometric template, reducing in the size of the database memory required and lows: very intrusive for some people, because it is still related to criminal identification, make mistakes with the dryness or dirty of the finger’s skin, as well as with the age (is not appropriate with children, because the

The face of the human is an important part of who that person is and how people identify him. Face recognition is an area of biometrics technologies that focuses in how we can recognize and identify the identity of a person using computer as a tool and face image as

Facial recognition systems are built on computer programs that analyze the characteristics of a person's face images which is the input generated through a digital camera ge, measure characteristics (Figure

13

1.4) such as the distance between the eyes, the length of the nose, the angle of the jaw and create a unique file called a "template". Using these templates, the software then compares that input image with other images in the database and produces a score that measures how similar the images are to each other.

Figure 1.4: Some of the main points used in feature measurements of face recognition

Facial recognition is the one of biometric technology that has wide public acceptance and is ranked as the least intrusive of all other technologies [26]. Face recognition is probably one of the most nonintrusive and user-friendly biometric authentication methods currently available [27][28][29][30][31].

A facial recognition system is a computer application or device that can identify individuals based on their unique facial characteristics. Unlike many other identification methods (e.g., fingerprints, voiceprint, signature), they do not need to make direct contact with an individual in order to verify their identity. This can be advantageous in clean environments, for surveillance or tracking, and in automation systems. A reference model of the individual, and captures their images for identification, there may be concerns about how the system is perceived by its users.

14 Voice recognition

Voice recognition is the technology by which sounds, words or phrases spoken by humans are converted into electrical signals, and these signals in turn are transformed into coding patterns to which meaning has been assigned [32]. While the concept more generally could be called "sound recognition", we focus here on the human voice because frequently and most naturally we use our voices to communicate our ideas to others in our immediate surroundings.

Each human voice is different, and identical words can have different meanings if spoken with different inflections or in different contexts. Several approaches have been tried, with varying degrees of success, to overcome these difficulties. The main advantages of this technology include: non intrusive with high social acceptability, verification time is about five seconds and cheap technology. But the disadvantages include: person’s voice can be easily recorded and used for unauthorized, low accuracy and an illness such as a cold can change in a person’s voice makes absolute identification difficult or impossible.

1.3.

Survey on face recognition methods

The problem of face recognition has been one of the most prominent areas of machine vision [23]. Face recognition is desirable in having a system that has the ability of learning to recognize unknown faces [33].

Face recognition is a difficult visual representation task in large part because it requires differentiating among objects which vary only subtly from each other. It represents a particularly interesting case within the context of object recognition. Moreover, it reconciles between rigidity and flexibility in the number of configurations as it is much less constrained compared to the case of the often used rigid machined objects and the general case which includes objects that could be posed in an infinite number of configurations with the least constraints [34].

Face Recognition has been an interesting issue for both neuroscientists and computer engineers dealing with artificial intelligence. A healthy human can detect a face easily and

15

identify that face, whereas for a computer to recognize faces, the face area should be first detected and recognition comes next. Hence, for a computer to recognize faces the photographs should be taken in a controlled environment; a uniform background and identical poses makes the problem easy to solve. These face images are called mug shots [35]. From these mug shots, canonical face images can be manually or automatically produced by some preprocessing techniques like cropping, rotating, histogram equalization and masking.

In physiology, the earliest work on face recognition can be traced back at least to the 1950s and in the engineering literature, the studies started on the 1960s [5]. On the other hand, the early research on machine recognition of faces initiated in the 1960s [36] [37]. The first work during this period is the work of Bledsoe [37], while the start of research on automatic machine recognition of faces was in the 1970 [5][38]. Since the early 1970's (Kelly, 1970). Face recognition has drawn the attention of researchers in fields from security, physiology, and image processing, to computer vision [38]. Also, there is the seminal work of Kanade done in 1973[5]. Kanade was the first developer of a fully automatic version of feature based system [37]. In the 1980s, the computer researchers focus on the visual representation which was reflected into the face recognition. They began experimenting the visual representation, making use of the appearance or texture of face images [37]. During that decade, many connectionist methods were developed. The 1990s can be considered as the main decade of face recognition boom. In this period there was growing interest in the technology at both state and private levels [36]. Numerous algorithms have been proposed for face recognition. The template or appearance-based techniques were submitted for further developments. These developments were prompted by the ground-breaking work of Kirby and Sirovich with Karhunen-Loeve Transform of faces, which led to the Principal Component Analysis (PCA) “eigenface” technique of Turk and Pentland [37].

1.3.1. Human recognition of face

When building artificial face recognition systems, scientists try to understand the

architecture of human face recognition system. Focusing on the methodology of human face recognition system may be useful to understand the basic system. However, the human face recognition system utilizes more than that of the machine recognition system which is just

2-16

D data. The human face recognition system uses some data obtained from some or all of the senses; visual, auditory and tactile are among the data sources. The collected data is used either individually or collectively for storage and remembering of faces.

In many cases, the surroundings also play an important role in human face recognition system. It is hard for a machine recognition system to handle so much data and their combinations. However, it is also hard for a human physical capacity to remember many faces due to storage limitations. A key potential advantage of a machine system is its memory capacity [5], whereas for a human face recognition system the important feature is its parallel processing capacity. The issue which features humans use for face recognition has been studied in the literature of biometric technology and it has been argued that both global and local features are useful for face recognition [5].

It is harder for humans to recognize faces which they consider as either attractive or otherwise. The low spatial frequency components are used to clarify the sex information of the individual whereas low and high frequency components are used to identify the individual. The low frequency components are used for the global description of the individual while the high frequency components are required for finer details needed in the identification process.

Both feature information and other environmental pertaining data are important for the human face recognition system. Studies suggest the possibility of global descriptions serving as a front end for better feature-based perception [5]. If there are dominant features present such as big ears or a small nose, then the feature information is suffice to describe the system. Hair, eyes, mouth, face outlines have been determined to be more important than nose for perceiving and remembering faces. It has also been found that the upper part of the face is more useful than the lower part of the face for recognition. Also, aesthetic attributes (e.g. beauty, attractiveness or pleasantness.) play an important role in face recognition; the more attractive the faces, the easily are to be remembered.

For humans, photographic negatives of faces are difficult to recognize. Still, there is not much study on why it is difficult to recognize negative images of human faces. Also, a study

17

on the direction of illumination [39] showed the importance of top lighting; it is easier for humans to recognize faces illuminated from top to bottom than the faces illuminated from bottom to top.

According to the neurophysicists, the analysis of facial expressions is done in parallel to face recognition in human face recognition system. Some prosopagnosia patients, who have difficulties in identifying familiar faces, seem to recognize facial expressions due to emotions. Patients who suffer from organic brain syndrome do poorly at expression analysis but perform face recognition quite well.

1.3.2. Machine recognition of faces

Although studies on human face recognition were expected to be a reference on machine recognition of faces, research on machine recognition of faces has been developed independently from studies on face recognition by humans. During 1970’s, typical pattern classification techniques, which use measurements between features in faces or face profiles, were used [4]. During the 1980’s, work on face recognition remained nearly stable. Since the early 1990’s, research interest on machine recognition of faces has grown tremendously. The reasons may be listed in (i) An increase in emphasis on civilian/commercial research projects (ii) The studies on neural network classifiers with emphasis on real-time computation and adaptation (iii)The availability of real time hardware (iv)The growing need for surveillance applications.

The basic question relevant for face classification is that; what form the structural code (for encoding the face) should take to achieve face recognition. Two major approaches are used for machine identification of human faces; geometrical local feature based methods, and holistic template matching based systems. Also, combinations of these two methods, namely hybrid methods, are used.

The geometrical local feature approach extracts and measures discrete local features (such as eye, nose, mouth, hair, to mention) for retrieving and identifying faces. Then, standard statistical pattern recognition techniques and /or neural network approaches are employed for matching faces using these measurements [40].

18

One of the well known geometrical-local feature based methods is the Elastic Bunch Graph Matching (EBGM) technique.

The holistic approach conceptually related to template matching, attempts to identify faces using global representations [2]. The holistic methods treat face image as a whole and try to extract features from the whole face region. In this approach, as in the previous approach, the pattern classifiers are applied to classify the image after extracting the features. One of the methods to extract features in a holistic system is applying statistical methods such as Principal Component Analysis (PCA) to the whole image.

Whichever method is used, the most important problem in face recognition is what is known as the curse of dimensionality problem. Appropriate methods should be applied for instance to reduce the dimension of the studied space, if that was the case. Working on higher dimension causes over fitting where the system starts to memorize. Also, computational complexity would be an important problem when working on large databases. In the following sections, the main pertaining studies shall be summarized. The recognition techniques are grouped as statistical and learning based approaches.

1.4.

Statistical approaches for face recognition

Statistical methods include template matching based systems where the training and test images are matched by measuring the correlation between them. Moreover, statistical methods include the projection based methods such as Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA). In fact, projection based systems came out due to the shortcomings of the straightforward template matching based approaches; that is, trying to carry out the required classification task in a space of extremely high dimensionality.

More technique Brunelli and Poggio [41], suggest that the optimal strategy for face recognition is holistic and corresponds to template matching. In their study, they compared a geometric feature based technique with a template matching based system. In the simplest form of template matching, the image (as 2-D intensity values) is compared with a single template representing the whole face using a distance metric.

19

Although recognition by matching raw images has been successful under limited circumstances, it suffers from the usual shortcomings of straightforward correlation-based approaches, such as sensitivity to face orientation, size, variable lighting conditions, and noise. The reason for this vulnerability of direct matching methods lies in their attempt to carry out the required classification in a space of extremely high dimensionality. In order to overcome the curse of dimensionality, the connectionists first employ the equivalent of data compression methods.

However, it has been successfully argued that the resulting feature dimensions do not necessarily retain the structure needed for classification, and that more general and powerful methods for feature extraction such as projection based systems are required. The basic idea behind projection based systems is to construct low dimensional projections of a high dimensional point cloud, by maximizing an objective function such as the deviation from normality.

1.4.1. Face detection and recognition by PCA

The Eigenface Method of Turk and Pentland [42] is one of the main methods applied in the literature which is based on the Karhunen- Loeve expansion. Their study is motivated by the earlier work of Sirowich and Kirby [43][44].

It is based on the application of Principal Component Analysis to the human faces. It treats the face images as 2-D data, and classifies the face images by projecting them to the eigenface space which is composed of eigenvectors obtained by the variance of the face images. The Eigenface method of facial recognition is considered to be the first working facial recognition technology [45].

When the method was first proposed by Turk and Pentland, they worked on the image as a whole. Also, they used Nearest Mean classifier to classify the face images [42]. By using the observation that the projection of a face image and non-face image are quite different, a method of detecting the face in an image is obtained. They applied the method on a database of 2500 face images of 16 subjects digitized at all combinations of 3 head orientations, 3 head sizes and 3 lighting conditions. They conducted several experiments to test the robustness of

20

their approach to illumination changes, variations in size, head orientation, and the differences between training and test conditions. They reported that the system was fairly robust to illumination changes, but degrades quickly as the scale changes [42]. This can be explained by the correlation between images obtained under different illumination conditions; the correlation between face images at different scales is rather low. The eigenface approach works well as long as the test image is similar to the training images used for obtaining the eigenfaces.

Later, derivations of the original PCA approach were proposed for different applications. In their study Moghaddam and Pentland used the Eigenface Method for image coding of human faces for potential applications such as video telephony, database image compression and face recognition [46]. Lee et al. [47] proposed a method using PCA which detects the head of an individual in a complex background and then recognize the person by comparing the characteristics of the face to those of known individuals. Lota et al. [48] proposed a method for generalizing the representational capacity of available face database.

In a study by Crowley and Schwerdt [49], PCA is used for coding and compression for video streams of talking heads. They suggest that a typical video sequence of a talking head can often be coded in less than 16 dimensions.

Another application method, is the Bayesian PCA method suggested by Moghaddam et al. [50][51][52][53][54][55][56][57]. By this system, the Eigenface Method based on simple subspace-restricted norms is extended to use a probabilistic measure of similarity. Also, another difference from the standard Eigenface approach is that method which uses the image differences in the training and test stages. The difference of each image belonging to the same individual with each other is fed into the system as intrapersonal difference, and the difference of one image with an image from different class is fed into the system as extra personal difference. Finally, when a test image comes, it is subtracted from each image in the database and each difference is fed into the system. For the biggest similarity (i.e. smallest difference) with one of the training images, the test image is decided to be in that class. The mathematical treatment is mainly studied in [58]. Moghaddam in [59] introduced his study on several techniques; Principal Component Analysis (PCA), Independent Component Analysis

21

(ICA), and nonlinear Kernel PCA (KPCA). He argued that the experimental results demonstrate the simplicity, computational efficiency and performance superiority of the Bayesian PCA method over other methods.

Chung et al. suggested the use of PCA and Gabor Filters together [60]. Their method consists of two parts: In the first part, Gabor Filters extracted facial features from the original image on predefined fiducially points.

In the second part, PCA is used optimally to classify the facial features. They suggest the use of combining these two methods in order to overcome the shortcomings of PCA. They argued that, when raw images are used as a matrix of PCA, the eigenspace cannot reflect well the correlation of facial feature, as original face images have deformation due to plane, in-depth rotation and illumination and contrast variation. Also they argued that, they have overcome these problems using Gabor Filters in extracting facial features.

Recently, Yang et al. developed a Two Dimensional PCA (2DPCA) [61]. In the use of 2DPCA they obtained many advantages over the PCA method.

1.4.2. Face recognition by LDA

Etemad and Chellappa [62], proposed a method on appliance of Linear/Fisher Discriminant Analysis for the face recognition process.

LDA is carried out via scatter matrix analysis. The aim is to find the optimal projection which maximizes between class scatter of the face data and minimizes within class scatter of the face data. As in the case of PCA, where the eigenfaces are calculated by the Eigen value analysis, the projections of LDA are calculated by the generalized Eigen value equation. An alternative method which combines PCA and LDA is studied [63][64][65][66]. This method consists of two steps; the face image is projected into the eigenface space which is constructed by PCA, and then the eigenface space projected vectors are projected into the LDA classification space to construct a linear classifier. In this method, the choice of the number of eigenfaces used for the first step is critical. The choice enables the system to generate classes of similar features via LDA from the eigenface space representation. The

22

generalization/over fitting problem can be solved in this manner. In these studies, a weighted distance metric guided by the LDA Eigen values was also employed to improve the performance of the subspace LDA method.

Several authors independently proposed techniques which extract features directly from 2D images without a factorization preprocessing. The methods were termed two-dimensional Discriminate Analysis (2DLDA) as generalization of classical LDA [67], [68], [69].

Moreover, in [70] it has shown that the class separability criterion that classical LDA maximizes is not necessarily representative of classification accuracy and the resulting projection will preserve the distances of already well separated classes, while causing unnecessarily overlap of neighboring classes. To solve this problem they have proposed an extended criterion by introducing a weighting scheme in the estimation of between-class scatter matrix. From the similar standpoint in [71] it has extended this concept to estimate the within-class scatter matrix by introducing the inter-class relationships as relevance weights. He has presented an LDA enhancements algorithm namely relevance weighted LDA (RW-LDA) by replacing the unweighted scatter matrices through the weighted scatter matrices in the classical LDA method. This was successful application in face recognition due to Chougdali et al. [72]. Waiyawut and Yuttapong explained efficient face recognition methods called the weighted
2DLDA which in incorporating weighted outlier class relationships in to the estimation of the overall between-class scatter matrix [73].

1.4.3. Feature-based approaches

Bobis et al. studies a feature based face recognition system. They suggested that a face can be recognized by extracting the relative position and other parameters of distinctive features such as eyes, mouth, nose and chin [74]. The system described the overall geometrical configuration of face features by a vector of numerical data representing position and size of main facial features. Starting by extracted eyes coordinates, they used the intraocular distance and eyes position to determine size and position of the areas of search for face features. In these areas binary threshold is performed, system modifies threshold automatically to detect features. In order to find their coordinates, discontinuities are searched for in the binary image. They claim that, their experimental results showed that their method

23

is robust, valid for numerous kind of facial images in real scene. Works in real time with low hardware requirements and the whole process is conducted automatically.

Cagnoni and Poggi suggested a feature-based approach instead of a holistic approach to face recognition [75]. They applied the eigenface method to sub-images (eye, nose, and mouth). They also applied a rotation correction to the faces in order to obtain better results. Guan and Szu compared the performance of PCA and ICA on face images [76]. They argue that, ICA encodes face images with statistically independent variables, which are not necessarily associated with the orthogonal axes, while PCA is always associated with orthogonal eigenvectors.

While PCA seeks directions in feature space that best represents the data in a sum-squared error sense, ICA seeks directions that are most independent from each other [77]. They also argue that, both pixel-based algorithms have the major drawback that they weight the whole face equally and therefore lack the local geometry information. Hence, Guan and Szu suggest approaching the face recognition problem with ICA or PCA applied on local features [78]. Martinez proposed a different approach based on identifying frontal faces [78]. His approach divides a face image into N different regions, analyzes each region with PCA, and then uses a Bayesian approach to find the best possible global match between a probe and database image. The relationship between the N parts is modeled by using Hidden Markov Models (HMMs).

1.5.

Media learning application

Er et al. [79], suggested the use of Radial Basis Function (RBF) Neural Networks on the data extracted by discriminant Eigen features. They used a hybrid learning algorithm to decrease the dimension of the search space in the gradient method, which is crucial on optimization of high dimension problem. First, they tried to extract the face features by both the PCA and LDA methods. Next, they presented a hybrid learning algorithm to train the RBF Neural Networks, so the dimension of the search space is significantly decreased in the gradient method.

24

Thomaz et al. [80], also studied on combining PCA and RBF neural network. Their system is a face recognition consisting of a PCA stage which maps the projections of a face image over the principal components into a RBF network acting as a classifier. Their main concern is to analyze how different network designs perform in a PCA+RBF face recognition system.

1.5.1. Face recognition by neural network

There are some other approaches which use both statistical pattern recognition techniques and Neural Network systems.

Neural Network approaches have been used in face recognition generally in a geometrical local feature based manner, but there are also some methods where neural networks are applied holistically.

Temdee et al. presented a frontal view face recognition method by using fractal codes which are determined by a fractal encoding method from the edge pattern of the face region (covering for instance eyebrows, eyes and nose) [81]. In their recognition system, the obtained fractal codes are fed as inputs to a back propagation neural network for identifying an individual. They tested their system performance on the ORL face database. They report their performance as 85 % correct recognition rate in the ORL face database.

Lades et al. presented an object recognition system based on Dynamic Link Architectures, which is an extension of the artificial neural networks [82]. The DLA uses correlations in the fine-scale cellular signals to group neurons dynamically into higher order entities.

These entities can be used to code high-level objects, such as a 2-D face image. The face images are represented by sparse graphs, whose vertices are labeled by a multi resolution description in terms of local power spectrum, and whose edges are labeled by geometrical distance vectors. Face recognition can be formulated as an elastic graph matching, which is performed in this study by stochastic optimization of a matching cost function.

25

Wiskott et al. presented a geometrical local feature based system for face recognition from single images out of a large database containing one image per person, which is known as Elastic Bunch Graph Matching (EBGM) [83]. In this system, faces are represented by labeled graphs, based on a Gabor Wavelet Transform (GWT). Image graphs of new faces are extracted by an Elastic Graph Matching process and can be compared by a simple similarity function. In this system, phase information is used for accurate node positioning and object-adapted graphs are used to handle large rotations in depth. The image graph extraction is based on the bunch graph, which is constructed from a small set of sample image graphs. In contrast to many neural-network systems, no extensive training for new faces or new object classes is required. Only a small number of typical examples have to be inspected to build up a bunch graph, and individuals can then be recognized after storing a single image.

The system inhibits most of the variances caused by position, size, expression and pose changes by extracting concise face descriptors in the form of image graphs. In these image graphs, some predetermined points on the face (such as eyes, nose and mouth) are described by sets of wavelet components (jets). The image graph extraction is based on the bunch graph, which is constructed from a small set of image graphs.

Yoshitomi et al. used thermal sensors to detect temperature distribution of a face [84]. In this method, the front-view face in input image is normalized in terms of location and size, followed by measuring the temperature distribution, the locally averaged temperature and the shape factors of face. The measured temperature distribution and the locally averaged temperature are separately used as input data to feed a Neural Network, while the values of shape factors are used for supervised classification. By integrating information from the Neural Network and supervised classification, the face is identified.

The disadvantage of visible ray image analysis that the accuracy of face identification is strongly influenced by lighting condition including variation of shadow, reflection and darkness is overcome by this method which uses infrared rays.

26

1.5.2. Face recognition by support vector machine SVM

Phillips applied SVM to face recognition. Face recognition is a K-class problem, where K is the number of known individuals; and SVM is a binary classification method [85]. By reformulating the face recognition problem and reinterpreting the output of the SVM classifier, they developed a SVM-based face recognition algorithm. They formulated the face recognition problem in difference space, which models dissimilarities between two facial images. In difference space, they formulated the face recognition as a two class problem. The classes are; dissimilarities between faces of the same person and dissimilarities between faces of different people. By modifying the interpretation of the decision surface generated by SVM, they generated a similarity metric between faces, learned from examples of differences between faces.

1.6. Other approaches

1.6.1. Transformation based systems

There are several studies on face recognition using different type of transformation. One is a method proposed by Podilchuk and Zang which can find the feature vectors using discrete-cosine transform (DCT) where their system tries to detect the critical areas of the face [86]. The system is based on matching the image to map of invariant facial attributes associated with specific areas of the face. They claim that this technique is quite robust, since it relies on global operations over a whole region of the face. A code book of feature vectors or code words is determined for each person from the training set.

They examine recognition performance based on feature selection, number of features or code book size, and feature dimensionality. For feature selection, they tried several block-based transformations and the K-means clustering algorithm to generate the code words for each code book [77]. They argue that the block-based DCT coefficients produce good low-dimensional feature vectors with high recognition performance.

This brings the possibility of performing face recognition directly on a DCT-based compressed bit stream without having to decode the image.