Face recognition with kernel sparse representation on Gabor features

In this paper, we propose a novel face recognition framework in which the image Gabor features are used for kernel sparse representation based classification (KSRC), i.e., Gabor features based KSRC (GKSRC). At first, each face image is convolved with a series of Gabor filters to extract Gabor features. To avoid careful selection of parameters for kernels, we propose to learn an optimal kernel through multiple kernel learning method (MKL) from a group of base kernels which are constructed from Gabor features. Then with the learned kernel, the kernel discriminant analysis (KDA) is used for dimension reduction. Finally, the query face image is recognized by minimizing the reconstruction error between the original image and its approximation in the kernel space. Experiments on AR, ORL and FERET databases demonstrate the effectiveness of the proposed GKSRC algorithm comparing with other face recognition schemes.