We propose a robust approach to discriminant\udkernel-based feature extraction for face recognition and verification.\udWe show, for the first time, how to perform the eigen analysis\udof the within-class scatter matrix directly in the feature space.\udThis eigen analysis provides the eigenspectrum of its range space\udand the corresponding eigenvectors as well as the eigenvectors\udspanning its null space. Based on our analysis, we propose a kernel\uddiscriminant analysis (KDA) which combines eigenspectrum\udregularization with a feature-level scheme (ER-KDA). Finally, we\udcombine the proposed ER-KDA with a nonlinear robust kernel\udparticularly suitable for face recognition/verification applications\udwhich require robustness against outliers caused by occlusions\udand illumination changes. We applied the proposed framework\udto several popular databases (Yale, AR, XM2VTS) and achieved\udstate-of-the-art performance for most of our experiments.
展开▼
