Kernel based 2D symmetrical principal component analysis for face classification

This paper presents a novel algorithm-kernel based 2D symmetrical principal component analysis (K2DSPCA), which takes full advantage of kernel method, the symmetrical property of facial image and the structural information of image (i.e., the advantage of two-dimensional PCA). Firstly, a facial image is decomposed into an even image and an odd image; Secondly, both the even image and the odd image are mapped to a high dimensional feature space (Reproducing Kernel Hilbert space, RKHS) by a nonlinear function; Thirdly, compute the eigenvectors and the eigenvectors of the even image and the odd image in RKHS, respectively; At last, select the eigenvectors with greater variance as the projection axis. We compare the performance of SPCA, 2DPCA, S2DPCA with K2DSPCA on CBCL database for binary classification, and on ORL face database for multi-category classification, respectively. The experimental results show the K2DSPCA is competitive with or superior to SPCA, 2DPCA and S2DPCA.