An approach for face recognition based on fusion of DTCWT and Manifold Regularized Orthogonal Discriminant Analysis

In this paper, a novel subspace learning method called manifold regularized orthogonal discriminant analysis (MRODA) is first proposed. Based on within-class local geometry preservation and least square regression framework for LDA, MRODA can encode both the local geometry and discriminant structures of face data manifolds, and can address the small sample size problem through pseudo-inverse resolution. The transform vectors are orthogonalized to improve their discriminatory performance. Based on the selected dual-tree complex wavelet transform features, an approach for face recognition based on the fusion of spatial and frequency features is developed. Experimental results on ORL, Yale and AR face databases show the effectiveness of the proposed approach.