Person identification using face and speech biometrics

Increasing security threats have recently highlighted the importance of efficient authentication systems. Although face and speech biometrics have shown good performance, there are key robustness issues which challenge the reliability of these systems. For instance illumination, expression, pose and occlusion remain open challenges in the paradigm of face recognition. To address these robustness issues, there is a dire need of novel algorithms in these areas. This dissertation investigates the recently proposed face recognition algorithm called Sparse Representation Classification (SRC) for key robustness issues. Since local features such as eyes, ears and lips have shown better performance compared to their global counterparts, the thesis successfully extends the SRC approach to the problem of ear recognition. In the paradigm of face recognition, three novel algorithms named Linear Regression Classification (LRC), Modular LRC and Robust Linear Regression Classification (RLRC) are proposed to address various issues of severe expression variations, adverse luminance variations, contiguous occlusion and random pixel corruption. Extensive experiments have been conducted on standard databases and excellent results have been reported. In particular, using the Modular LRC approach we achieve the best result ever reported for the challenging scarf occlusion problem. Addressing the problem of luminance, the proposed RLRC algorithm is able to achieve 100% recognition accuracy on the most adversely distorted Subset 5 of the Yale Database B outperforming the contemporary illumination invariant face recognition algorithms. In the paradigm of speaker recognition, we propose two novel algorithms based on sparse representation and linear regression. These algorithms are tested on subsets of TIMIT database achieving competitive performance index compared to the state-of-art approaches. The dissertation is presented as a compilation of publications.