Harin Sellahewa

Image-Quality-Based Adaptive Face Recognition

The accuracy of automated face recognition systems is greatly affected by intraclass variations between enrollment and identification stages. In particular, changes in lighting conditions is a major contributor to these variations. Common approaches to address the effects of varying lighting conditions include preprocessing face images to normalize intraclass variations and the use of illumination invariant face descriptors. Histogram equalization is a widely used technique in face recognition to normalize variations in illumination. However, normalizing well-lit face images could lead to a decrease in recognition accuracy. The multiresolution property of wavelet transforms is used in face recognition to extract facial feature descriptors at different scales and frequencies. The high-frequency wavelet subbands have shown to provide illumination-invariant face descriptors. However, the approximation wavelet subbands have shown to be a better feature representation for well-lit face images. Fusion of match scores from low- and high-frequency-based face representations have shown to improve recognition accuracy under varying lighting conditions. However, the selection of fusion parameters for different lighting conditions remains unsolved. Motivated by these observations, this paper presents adaptive approaches to face recognition to overcome the adverse effects of varying lighting conditions. Image quality, which is measured in terms of luminance distortion in comparison to a known reference image, will be used as the base for adapting the application of global and region illumination normalization procedures. Image quality is also used to adaptively select fusion parameters for wavelet-based multistream face recognition.

A lightweight approach for biometric template protection

Privacy and security are vital concerns for practical biometric systems. The concept of cancelable or revocable biometrics has been proposed as a solution for biometric template security. Revocable biometric means that biometric templates are no longer fixed over time and could be revoked in the same way as lost or stolen credit cards are. In this paper, we describe a novel and an efficient approach to biometric template protection that meets the revocability property. This scheme can be incorporated into any biometric verification scheme while maintaining, if not improving, the accuracy of the original biometric system. However, we shall demonstrate the result of applying such transforms on face biometric templates and compare the efficiency of our approach with that of the well-known random projection techniques. We shall also present the results of experimental work on recognition accuracy before and after applying the proposed transform on feature vectors that are generated by wavelet transforms. These results are based on experiments conducted on a number of well-known face image databases, e.g. Yale and ORL databases.

Nonintrusive multibiometrics on a mobile device: a comparison of fusion techniques

In this article we test a number of score fusion methods for the purpose of multimodal biometric authentication. These tests were made for the SecurePhone project, whose aim is to develop a prototype mobile communication system enabling biometrically authenticated users to deal legally binding m-contracts during a mobile phone call on a PDA. The three biometrics of voice, face and signature were selected because they are all traditional non-intrusive and easy to use means of authentication which can readily be captured on a PDA. By combining multiple biometrics of relatively low security it may be possible to obtain a combined level of security which is at least as high as that provided by a PIN or handwritten signature, traditionally used for user authentication. As the relative success of different fusion methods depends on the database used and tests made, the database we used was recorded on a suitable PDA (the Qtek2020) and the test protocol was designed to reflect the intended application scenario, which is expected to use short text prompts. Not all of the fusion methods tested are original. They were selected for their suitability for implementation within the constraints imposed by the application. All of the methods tested are based on fusion of the match scores output by each modality. Though computationally simple, the methods tested have shown very promising results. All of the 4 fusion methods tested obtain a significant performance increase.

Improving performance and security of biometrics using efficient and stable random projection techniques

The challenges in biometrics research activities have expanded recently to include the maintenance of security and privacy of biometric templates beside the traditional work to improve accuracy, speed, and robustness. Revocable biometric templates and biometric cryptosystems have been developed as template protection measures. Revocability means that biometric templates could be revoked in the same way as lost or stolen credit cards are, while biometric cryptosystems aim to generate biometric keys and hashes to be used as proof of identity. Recently developed biometric protection schemes involve the use of random projections (RP) onto secret personalised domains. In this paper, we propose a novel and efficient orthonormal RP scheme to be used for the generation of revocable biometrics. We shall demonstrate the result of applying our RP transforms on face biometrics and compare its efficiency with that of the widely used RP technique based on the Gram-Schmidt process. We shall also present the results of experimental work on recognition accuracy before and after applying the proposed transform on feature vectors that are generated by wavelet transformed face images. Some security analysis of our scheme will also be presented.

Illumination and Expression Invariant Face Recognition: Toward Sample Quality-based Adaptive Fusion

The performance of face recognition schemes is adversely affected as a result of significant to moderate variation in illumination, pose, and facial expressions. Most existing approaches to face recognition tend to deal with one of these problems by controlling the other conditions. Beside strong efficiency requirements, face recognition systems on constrained mobile devices and PDAs are expected to be robust against all variations in recording conditions that arise naturally as a result of the way such devices are used. Wavelet-based face recognition schemes have been shown to meet well the efficiency requirements. Wavelet transforms decompose face images into different frequency subbands at different scales, each giving rise to different representation of the face, and thereby providing the ingredients for a multi-stream approach to face recognition which stand a real chance of achieving acceptable level of robustness. This paper is concerned with the best fusion strategy for a multi-stream face recognition scheme. By investigating the robustness of different wavelet subbands against variation in lighting conditions and expressions, we shall demonstrate the shortcomings of current non-adaptive fusion strategies and argue for the need to develop an image quality based, intelligent, dynamic fusion strategy.

Wavelet-based face verification for constrained platforms

Human Identification based on facial images is one of the most challenging tasks in comparison to identification based on other biometric features such as fingerprints, palm prints or iris. Facial recognition is the most natural and suitable method of identification for security related applications. This paper is concerned with wavelet-based schemes for efficient face verification suitable for implementation on devices that are constrained in memory size and computational power such as PDA’s and smartcards. Beside minimal storage requirements we should apply as few as possible pre-processing procedures that are often needed to deal with variation in recoding conditions. We propose the LL-coefficients wavelet-transformed face images as the feature vectors for face verification, and compare its performance of PCA applied in the LL-subband at levels 3,4 and 5. We shall also compare the performance of various versions of our scheme, with those of well-established PCA face verification schemes on the BANCA database as well as the ORL database. In many cases, the wavelet-only feature vector scheme has the best performance while maintaining efficacy and requiring minimal pre-processing steps. The significance of these results is their efficiency and suitability for platforms of constrained computational power and storage capacity (e.g. smartcards). Moreover, working at or beyond level 3 LL-subband results in robustness against high rate compression and noise interference.

A wavelet-based approach to face verification/recognition

Face verification/recognition is a tough challenge in comparison to identification based on other biometrics such as iris, or fingerprints. Yet, due to its unobtrusive nature, the face is naturally suitable for security related applications. Face verification process relies on feature extraction from face images. Current schemes are either geometric-based or template-based. In the latter, the face image is statistically analysed to obtain a set of feature vectors that best describe it. Performance of a face verification system is affected by image variations due to illumination, pose, occlusion, expressions and scale. This paper extends our recent work on face verification for constrained platforms, where the feature vector of a face image is the coefficients in the wavelet transformed LL-subbands at depth 3 or more. It was demonstrated that the wavelet-only feature vector scheme has a comparable performance to sophisticated state-of-the-art when tested on two benchmark databases (ORL, and BANCA). The significance of those results stem from the fact that the size of the k-th LL- subband is 1/4k of the original image size. Here, we investigate the use of wavelet coefficients in various subbands at level 3 or 4 using various wavelet filters. We shall compare the performance of the wavelet-based scheme for different filters at different subbands with a number of state-of-the-art face verification/recognition schemes on two benchmark databases, namely ORL and the control section of BANCA. We shall demonstrate that our schemes have comparable performance to (or outperform) the best performing other schemes.

Enhancing face recognition at a distance using super resolution

The characteristics of surveillance video generally include low-resolution images and blurred images. Decreases in image resolution lead to loss of high frequency facial components, which is expected to adversely affect recognition rates. Super resolution (SR) is a technique used to generate a higher resolution image from a given low-resolution, degraded image. Dictionary based super resolution pre-processing techniques have been developed to overcome the problem of low-resolution images in face recognition. However, super resolution reconstruction process, being ill-posed, and results in visual artifacts that can be visually distracting to humans and/or affect machine feature extraction and face recognition algorithms. In this paper, we investigate the impact of two existing super-resolution methods to reconstruct a high resolution from single/ multiple low-resolution images on face recognition. We propose an alternative scheme that is based on dictionaries in high frequency wavelet subbands. The performance of the proposed method will be evaluated on databases of high and low-resolution images captured under different illumination conditions and at different distances. We shall demonstrate that the proposed approach at level 3 DWT decomposition has superior performance in comparison to the other super resolution methods.

Face recognition in the presence of expression and/or illumination variation

Face recognition is a technically difficult task that forms part of an ever-growing number of applications. The challenge becomes more complex by requirement of robustness against variation in facial expressions, pose and lighting condition. In this paper, we present a wavelet-based face recognition scheme and report on its performance on 2 databases. We shall also demonstrate that it is more robust against varying facial expressions than a known scheme that has been developed specifically for that purpose. We shall also report on the positive effect of a simple procedure to reduce the effect of variation in illumination level on accuracy, in contrast to histogram equalization.

Multi-factor biometrics for authentication: a false sense of security

Multi-factor biometric authentications have been proposed recently to strengthen security and/or privacy of biometric systems in addition to enhancing authentication accuracy. An important approach to multi-factor biometric authentication is to apply User-Based Transformations (UBTs) on biometric features. Typically, UBTs rely on generating user-based transformation keys from a password/PIN or retrieved from a token. One significant advantage of employing UBTs is its ability to achieve zero or near zero Equal Error Rate (EER) i.e. a clear separation of genuine and imposter distributions. However, the effect of compromised transformation keys on authentication accuracy has not been tested rigorously. In this paper, we challenge the myth that has been reported in the literature that in the case of stolen transformation key(s), accuracy drops but remains close to the accuracy of biometric only system. Moreover, we shall show that a multi-factor authentication system setup to operate at a zero EER has a serious security lapse in the event of stolen or compromised keys. In such a scenario, the False Acceptance Rate (FAR) of the system reaches unacceptable levels. We shall demonstrate this by experiments conducted on face and fingerprint biometrics, and show that an imposter with a stolen key needs no more than two attempts on average to be falsely accepted by the biometric system.