Sparse representation via multi-feature based Fisher Discrimination Dictionary Learning

In this paper, we propose a multi-feature based sparse representation method named multi-feature Fisher Discrimination Dictionary Learning (MFDDL) and apply it to face recognition. In the new proposed method, firstly, to extract the texture information, multi-scales and multi-orientations Gabor Wavelet Transform is proposed for feature representation. Then the local characteristics of the face Gabor feature is future enhanced by multi-block rotation invariant LBP, which extracts statistically-significant histogram feature and meanwhile, reduces the dimension of the extracted Gabor features. Finally, the Fisher Discrimination Dictionary Learning is utilized to achieve face recognition. Experimental results on the AR face database show that the proposed method can effectively overcome the effect of light variation and occlusion, and can improve the face image recognition performance.