This PDF file contains the front matter associated with SPIE Proceedings Volume 6539, including the Title Page, Copyright information, Table of Contents, Introduction, and the Conference Committee listing.

Significant advances continue to be made in biometric technology. However, the global war on terrorism and our increasingly electronic society have created the societal need for large-scale, interoperable biometric capabilities that challenge the capabilities of current off-the-shelf technology. At the same time, there are concerns that large-scale implementation of biometrics will infringe our civil liberties and offer increased opportunities for identity theft. This paper looks beyond the basic science and engineering of biometric sensors and fundamental matching algorithms and offers approaches for achieving greater performance and acceptability of applications enabled with currently available biometric technologies. The discussion focuses on three primary biometric system aspects: performance and scalability, interoperability, and cost benefit. Significant improvements in system performance and scalability can be achieved through careful consideration of the following elements: biometric data quality, human factors, operational environment, workflow, multibiometric fusion, and integrated performance modeling. Application interoperability hinges upon some of the factors noted above as well as adherence to interface, data, and performance standards. However, there are times when the price of conforming to such standards can be decreased local system performance. The development of biometric performance-based cost benefit models can help determine realistic requirements and acceptable designs.

We address the problem of score level fusion of intramodal and multimodal experts in the context of biometric identity verification. We investigate the merits of confidence based weighting of component experts. In contrast to the conventional approach where confidence values are derived from scores, we use instead raw measures of biometric data quality to control the influence of each expert on the final fused score. We show that quality based fusion gives better performance than quality free fusion. The use of quality weighted scores as features in the definition of the fusion functions leads to further improvements. We demonstrate that the achievable performance gain is also affected by the choice of fusion architecture. The evaluation of the proposed methodology involves 6 face and one speech verification experts. It is carried out on the XM2VTS data base.

Fingerprints are one of the most commonly used and relied-upon biometric technology. But often the captured fingerprint image is far from ideal due to imperfect acquisition techniques that can be slow and cumbersome to use without providing complete fingerprint information. Most of the diffculties arise due to the contact of the fingerprint surface with the sensor platen. To overcome these diffculties we have been developing a noncontact scanning system for acquiring a 3-D scan of a finger with suffciently high resolution which is then converted into a 2-D rolled equivalent image. In this paper, we describe certain quantitative measures evaluating scanner performance. Specifically, we use some image software components developed by the National Institute of Standards and Technology, to derive our performance metrics. Out of the eleven identified metrics, three were found to be most suitable for evaluating scanner performance. A comparison is also made between 2D fingerprint images obtained by the traditional means and the 2D images obtained after unrolling the 3D scans and the quality of the acquired scans is quantified using the metrics.

A comparative study on multiple participants was undertaken to quantify the ability of a multispectral imaging fingerprint sensor to perform reliable biometric matching in the presence of extreme sampling conditions. These extreme conditions included finger wetness, dirt, chalk, acetone, bright ambient light, and low contact pressure during image acquisition. The comparative study included three commercially available total internal reflectance sensors, run in parallel with the multispectral imaging sensor and under identical sampling conditions. Performance assessments showed that the multispectral imaging sensor was able to provide fingerprint images that produced good biometric performance even under conditions in which the performance of the total internal reflectance sensors was severely degraded. Additional analysis showed that the performance advantage of the multispectral images taken under these conditions was maintained even when matched against enrollment images collected on total internal reflectance sensors.

In this paper we propose a novel 3D face recognition system. Furthermore we propose and discuss the development of a 3D reconstruction system designed specifically for the purpose of face recognition. The reconstruction subsystem utilises a capture rig comprising of six cameras to obtain two independent stereo pairs of the subject face during a structured light projection with the remaining two cameras obtaining texture data under normal lighting conditions. Whilst the most common approaches to 3D reconstruction use least square comparison of image intensity values, our system achieves dense point matching using Gabor Wavelets as the primary correspondence measure. The matching process is aided by Voronoi segmentation of the input images using strong confidence correlations as Voronoi seeds. Additional matches are then propagated outwards from the initial seed matches to produce a dense point cloud and surface model. Within the recognition subsystem models are first registered to a generic head model, and then an ICP variant is applied between the recognition subject and each model in the comparison database, using the average point-to-plane error as the recognition metric. Our system takes full advantage of the additional information obtained from the shape and structure of the face, thus combating some of the inherent weaknesses of traditional 2D methods such as pose and illumination variations. This novel reconstruction / recognition process achieves 98.2% accuracy on databases containing in excess of 175 meshes.

An automatic classification scheme of on-line handwritten signatures is presented. A Multilayer Perceptron (MLP) with a hidden layer is used as classifier, and two different signature classes are considered, namely: legible and non-legible name. Signatures are represented considering different feature subsets obtained from global information. Mahalanobis distance is used to rank the parameters and feature selection is then applied based on the top ranked features. Experimental results are given on the MCYT signature database comprising 330 signers. It is shown experimentally that automatic on-line signature classification based on the name legibility is feasible.

Hand-based authentication is a key biometric technology with a wide range of potential applications both in industry and government. Traditionally, hand-based authentication is performed by extracting information from the whole hand. To account for hand and finger motion, guidance pegs are employed to fix the position and orientation of the hand. In this paper, we consider a component-based approach to hand-based verification. Our objective is to investigate the discrimination power of different parts of the hand in order to develop a simpler, faster, and possibly more accurate and robust verification system. Specifically, we propose a new approach which decomposes the hand in different regions, corresponding to the fingers and the back of the palm, and performs verification using information from certain parts of the hand only. Our approach operates on 2D images acquired by placing the hand on a flat lighting table. Using a part-based representation of the hand allows the system to compensate for hand and finger motion without using any guidance pegs. To decompose the hand in different regions, we use a robust methodology based on morphological operators which does not require detecting any landmark points on the hand. To capture the geometry of the back of the palm and the fingers in suffcient detail, we employ high-order Zernike moments which are computed using an effcient methodology. The proposed approach has been evaluated on a database of 100 subjects with 10 images per subject, illustrating promising performance. Comparisons with related approaches using the whole hand for verification illustrate the superiority of the proposed approach. Moreover, qualitative comparisons with state-of-the-art approaches indicate that the proposed approach has comparable or better performance.

The paper proposes an efficient indexing scheme for binary feature template using B+ tree. In this scheme the input image is decomposed into approximation, vertical, horizontal and diagonal coefficients using the discrete wavelet transform. The binarized approximation coefficient at second level is divided into four quadrants of equal size and Hamming distance (HD) for each quadrant with respect to sample template of all ones is measured. This HD value of each quadrant is used to generate upper and lower range values which are inserted into B+ tree. The nodes of tree at first level contain the lower and upper range values generated from HD of first quadrant. Similarly, lower and upper range values for the three quadrants are stored in the second, third and fourth level respectively. Finally leaf node contains the set of identifiers. At the time of identification, the test image is used to generate HD for four quadrants. Then the B+ tree is traversed based on the value of HD at every node and terminates to leaf nodes with set of identifiers. The feature vector for each identifier is retrieved from the particular bin of secondary memory and matched with test feature template to get top matches. The proposed scheme is implemented on ear biometric database collected at IIT Kanpur. The system is giving an overall accuracy of 95.8% at penetration rate of 34%.

This paper investigates a new approach for human ear identification using holistic grey-level information. We employ Log-Gabor wavelets to extract the phase information, i.e. ear-codes, from the 1D gray-level signals. Thus each ear is represented by a unique ear code or (phase template). The query ear images are compared with those in the database using Hamming distance. The minimum Hamming distance obtained from the rotation of ear template is used to authenticate the user. Our experiments on two different public ear databases achieve promising results and suggest its utility in ear-based authentication. This paper also illustrates that the phase information extracted from ear images can achieve significant performance improvement as compared to appearance-based approach employed in the literature.

To compensate for the different orientations of two fingerprint images, matching systems use a reference point and a set of transformation parameters. Fingerprint minutiae are compared on their positions relative to the reference points, using a set of thresholds for the various matching features. However a pair of minutiae might have similar values for some of the features compensated by dissimilar values for others; this tradeoff cannot be modeled by arbitrary thresholds, and might lead to a number of false matches. Instead given a list of potential correspondences of minutiae points, we could use a static classifier, such as a support vector machine (SVM) to eliminate some of the false matches. A 2-class model is built using sets of minutiae correspondences from fingerprint pairs known to belong to the same and different users. For a test pair of fingerprints, a similar set of minutiae correspondences is extracted and given to the recognizer, using only those classified as genuine matches to calculate the similarity score, and thus, the matching result. We have built recognizers using different combinations of fingerprint features and have tested them against the FVC 2002 database. Using this recognizer reduces the number of false minutiae matches by 19%, while only 5% of the minutiae pairs corresponding to fingerprints of the same user are rejected. We study the effect of such a reduction on the final error rate, using different scoring schemes.

Given a query fingerprint, the goal of indexing is to identify and retrieve a set of candidate fingerprints from a large database in order to determine a possible match. This significantly improves the response time of fingerprint recognition systems operating in the identification mode. In this work, we extend the indexing framework based on minutiae triplets by utilizing ridge curve parameters in conjunction with minutiae information to enhance indexing performance. Further, we demonstrate that the proposed technique facilitates the indexing of fingerprint images acquired using different sensors. Experiments on the publicly available FVC database confirm the utility of the proposed approach in indexing fingerprints.

Matching of partial fingerprints has important applications in both biometrics and forensics. It is well-known that the accuracy of minutiae-based matching algorithms dramatically decrease as the number of available minutiae decreases. When singular structures such as core and delta are unavailable, general ridges can be utilized. Some existing highly accurate minutiae matchers do use local ridge similarity for fingerprint alignment. However, ridges cover relatively larger regions, and therefore ridge similarity models are sensitive to non-linear deformation. An algorithm is proposed here to utilize ridges more effectively- by utilizing representative ridge points. These points are represented similar to minutiae and used together with minutiae in existing minutiae matchers with simple modification. Algorithm effectiveness is demonstrated using both full and partial fingerprints. The performance is compared against two minutiae-only matchers (Bozorth and k-minutiae). Effectiveness with full fingerprint matching is demonstrated using the four databases of FVC2002- where the error rate decreases by 0.2-0.7% using the best matching algorithm. The effectiveness is more significant in the case of partial fingerprint matching- which is demonstrated with sixty partial fingerprint databases generated from FVC2002 (with five levels of numbers of minutiae available). When only 15 minutiae are available the error rate decreases 5-7.5%. Thus the method, which involves selecting representative ridge points, minutiae matcher modification, and a group of minutiae matchers, demonstrates improved performance on full and especially partial fingerprint matching.

The increasing use of biometrics in different environments presents new challenges. Most importantly, biometric data are irreplaceable. Therefore, storing biometric templates, which is unique to individual user, entails significant security risks. In this paper, we propose a geometric transformation for securing the minutiae based fingerprint templates. The proposed scheme employs a robust one-way transformation that maps geometrical configuration of the minutiae points into a fixed-length code vector. This representation enables efficient alignment and reliable matching. Experiments are conducted by applying the proposed method on a synthetically generated minutiae point sets. Preliminary results show that the proposed scheme provides a simple and effective solution to the template security problem of the minutiae based fingerprint.

Reliable person recognition is important for secure access and commercial applications requiring human identification. Face recognition (FR) is an important technology being developed for human identification. Algorithms and systems for large population face recognition (LPFR) are of significant interest in applications such as watch lists and video surveillance. In this paper, we present correlation filter-based feature analysis methods that effectively exploit available generic training data to represent a large number of subjects and thus improve the performance for LPFR. We first introduce a general framework - class-dependence feature analysis (CFA), which uses correlation filters to provide a discriminant feature representation for LPFR. We then introduce two variants of the correlation filter-based CFA methods: 1) the kernel correlation filter CFA (KCFA) that generates nonlinear decision boundaries and significantly improves the recognition performance without greatly increasing the computational load, and 2) the binary coding CFA that uses binary coding to reduce the number of correlation filters and applies error control coding (ECC) to improve the recognition performance. These two variants offer ways to tradeoff between the computational complexity and the recognition accuracy of the CFA methods. We test our proposed algorithms on the face recognition grand challenge (FRGC) database and show that the correlation filter-based CFA approach improves the recognition rate and reduces the computational load over the conventional correlation filters.

Most existing face recognition algorithms require face images with a minimum resolution. Meanwhile, the rapidly emerging need for near-ground long range surveillance calls for a migration in face recognition from close-up distances to long distances and accordingly from low and constant resolution to high and adjustable resolution. With limited optical zoom capability restricted by the system hardware configuration, super-resolution (SR) provides a promising solution with no additional hardware requirements. In this paper, a brief review of existing SR algorithms is conducted and their capability of improving face recognition rates (FRR) for long range face images is studied. Algorithms applicable to real-time scenarios are implemented and their performances in terms of FRR are examined using the IRISLRHM face database [1]. Our experimental results show that SR followed by appropriate enhancement, such as wavelet based processing, is able to achieve comparable FRR when equivalent optical zoom is employed.

In this paper, we propose a novel method for performing robust super-resolution of face images by solving the practical problems of the traditional manifold analysis. Face super-resolution is to recover a high-resolution face image from a given low-resolution face image by modeling the face image space in view of multiple resolutions. In particular, face super-resolution is useful to enhance face images captured from surveillance footage. Face super-resolution should be preceded by analyzing the characteristics of the face image distribution. In literature, it has been shown that face images lie on a nonlinear manifold by various manifold learning algorithms, so if the manifold structure is taken into consideration for modeling the face image space, the results of face super-resolution can be improved. However, there are some practical problems which prevent the manifold analysis from being applied to super-resolution. Almost all of the manifold learning methods cannot generate mapping functions for new test images which are absent from a training set. Also, there exists another significant problem when applying the manifold analysis to super-resolution; superresolution seeks to recover a high-dimensional image from a low-dimensional one while manifold learning methods perform the exact opposite for dimensionality reduction. To break those limitations of applying the manifold analysis to super-resolution, we propose a novel face superresolution method using Locality Preserving Projections (LPP). LPP gives an advantage over other manifold learning methods in that it has well-defined linear projections which allow us to formulate well-defined mappings between highdimensional data and low-dimensional data. Moreover, we show that LPP coefficients of an unknown high-resolution image can be inferred from a given low-resolution image using a MAP estimator.

In this paper we demonstrate the subspace generalization power of the kernel correlation feature analysis (KCFA) method for extracting a low dimensional subspace that has the ability to represent new unseen datasets. Examining the portability of this algorithm across different datasets is an important practical aspect of real-world face recognition applications where the technology cannot be dataset-dependant. In most face recognition literature, algorithms are demonstrated on datasets by training on one portion of the dataset and testing on the remainder. Generally, the testing subjects' dataset partially or totally overlap the training subjects' dataset however with disjoint images captured from different sessions. Thus, some of the expected facial variations and the people's faces are modeled in the training set. In this paper we describe how we efficiently build a compact feature subspace using kernel correlation filter analysis on the generic training set of the FRGC dataset and use that basis for recognition on a different dataset. The KCFA feature subspace has a total dimension that corresponds to the number of training subjects; we chose to vary this number to include up to all of 222 available in the FRGC generic dataset. We test the built subspace produced by KCFA by projecting other well-known face datasets upon it. We show that this feature subspace has good representation and discrimination to unseen datasets and produces good verification and identification rates compared to other subspace and dimensionality reduction methods such as PCA (when trained on the same FRGC generic dataset). Its efficiency, lower dimensionality and discriminative power make it more practical and powerful than PCA as a robust lower dimensionality reduction method for modeling faces and facial variations.

Automatic face recognition (AFR) is a challenging task that is increasingly becoming the preferred biometric trait for identification and has the potential of becoming an essential tool in the fight against crime and terrorism. Closed-circuit television (CCTV) cameras have increasingly been used over the last few years for surveillance in public places such as airports, train stations and shopping centers. They are used to detect and prevent crime, shoplifting, public disorder and terrorism. The work of law-enforcing and intelligence agencies is becoming more reliant on the use of databases of biometric data for large section of the population. Face is one of the most natural biometric traits that can be used for identification and surveillance. However, variations in lighting conditions, facial expressions, face size and pose are a great obstacle to AFR. This paper is concerned with using waveletbased face recognition schemes in the presence of variations of expressions and illumination. In particular, we will investigate the use of a combination of wavelet frequency channels for a multi-stream face recognition using various wavelet subbands as different face signal streams. The proposed schemes extend our recently developed face veri.cation scheme for implementation on mobile devices. We shall present experimental results on the performance of our proposed schemes for a number of face databases including a new AV database recorded on a PDA. By analyzing the various experimental data, we shall demonstrate that the multi-stream approach is robust against variations in illumination and facial expressions than the previous single-stream approach.

This paper presents a fully automatic real-time face recognition system from video by using Active Appearance Models (AAM) for fitting and tracking facial fiducial landmarks and warping the non-frontal faces into a frontal pose. By implementing a face detector for locating suitable initialization step of the AAM shape searching and fitting process, new facial images are interpreted and tracked accurately in real time (15fps). Using an Active Appearance Model (AAM) for normalizing facial images under different poses and expressions is crucial to providing improved face recognition performance as most systems degrade matching performance with even smallest pose variation. Furthermore the AAM is a more robust feature registration tracking approach as most systems detect and locate the eyes while AAMs detect and track multiple fiducial points on the face holistically. We show examples of AAM fitting and tracking and pose normalization including an illumination pre-processing step to remove specular and cast shadow illumination artifacts on the face. We show example pose normalization images as well as example matching scores showing the improved performance of this pose correction method.

As biometric recognition systems are widely applied in various application areas, security and privacy risks have recently attracted the attention of the biometric community. Template protection techniques prevent stored reference data from revealing private biometric information and enhance the security of biometrics systems against attacks such as identity theft and cross matching. This paper concentrates on a template protection algorithm that merges methods from cryptography, error correction coding and biometrics. The key component of the algorithm is to convert biometric templates into binary vectors. It is shown that the binary vectors should be robust, uniformly distributed, statistically independent and collision-free so that authentication performance can be optimized and information leakage can be avoided. Depending on statistical character of the biometric template, different approaches for transforming biometric templates into compact binary vectors are presented. The proposed methods are integrated into a 3D face recognition system and tested on the 3D facial images of the FRGC database. It is shown that the resulting binary vectors provide an authentication performance that is similar to the original 3D face templates. A high security level is achieved with reasonable false acceptance and false rejection rates of the system, based on an efficient statistical analysis. The algorithm estimates the statistical character of biometric templates from a number of biometric samples in the enrollment database. For the FRGC 3D face database, the small distinction of robustness and discriminative power between the classification results under the assumption of uniquely distributed templates and the ones under the assumption of Gaussian distributed templates is shown in our tests.

We define the biometric performance invariance under strictly monotonic functions on match scores as normalization symmetry. We use this symmetry to clarify the essential difference between the standard score-level fusion approaches of sum rule and Neyman-Pearson. We then express Neyman-Pearson fusion assuming match scores defined using false acceptance rates on a logarithmic scale. We show that by stating Neyman-Pearson in this form, it reduces to sum rule fusion for ROC curves with logarithmic slope. We also introduce a one parameter model of biometric performance and use it to express Neyman-Pearson fusion as a weighted sum rule.

A nonparametric inferential statistical data analysis is presented. The utility of this method is demonstrated through analyzing results from minutiae exchange with two-finger fusion. The analysis focused on high-accuracy vendors and two modes of matching standard fingerprint templates: 1) Native Matching - where the same vendor generates the templates and the matcher, and 2) Scenario 1 Interoperability - where vendor A's enrollment template is matched to vendor B's authentication template using vendor B's matcher. The purpose of this analysis is to make inferences about the underlying population from sample data, which provide insights at an aggregate level. This is very different from the data analysis presented in the MINEX04 report in which vendors are individually ranked and compared. Using the nonparametric bootstrap bias-corrected and accelerated (BCa) method, 95 % confidence intervals are computed for each mean error rate. Nonparametric significance tests are then applied to further determine if the difference between two underlying populations is real or by chance with a certain probability. Results from this method show that at a greater-than-95 % confidence level there is a significant degradation in accuracy of Scenario 1 Interoperability with respect to Native Matching. The difference of error rates can reach on average a two-fold increase in False Non-Match Rate. Additionally, it is proved why two-finger fusion using the sum rule is more accurate than single-finger matching under the same conditions. Results of a simulation are also presented to show the significance of the confidence intervals derived from the small size of samples, such as six error rates in some of our cases.

Facial recognition is fast becoming one of the more popular and effective modalities of biometrics when used in controlled environments. Controlled environments are those in which factors such as facial expression, pose, camera position, and in particular illumination effects are controlled to some degree with respect to better performance. Regulation or normalization of such factors has effects on all facial recognition algorithms, and the factor of illumination effects is one of significant importance. In this paper we describe a method to address illumination effects in the biometric modality of face recognition using Empirical Mode Decomposition (EMD) to identify illumination modes that compose the image. Following identification of intrinsic mode functions that correspond to the dominant illumination factors, we reconstruct the facial image minus these negative factors to synthesize a more neutral facial image. We then perform recognition and verification experiments using different algorithms such as Principal Component Analysis (PCA), Fisher Linear Discriminant Analysis (FLDA), and Correlation Filters (CF's) to demonstrate the fundamental effectiveness of EMD as an illumination compensation method. Results are reported on the Carnegie Mellon University Pose-Illumination-Expression (CMU PIE) Database and the Yale Face Database B.

Quality of a biometric system is directly related to the performance of the dissimilarity measure function. Frequently a generalized dissimilarity measure function such as Mahalanobis distance is applied to the task of matching biometric feature vectors. However, often accuracy of a biometric system can be greatly improved by introducing a customized matching algorithm optimized for a particular biometric. In this paper we investigate two tailored similarity measure functions for behavioral biometric systems based on the expert knowledge of the data in the domain. We compare performance of proposed matching algorithms to that of other well known similarity distance functions and demonstrate superiority of one of the new algorithms with respect to the chosen domain.

Gaze estimation is an important component of computer vision systems that monitor human activity for surveillance, human-computer interaction, and various other applications including iris recognition. Gaze estimation methods are particularly valuable when they are non-intrusive, do not require calibration, and generalize well across users. This paper presents a novel eye model that is employed for efficiently performing uncalibrated eye gaze estimation. The proposed eye model was constructed from a geometric simplification of the eye and anthropometric data about eye feature sizes in order to circumvent the requirement of calibration procedures for each individual user. The positions of the two eye corners and the midpupil, the distance between the two eye corners, and the radius of the eye sphere are required for gaze angle calculation. The locations of the eye corners and midpupil are estimated via processing following eye detection, and the remaining parameters are obtained from anthropometric data. This eye model is easily extended to estimating eye gaze under variable head pose. The eye model was tested on still images of subjects at frontal pose (0^o) and side pose (34^o). An upper bound of the model's performance was obtained by manually selecting the eye feature locations. The resulting average absolute error was 2.98^o for frontal pose and 2.87^o for side pose. The error was consistent across subjects, which indicates that good generalization was obtained. This level of performance compares well with other gaze estimation systems that utilize a calibration procedure to measure eye features.