Image feature representation with orthogonal symmetric local weber graph structure

Image feature representation is a hot topic in the computer vision field. Inspired by Weber's law and local graph structure (LGS), we propose a novel image feature representation descriptor, called orthogonal symmetric local weber graph structure (OSLWGS). It contains two components: differential excitation pattern (DEP) and orthogonal symmetric LGS (OSLGS). In particular, DEP is extended by bringing difference of Gaussian (DoG), which can make OSLWGS robust to image noise. In addition, OSLGS can overcome some defects of LGS including non-symmetric and single horizontal structure problems. Furthermore, 2D OSLWGS histogram is generated by fusing DEP and OSLGS to improve the discriminative power and obtain more precise image description. And then, it is further encoded into 1D histogram and classified via sparse representation. Extensive experiments on FERET, CMUPIE, LFW, Yale B, simulated YALE partial occlusion, RawFooT and PhoTex databases validate the effectiveness of the proposed OSLWGS. Experimental results demonstrate that the proposed algorithm is an efficient and robust method compared with some state-of-the-art approaches.