Class-dependence feature analysis for large population face recognition

Reliable person recognition is important for secure access and commercial applications requiring identification. Face recognition (FR) is an important technology being developed for human identification. However, most current FR studies are based on databases with relatively small number of subjects. Developing algorithms/systems for large population face recognition (LPFR) is of significant interest in many applications such as watch lists and video surveillance. In this thesis, we propose a new method to effectively exploit the generic training data to represent a large number of subjects and thus improve the performance of correlation filters for LPFR. We propose a new general framework---class-dependence feature analysis (CFA), which can embed many common classification methods (e.g., correlation filters, support vector machines, principal component analysis etc.) to provide a discriminant feature representation for LPFR. In the CFA approach, a face discrimination model is obtained based on the subjects in the generic training set, but is applied to a larger set of subjects. In our experiments, the correlation filter based CFA approach greatly improves the face recognition rate and reduces the computational load over conventional correlation filters. Because nonlinear classifiers usually outperform their linear counterparts for classification accuracy, under the CFA framework we developed the kernel correlation filters that generate nonlinear decision boundaries. The kernel CFA (KCFA) approach significantly improves the recognition performance in our experiments, without greatly increasing the computational load. In order to further reduce the computational complexity, we extend the CFA approach to reduce the number of classifiers by using binary coding and applying error control coding (ECC) to preserve the recognition performance. This approach offers a systematic tradeoff between the computational complexity and the recognition accuracy. We test our proposed algorithms on the face recognition grand challenge (FRGC) database and show significant improvements on both verification performance and computational speed.