Learning kernel subspace classifier for robust face recognition

Subspace classifiers are very important in pattern recognition in which pattern classes are described in terms of linear subspaces spanned by their respective basis vectors. To overcome the limitations of linear methods, kernel based subspace models have been proposed in the past by applying the Kernel Principal Component Analysis (KPCA). However, the projection variance in the kernel space as applied in the previously proposed kernel subspace methods, is not a good criteria for the data representation and they simply fail in many recognition problems. We address this issue by proposing a learning kernel subspace classifier which attempts to reconstruct data in the input space through the kernel subspace projection. Comparing with the pre-image methods, we emphasize the problem of how to use a kernel subspace as a model to describe input space rather than finding an approximate pre-image for each input by minimization of the reconstruction error in the kernel space. Experimental results on occluded face recognition demonstrated the efficiency of the proposed method.