A New Regularized Orthogonal Local Fisher Discriminant Analysis for Image Feature Extraction

Local Fisher Discriminant Analysis (LFDA) is a feature extraction method which combines the ideas of Fisher discriminant analysis (FDA) and locality preserving projection (LPP). It works well for multimodal problems. But LFDA suffers from the under-sampled problem of the linear discriminant analysis (LDA). To deal with this problem, we propose a regularized orthogonal local Fisher discriminant analysis (ROLFDA) to improve the performance of LFDA. ROLFDA finds the optimal projection vectors in the range space of local mixture scatter matrix. The new local within-class scatter matrix is approximated by regularization method in order to regulate the eigenvalues of the newly local within-class scatter matrix. The projection vectors are orthogonal by utilizing the trace ratio criterion. Experimental results on the Yale, ORL and CMU PIE face databases demonstrate that the ROLFDA is an effective algorithm.