Biometric Systems Biometric Modalities Biometric Performance

Published Date: 02 Nov 2017

INTRODUCTION

In this chapter introduction to Biometric systems,Biometric modalities,Biometric performance, multimodal biometric system,Literature survey,problem statement,scope and objectives,organization of the thesis are discussed.

INTRODUCTION

Establishing person identity is a critical issue to a wide variety of applications such as access control,electronic commerce,banking,communications etc.,.Knowledge based methods like pass words or token based methods like ID cards,Pan cardsetc., can be stolen,shared or forgotten.Biometric person identification is described as identifying a person based on his/her physical or behavioral characteristics.A biometric system verifies the user characteristics like face,finger print,palm,iris,voice,gait,signature etc., for a reliable authentication system.Unimodal biometric system uses a single modality of a person for identity verification.Claimant input identity is verified against stored template in the data base.Biometric system performs in four levels of operation.Sensor level which captures the input data,feature level which extracts the salient features of the captured image,score level which finds the match score by comparing it with the stored template in the data base and decision level which helps in establishing the identityof the input image.Unimodal biometric system is subjected to various limitations such as noisy sensor data,spoof attacks,inter class variations,intraclass variations etc.,.These can be overcome by using multiple source of information of a person in in establishing his identity called a multimodal biometric system.Designing a biometric system is a research challenge as it has to fulfill the conditions of universality,uniqueness,permanence, acceptability, performance, circumvention, and collectability. Biometric systems are deployed in various commercial, civilian, forensic, communication and networking which demands a reliable, highly secured systems for authentication.Different approaches using various biometric characteristics are proposed by researchers. No biometric system is optimal. Biometric system efficiency can be increase by with multimodal fusions, selecting a compatible biometric characters, and approaches.

Biometric system

A Biometric system is a pattern recognition system which works in verification or identification mode. In verification mode system compares the query image against its stored template in the database (one to one). In identification mode the query image is compared against all the templates stored in the data base (oneto many). A Biometric system uses physical characters like Iris, face, palm print, finger print or behavioral characteristics like voice, signature, gait for person identification[1]. The four levels of operation of a biometric system namely sensor level, feature level, match score level, decision level is shown in fig 1.1.

http://t2.gstatic.com/images?q=tbn:ANd9GcQxKDg97EqiH2MmK22xL-K6qX1rWErmME2hy2EHyF-q_Qb5CWhR

FIG 1.1A Basic Biometric System

Sensorlevel: The biometric images like iris , face , fingerprint , palm, is captured with sensors like digital cameras, scanners, video clipping, CCTV cameras, depending on the application scenario. For a commercial, civilian applications a low resolutioncamera, and for forensic and criminal detection a high resolution cameras are used.

Feature extraction level: The image captured by the sensor is pre- processed and features are extracted and stored as a template in the data base. During verification or identification the input image features are compared with the stored template features. The features extracted can be shape of an iris, textures or ridges of an finger print, shape of mouth in a face, principal lines of a palm, etc, which are unique for an individual.

Match score level: The extracted features are compared with the stored templates and degree of similarity is measured by estimating the match score. Higher the match score indicates closer with the similarity.

Decision level: Depending on the degree of matching a final decision is taken by the biometric system to decide whether the input image is of a genuine user or of an imposter and system either accepts or rejects the claim.[1],[3].

Biometric system analysis

Biometric system establishes an individual identity based on "who she is" rather than "what she remembers". With the growing security concerns there is a need for reliable biometric system to be designed. Depending on the application context, a biometric system is designed. Biometric system is analyzed taking into consideration the type of the modality used, performance of the system, cost and limitations.

Biometric modalities

Biometric system uses number of physical or behavioral characteristics of a person for identification. Each character used depends on the application environment and has its own strengths and weakness [2] [7]. The various biometric characters are

FACE: Face as a biometric has a universal acceptability and the most common biometric character used for identifying a person. Face biometrics uses facial features like eyes, eyebrows, shape of the face, shape of mouth, model of the face etc., establishing the identity of a person. Face recognition is an active research area as face verification is used in applications like Face tracking, CCTV, criminal detection, Forensic evidence etc. Face recognition is non-intrusive and has high accuracy in recognition in environment.

PALM: Study of palm print is an ancient practice to know the personal details of a person and also for the astrologers. As palm features are unique for an individual and has rich information on it like principal lines, textures, ridges, etc., and are used for identifying a person. Palm as a biometric has high acceptance universally as it is considered as a reliable, user friendly biometric character. Palm features are extracted using high or low resolution cameras.Latent palm print is of growing importance in forensic applications.

Finger print: Finger prints are more successful and popular biometric identifier. Finger print has pattern of ridges and valleys on the surface of the finger tips. It is most widely used low cost biometric system. Finger print of a person has unique information and is used in legal documents for person identification. Finger print of a person may also vary due to skin texture, aging, working conditions (laborer working with cuts and bruces). Finger print biometrics is suitable for a small data base verification and time taken for computation is large.

IRIS: Iris information is highly complex yet a reliable source used as a biometric. For a high secured applications iris biometric character gives a promising results. Iris characters like stripes, pits and furrows are used for personal identification.Government ADHAR UNIQUE IDENTITYscheme uses Iris as a biometric trait. User’s non acceptance due to sensitiveness of eyes, expensive sensors, large failure to enroll rate limits the use of iris as a biometric identifier.

SIGNATURE: Signature is widely accepted biometric character used to identify a person in a day to day transactions. Various Government, Banking, organizations use signature for a person enrollment to system activity. Signature is a behavioral characteristics of a person. Signatures stroke, speed, shape, acceleration helps in person identification. Signature is an easy way of identifying a person. Behavioral character of a person is also analyzed with signatures.

VOICE: Voice is combination of physiological and behavioral characteristics of a person. The shape and size of vocal track, nose, mouth, contributes the voice character. For a secured biometric system voice character has a limited application as voice could be imitated. Noise background, variations in voice due to health hazards, change of voice over a period of time from a child to an adult, contributes for limited applications of voice as a reliable biometric character.

DNA:DNA acquisition and analysis is time consuming and cannot be used for automatic person identification. DNA structure is unique and has highly reliable information that are used for forensic applications, medical analysis of a person. DNA samples are taken from finger nails, blood samples, hair and saliva. DNA is behavioral characteristics of a person.

HAND VEINS:Biometric identification with vein pattern is a recent approach that uses network of blood vessels underneath the skin. Vein pattern of palm is unique and do not vary over a period of time. The property of uniqueness, non-variability, stable, immune to spoof makes the biometric more secured. Many commercial applications uses hand vein as biometric. Vein is not observed by visible light .It is captured using infrared sensors.

http://www.biometria.sk/images/sigma4.jpghttp://machineslikeus.com/sites/default/files/styles/homepage_image/public/iris.jpghttp://www.information-age.com/article_images/articledir_609/304936/5_fullsize.gifhttp://www.idahocriminaldefenselaw.com/uploads/image/fingerprint_0_thumbnail.jpg

http://t3.gstatic.com/images?q=tbn:ANd9GcTd1hVDhiXvWQa1cHKwU3jMOhjbAtqFw_QDC2c4DkXVZ9epflJnhttp://t1.gstatic.com/images?q=tbn:ANd9GcQ-WRsguo4dct4abimBLy4b3P0Wep_15a7J4vlC7wW-xzkk36GDgwhttp://www.inttelix.com/images/face_recognition.jpghttp://4.bp.blogspot.com/-iQobP1PcFcs/T9VKCc0guQI/AAAAAAAAAHI/JMJH4W4BkHc/s1600/6a00d8341c562c53ef0163058d122b970d-800wi.jpg

Fig 1.2 BIOMETRIC MODALITIES

Biometric

Acceptability

Application

Face

High

Access control, surveillance

Finger print

High

Law enforcement, access control

Palm print

Medium

Access control

Iris

Low

ID recognition

Signature

High

ID recognition

Table 1.1 Biometric characters acceptability and application

Biometric performance

Performance of a biometric system is measured with two types of errors associated namely false match FAR(accepting a false user) and falsemismatch FRR (rejecting an actual user) [12]. Biometric system should have small FAR and small FRR. A high FAR indicates system failure and low FRR indicates poor performance of the system accepting all the users. Thetradeoff between these two is EER. Biometric error estimation are also used to rate the accuracy of the system. ROC plots, DET plots, CMC plots, EER plots, tradeoffs are various parameters used in analyzing a biometric system.

http://urreward.com/wp-content/uploads/2012/10/im11.png

Fig 1.3 EER , ROC Plots

COST :Biometric system cost includes the hardware , software , maintainance. Sensor costs depends on the biometric trait used. Finger print sensors are inexpensive compared to hand vein sensors. With the rapid decrease in prices and better performance have made biometric technology a practical use in various applications like ATM, banking, commercial appications. For a high secured applications like airport safety, surviellance, criminal detection etc cost becomes secondary with respect to safety and security.

1.3.4 Biometric limitations:

Unimodal biometric system is subjected to various limitations resulting into errors in the system. Limitations are to be analyzed to improve the performance of the existing system. Limitations of biometric system can be listed as

Noisy sensor data: Data acquired may be noisy due to defective sensor, improper interaction of the user with the sensor (head pose variations, placement of the finger etc)

Intra class variation: A query image is rejected by the system due to its mismatch with its own template stored in the data base. This may be due to change in features over a period of time, poor image quality etc.

Inter class similarity: A query image may be wrongly accepted by the system due to its similarity with the features of other person stored in the data base.

Spoof attacks: An imposter may attempt to get an access into the system by spoof attack. Voice imitation, signatureforgery, facial mask, silicon finger prints, are such examplesfor spoof etc.

MULTIMODA BIOMETRIC SYSTEM

Limitations imposed by unimodal biometric systems can be overcome using multiple source of information of a person in establishing his identity. Biometric system using multiple modalities, multiple information are called Multimodal biometric system. Integrating information from different traits (likeface and fingerprint) or different samples of same trait( like multiple face images of person in different angles)results into more reliable, accurate biometric system. Information fusion of a multimodal biometrics can be consolidated in any one of the level stated [21].

Sensor level fusion: The biometric data is acquired with the help of a sensor. Information from multiple samples of same trait(samples of face in different pose, right and left hand finger prints, right and left iris of a same person) or multiple traits of same person(face and finger print, face and palm print of single person) are fused in sensor level fusion. The fused data is processed before feature extraction and matching.

Feature level fusion: The features of the data acquired from the multiple biometric characters are fused to get resultant feature set. The compatibility of the features of different biometric characters is a critical requirement as the fused data is to be expressed in to a single feature vector.

Matchingscore level: Input image features are compared with its stored template features. The degree of similarity is estimated with match score.For a multimodal biometric system match scores of different biometric characters are calculated and fused to get final match score.

Decision level: A threshold set by the system decides how close is the matching between the query and template. The system decides to accept or reject the user depending on the match score [22].

C:\Documents and Settings\seshikala\Desktop\kiran pictures modified\22.jpg

Fig 1.4 Multimodal Biometric characters

Any human physical or behavioral character can be used as biometric as long as it satisfies the condition of universality, distinctiveness, permanence, and collectability. Applications which require a more reliable authentications like law enforcement, forensics, criminal identification etc.,needsto be designed with high accuracy. Performance improvement can be achieved by adapting multiple character, multiple approaches in the design of biometric system [26],

LITERATURE REVIEW

In this section of literature survey the following topics are discussed

Biometrics, Biometric limitations, secured biometrics, biometric performance analysis, accurate biometric system, multiple modalities, multimodal biometrics, fusion levels, fusion strategies, fusion scenario, Palm as a biometric, palm feature, palm feature extractions methods, edge detection methods, Face as a biometric, face recognition system, face detection, face feature extraction, approaches, classifiers, multi resolution methods, feature level fusion, feature compatibility and approaches.

Biometrics is a Greek term defining bios as life and metric as a measurement. Biometrics is basically a pattern recognition system using human characters recognition. History says that astrologers studied palm prints in predicting the future and finger prints were used in olden days as a mark of authentication of document. Behavioral characters like signature are used from olden days to identify a person. With the growing technology and rapid growth in applications of various fields like communication, networking, banking, and highly secure applications like criminal detection, forensic, military security, we can see tremendous growth in the field of biometrics to fulfill the demands of get reliable, cost effective, user friendly system.

Anil K Jain [1] introduces the biometric recognition which works in verification mode or identification mode. Identity of a person is established by comparing the input data with the stored template data. Author gives the basic biometric system operation modes as sensor level which captures the input data, feature level which extracts the features of collected data, match score level which estimates the degree of matching between input data and stored data, and a decision level which decides to either accept or reject the input. Author presents an overview of errors that could occur in the biometrics as false matchand false mismatch. Biometric limitations like noisy sensor data, inter and intra user variations, spoof attacks, etc are presented. A brief review on multimodal biometrics was given. The paper gives an overview of the biometrics. The approaches to implement, performance of different modalities, measurements to be done are not discussed.

Anil Jain [10] presented biometrics as apromising frontier for the identification. He describes biometric can be knowledge based like pass words or token based like ID cards. A comparison of face, finger print, hand, iris based on universality, acceptability, permanence, uniqueness is presented as a case study and summarized stating biometrics will likely to be used in almost every transactions. Years after the authors observation indeed biometrics has found its place in every transaction of a modern world.

Publicationsrelated to biometrics are found in the literature from 1998 onwards and later within a decade a tremendous growth in research publications are seen.

James Wayman [4] describes taxonomy of uses, applications, system model in his work. Various issues like performance of a biometric system, types of identification like positive vs negative, open vs closed, attended vs non attended etc are discussed. This paper gives a scenario of various applications of a biometric system.

R. Brunelli [7] in his patent presented an integrated multisensory recognition system using acoustic and visual features for person identification. Integration of multipleinformation was a key issue in implementation of a reliable system. The work was carried out using acoustic and visual features of a person for identification. The speaker and face recognition system are decomposed into two and three single feature classifier respectively. The resulting five classifiers produce non homogenous scores which are combined using different approaches. The speaker recognition is based on vector quantization of the acoustic parameter space. Face recognition is based on comparison of facial features at the pixel level. Integration of information, using multiple classifiers is considered as learning task.

R.Michael Mc Cabe [16] presented the standards for identifying biometric accuracy in 2003. In his work carried under NIST (National institute of standards and technology, tests were conducted using image based biometrics under FRVT facial recognition vendors test, finger print verification. Standards were set and it is observed that not all subjects can easily finger printed and 2 to 3.5% have damaged friction ridges. This work gave a platform to indicate that a dual biometric system including more images may be needed to meet the existing system requirement. Biometric performance analysis was also presented in the work of [11] [15][17].

Marco Gamassi [19] presented a critical analysis of the measurement of accuracy and performance of a biometric system. He compared and criticized the current approaches stating that the performance under larger data bases must be tested. He also stated that second source of uncertainty which effects the overall accuracy should be considered. It is also observed that accuracy depends on how information of the biometrics features are used, but not on the complexity of the design. The author also presents typical accuracy indexes like symmetry, asymmetry,, matching scores, false match rate, false non match rate. EER is plotted at gives the performance of the system for false match vs false non match.

P.J.Phillips[12] presented that by continuously evaluating, it encourages the progress of improving the accuracy of a system. The performance limitations that are nearly impossible to work around are to be analyzed and working towards the multiple biometrics for performance improvement was stated in his work. The paper also presents the tools and techniques for biometric testing. Li Hong[23] in his work presented the architecture of integration of face and finger prints where a case study was performed using score based recognition. The integration systemretrieves first top five matches for facerecognition. Finger print verification was applied to each of the resulting top five matches and final decision is made by decision fusion scheme. Experiments were calculated on a small data base with 64 individuals and it was presented that an integration scenario gave a better results compared to finger or face taken separately.

Biometric recognition system works in four levels namely sensor level, feature level, match score level, and decision level.

U. Dieckman [33] proposed person authentication system SESAM in which three different modalities sound of voice, lip motion, and fixed image of face are used as cues in identifying a person. The system uses three different sensors. Optical cue are recorded with grey level CCD camera(768x572 pixels). A second video recording is triggered by acoustic signal so that voice and lip movements are extracted. Each cue is recorded separately and preprocessed independently. The data is trained by three separate classifiers. The work presented includes multiple classifiers and multiple sensors. The drawback of the implementation was the large storage space of learning patterns.

A.Ross[22] provides the information of fusion scenario of biometric system. A biometric system limitations such as noisy sensor data, spoof attacks, inter class similarity, intra- class variations can be reduced using fusion of information. Ross presents the fusion in the context of biometric as single biometric multiple samples, multi biometric samples, multi classifiers, multiple approaches. Ross presented fusion in match score level increases the performance of the biometric system. Finger print, face data is obtained from 50 users with 5 samples each to generate 500(50x10) genuine scores and is compared with 12250 imposters. Sum rule is used to find the weighted average of the final score. Decision tree, linear discriminant methods were also used to compare it with sum rule. User specific applications with widelyacceptable biometric character selection was addressed.

Ross and Jain [21] gave an over view of multimodal biometrics presenting levels of fusion, fusion scenario using multiple sensors, multiple classifiers, multiple approaches. Integration strategies for feature level fusion,match score level fusion, rank level fusion and decision level fusion are presented. Multimodal biometrics addresses several limitations of unimodal system. The performance of unimodal can be improved with integration of multiple source of information. Researchers presented various papers related to application, Implementation of multimodal biometric system in their work [30] [31]

Some of the fusion level implementations contributed for biometric systems are

D.Kisku [60] proposed feature level fusion of face and ear. S. Prabhakar [ 29 ] decision level fusion in finger print verification, Kittler[35 ] presented combining classifiers for decision level fusion, P.Verlinde[32] on comparing decision fusion using KNN classifier, decision trees and logistic regression, E. Bigun[ ] with multimodal system using Bayesian statistics and Y. Zue [ ] presented to increase the decision reliability.

Researches worked on various levels of fusion and literature finds match score level fusion was the most studied, presented with different approaches. Feature level fusion is an understudied problem with limited publications.

Feature level fusion involves the fusion of features extracted from the multiple biometrics. The features are to be represented in a feature vector space and the salient feature are extracted and represented as resultant feature of the modality. Fusion at feature level involves consolidation of features of multiple biometrics and representing in a common domain. Dimensionality of the fused vectors has to be reduced to suit the representation. Selecting the biometric traits to suit the requirement is a challenging task. As a result feature level fusion is still at an ongoing research issue for multimodal biometrics.

Selection of biometric characters for feature level fusion is based on the condition that the fusion should satisfy the condition of resolution, localization, sampling directionality and anisotropy.Feature level fusion involves, identifying the region of interest, preprocessing, extracting the features of multiple biometrics, concatenating the multiple features in to single feature set, identifying the vector space to represent the larger feature set, dimension reduction of larger feature sets and finally storing the resultant feature as a template in the data base. Stored templates are unique.

Selecting biometric characters for feature level fusion involves two constraints

Compatibility of feature set

Dimensionality

Compatibility of a feature set is based on selection of biometric character to be fused. In the present thesis work, palm and face are chosen as biometric characters whose features are compatible and can be used for biometric fusion.

Palm print based biometrics is one of the low cost biometricsystems used. Palm print has rich information like principal lines, textures, ridges, minutiae and is a user friendly biometric character, which are easy to obtain from a low cost sensor.It is used in both unimodal and multimodal biometric system.

Zhang et al and Han were first research team developed CCD based palm print scanner for Hong Kong polytechnic university. CCD based scanners align the palm print and capture high quality image in a controlled environment. Palm image in a more hygienic environment wascaptured bycontactless sensors and are used for person identification [ ]. Digital cameras, video cameras were also used as palm image sensors but were found to have a low quality image.

X. Wu [41] presents a novel approach for palm line extraction for on line palm prints. It was presented in two stages, coarse level extraction and fine level extraction. Morphological operations are used to extract palm lines in different directions in first stage and for each extracted line a recursive process is used to extract further features. Palm lines form one of the most important features for palm print recognition. Palm lines are bold, thick compared to other regions of pam surface.

A.Kumar and D..Zhang[42] investigates the performance of bimodal biometric system using fusion of shape and texture. To improve the performance of the hand shape user authentication, Discrete courier transform coefficients for feature extraction of palm is used. Score level fusion was used to fuse the hand shape and palm features scores. Product rule is used for calculating the match scores of palm and hand shape. The results are compared with sum and max rules. Application of multiple approaches were used and compared against their performance.

Application of various approaches used by researchers enhanced the performance of palm print recognition system. Approaches like line based, sub space based, statistic based were used to extract palm features. [37][38][39][40][41].

Researchers proposed combination of palm print for better recognition than a single palm print. A. Kumar and D.Zhang [45] proposed a new approach where simultaneous use of three approaches were used for better results. Gabor, line and appearance based palm print representations with score and decision level fusion was used. Integrating shape and texture was also proposed by the author. Using a common sensor palm and hand geometry features are acquired from the same image and at the same time. Features are examined for individual and combined performance.

Template security and data base attacks was addressed by various researchers

A.Kong [48] addresses the three relevant security issues like data attack, replay attack, template reuse attacks. The paper proposes random orientation filter bank as a feature extraction to generate noise like feature codes called competitive codes. Secret messages are hidden in template to prevent data base attacks. Author proposed palm print recognition system based on competitive code for template security. The palm print images collected from the scanner are aligned to suit to the coordinate system and preprocessed for further feature extraction and template storage. Competitive codes are generated and hidden in the templates to prevent replay and data base attacks. Authors in their work also addressed the secured measures for template storing[87], [28].

Multi resolution methods and its application in pattern recognition and image processing gavea promising results. Multi resolution analysis tools, more popularly wavelets have been found quite useful in analyzing the information content of images.

Zhang [ ] used complex wavelets to decompose palm print images and proposed complex wavelet structural similarity(CW-SSIM) index for measuring the local similarity of two images. The overall similarity of two palm prints is estimated as the average of all local modified CW-SSIM .

Zhou [50] used wavelet to decompose palm prints and used support vector machine as a classifier. The input of the SVM is low subband images. This approach lacks the information in muddle frequency spectrum.

Chen et al [52] proposed two dimensional dual tree complex transform on the preprocessed palm print image to resolve the weakness of traditional wavelet transform like shift invariant.Symbolic aggregate approximation is used to represent the features and minimum distance to compare the two feature vector.

The first work of feature level fusion of palm print was proposed by Han [93 ]. Author extracted seven specified line profiles from a preprocessed palm prints and three fingers and used wavelets to compute low frequency information. A new feature vector is formed using the information and the dimensionality is reduced using PCA. A learning vector quantization and optimal positive Boolean function are used to make final decision.

Statistical approaches, transform approaches were used to extract palm print features. Using either Fouriertransform or wavelet transform the directional features were not represented accurately. Contour transforms, complex wavelet transform were used to represent the directional property but has the limitation of larger dimensionality.

Curvelet transform overcomes the various limitations set by other transform methods and its application to image processing is being introduced only recently. The present thesis work is to use an approach which can be used for face and palm printfeature extraction and later use for feature level fusion.

Facerecognitionis most studied topic in the computer vision literature past over 30 years. Face recognition is still an active research due to its application in various human activity involved applications like face tracking, face surveillance, criminal detection, etc. Face biometrics is most preferred biometric character as it is nonintrusive and can also be tan en without user knowledge for identification and verification. Face recognition system stages involves face detection, preprocessing, feature extraction and face recognition.

The main goal of face detection is to find presence or absence of a face from a picture, video or in a surveillance. Automatic face recognition is to identify a given face image from the stored data base. The limitations that are encountered are variations of given image with respect to the stored data base image of a genuine user. Variations can be due to aging, pose, illumination, clutter in the background, occlusions due to accessories etc. Various approaches were proposed by researchers to overcome the limitations.

Viola and Jones [113] has made face detection practically feasible to real world applications. The viola jones face detector contains three main ideas that make it possible to build a successful face detector that can run in real time.i.e integral image, classifier learning with Ada Boost and a cascade structure.

Face detection methods in literature survey shows that they are classified into two groups, knowledgebased and image based. Knowledge based methods uses facial features (like shape of face, eyes, eyebrows) template matching, skin color etc. Many detection algorithms are based on facial features.

Zhi-fang [ ] detected face and facial features by extraction of skin like region with YCbCr color space and edges are detected in skin like region. Eyes, mouth are found with geometrical information. Researchers proposed various approaches using skin color to detect the face features. [ ], [ ], [ ].

Human face detection using template matching was proposed by various researchers in the literature study. Chen et al [ ] used half face template instead of full face template to reduce the computation time. The template based methods compute the correlation between a face and one or more model templates to find the face identity. Brunelli and Poggio suggested that the optimal strategy for face recognition is holistic and corresponds to template matching. The author compared geometric feature based technique with template matching based system and achieved a very high accuracy. Principal component analysis (PCA), Linear discriminant analysis(LDA), Neural networks , kernel methods were used to construct suitable set of face templates [ ],[ ],[ ],[ ],[ ].

Geometric feature based methods analyze explicit local facial features and their geometric relationships.Yuille proposed a shape model for face image and later the extension of work can be seen by Cortes et al [110]. Wiskott [95] was first to develop elastic bunch graph model and application of PCA into local features was developed by Penev et al [96]. The work proposed by these researchers gave a basic for many applications and development in the face recognition and image processing.

Phillips[84] presented FERET evaluation tests in his work and the main objective of the test was to assess the state of art, identify the future area of research and to measure the algorithm performance. Later FRVT 2000, FRVT 2002, FAT 2004, were conducted to address age, pose, illumination variations and standards were set for future research.

Extraction of facial features is an integral process involved in face detection, face modeling, face recognition, animation and facial expression determination. Featured-based approaches first process the input image to identify and extract and measure distinctive facial features such as the eyes, mouth, nose, etc. and then compute the geometric relationships among those facial points. Standard statistical pattern recognition techniques are then employed to match faces using these measurements.

*Kanade[ ] whose work was an initial state of recognition, extracted 16 facial parameters which were ration of distance, angle, area and used Euclidean distance measure for matching and achieved 75% of accuracy with a data base of 20 people and with two images per person. Brunelli and Poggio [ ] extended the Kanades work with 35 geometric features with a90% recognition rate.

Li and Yin [ ] introduced a system in which a face image is first decomposed with a wavelet transform to three levels. The Fisher faces method [ ] is then applied to each of the three low-frequency sub-images. Then, the individual classifiers are fused using the RBF neural network.

Melin et al. [ ] divided the face into three regions (the eyes, the mouth, and the nose) and assigned each region to a module of the neural network. A fuzzy Sugeno integral was then used to combine the outputs of the three modules to make the final face recognition decision. They tested it on a small database of 20 people and reported that the modular network yielded better results than a monolithic one.

Multiple classifiers were used and their information was fused to give more accurate performance. Lu et al. [ ] fused the results of PCA, ICA and LDA using the sum rule. Bayes combination rule to integrate the weighted outcomes of three classifiers based on frontal and profileviews of faces.

Wan et al. [ ] used SVM and HMM hybrid model, Kwak and Pedrycz [ ] divided the

face into three regions, applied the fisherfaces method to the regions as well as to the whole face and then integrated the classification results using the fuzzy integral [ ], Haddadnia et. al. [] used PCA, the Pseudo Zernike Moment Invariant (PZMI) [] and theZernike Moment Invariant (ZMI) to extract feature vectors in parallel, which were then classified simultaneously by separate RBF neural networks and the outputs of these networks were then combined by a majority rule todetermine the final identity of the individual in the input.

Principal component analysis used to represent Eigen faces with a lower dimension suffers from computational load and correlation of facial features. Image representation should satisfy the condition of multi resolution, localization, critical sampling, directionality and anisotropy. Multi resolution analysis tools more popularly wavelets have been found quite useful in analyzing the information content of images. Wavelets have been successfully used in image processing and its ability to capture localized time-frequency information of image motivates its use for feature extraction. Wavelet-based methods focus on the sub bands that contain the most relevant information for better representation of the data. The wavelet transform can be interpreted as a multi scale differentiator or edge detector that represents the singularity of an image at multiple scales and three different orientations as horizontal, vertical, and diagonal.

Zhang et al.[ ] proposed a modular face recognition scheme by combining the techniques of wavelet subband representations and kernel associative memories. By the wavelet transform, face images are decomposed and the computational complexity is substantially reduced by choosing a lower spatial-frequency sub band image. Then a kernel associative memory (KAM) model is built up for each subject, with the corresponding prototypical images without any counter examples involved. Multiclass face recognition is thus obtained by simply holding these associative memories. When a probe face is presented, the KAM model gives the likelihood that the probe is from the corresponding class by calculating the reconstruction errors or matching scores. Various researchers implemented wavelets as feature extraction method. [ ][ ][ ][ ] [ ]

Wavelet transform has been found quite useful for analyzing the information content of an image, has the limitations of identifying only the point singularities in an image. The edges and curves are not well defined by the wavelet transform. Application of ridge transform ,contour transform was able to give information of edges and curves but suffers with the dimensionality problem of representing the coefficients. Curvelets which is a multi-scale, multi-resolution transform overcomes the limitations of wavelets and provides optimal representation of objects with curve singularities. Curvelets requires relatively small number of coefficients to represents a line or a curve in a given image.

Literature study shows that application of curvelet transform was a recent introduction to biometric face recognition. Its application to multimodal fusion is still under research.

Selecting a biometric character for multimodal fusion depends on various criteria. The level of fusion selected must be compatible for integration of information from the different biometric traits. Multimodal fusion is categorized as sensor level, feature level, match score level and decision level. Feature level fusion is an understudied problem by researchers due to the following limitations

Compatibility of feature sets of different biometric characters

Large dimension of feature set of fused data

The present research work was carried out by addressing the initial problem of selecting a biometric character, which are compatible in their feature representation.

Palm print with its rich information and face as a universal biometric identifier are used for fusion. Literature survey shows that fusion of palm and face under feature level is a under studied problem. Fusion under match score level was used extensively by researchers with different approaches.

D.kisku[61] proposed feature level fusion of palm and face by isomorphic graph based K -medoids partitioning. Partitioning the palm and face images with K invariant SIFT points, forming number of clusters. On each cluster an isomorphic graph is drawn. The most probable pair of graph is searched using iterative relaxation algorithm from all possible graph pair of face and palm print images. Graphs are fused into augmented groups using concatenation rule. The accuracy of the system is very high when compared with other level of fusion.

D.Kisku [ ] proposed feature level fusion of face and finger prints by extracting feature independently from two modalities and making two point sets compatible for concatenation. Feature reduction techniques are employed prior and after feature set fusion. Comparative experiments were also conducted with match score level fusion. Face features were extracted using SIFT and fingerprint verification on minutiae matching and K means clustering was employed in feature reduction. The proposed technique was also tested on different data bases. FAR, FRR and accuracy obtained was compared with unimodal, multimodal and after using reduction techniques.

Yucheng wang [ ] proposed feature level fusion of palm and face based on kernel fisher discriminant analysis (KFDA) and Linear discriminant analysis( LDA). The results showed that multimodal fusion had a higher accuracy when compared with unimodal.

Anil Jain and Ling Hong [71] used integration of face finger print and speaker verification in making personal identification. The fusion level used was decision level. Face recognition was implemented using Eigen faces. The speaker recognition was based on text dependent speaker recognition. The minutiae of ridges were used for extraction and matching. The aim was to show that identity established in multimodal system is more reliable than using a single modality alone. The system was tested only on a small data set and its performance for a larger data set was not tested.

Raghavendra R [113] Introduced Particle swam optimization for feature level fusion of face and palm print. Face and palm print features are represented using Log Gabor features which are then concatenated to form a fused feature vector. PSO is used to compute weights for each of the features qualitatively. Binary PSO is used to select most discriminant features from fused features. The feature level was compared with match score level. The proposed scheme out- performed the state of art scheme.

By studying and analyzing the literature one can summarize the main requirement of a biometric system as

Selection of a biometric trait depends on the application.

Biometric system performance is not optimum. Biometric system has limitations.

Using multimodal biometrics the performance accuracy can be improved.

Multiple biometric characters selected must be compatible in fusion level used.

Performance and accuracy of a biometric system can be improved using multiple approaches.

At the end of literature it was found that

Application of multi scale wavelet edge detection methods was not used for palm print recognition.

Application of curvelet transform for feature level fusionofpalm and face was not found in the literature.

Problem Statement:

The problem of the present dissertation can be stated as

To study the basic biometric system, modalities, limitations, multimodal biometrics and fusion levels.

To study the requirements of a biometric system by analyzing its parameters, limitations and errorsassociated with each biometric character used.

To develop and implement palm print recognition system.

To develop and implement face recognition system.

To develop and implement feature level fusion of palm and face biometrics.

Objectives:

In this research work an attempt has been made to study the basic biometric system processing. The biometric levels, modalities used, limitations, multimodal biometrics and fusion levels are investigated. Efforts are made to analyze the performance of biometric system using various modalities. Biometric characters palm and face are chosen to develop and implement a recognition system using MATLAB software. Palm print recognition system using multi scale wavelets and multi resolution Curvelet transform is developed for which a considerable improvement in efficiency is seen compared to other approaches found in the literature. Application of curvelet for face feature extraction was implemented and found to have remarkable increase in accuracy. Feature level fusion of multimodal biometric system using curvelet transform was developed which are not available in the literature.The data bases used are POLY U, IIT K , CASIA for palm and face respectively. The performance is analyzed with ROC plots and recognition rates.

Organization of the Thesis:

The thesis is organized in to 7 chapters

The chapter-2 gives Biometric performance analysis. The biometric parameters, system scenario, system requirement and performance of various modalities are presented with results and conclusions.

The chapter-3 Explains Palm print characters, Palm print recognition system, processing levels, data bases used, analysis and discussion.

The chapter-4Explains developmentand implementation of palm print recognition system using edge detection techniques like sobel, canny , multi scale edge detection with results and discussion.

The chapter-5explains development and implementation of Face recognition system using multi resolution curvelet transform with results and discussion

The chapter-6gives the multimodal biometric system, fusion levels, fusion scenarios, feature level fusion analysis, and discussion.

The chapter 7explains the implementation of feature level fusion of palm and face biometric character using curvelet transform with results and conclusion.

The chapter-8 gives Results and conclusion, Scope for future work.