Description method of Illumination invariant image features

Abstract

Image feature extraction technology is one of main topics in the field of computer vision, which has been widely applied in biological recognition, image retrieval, target detection and other fields. To overcome the drawbacks of WLD under complex illumination condition, we propose a novel illumination–insensitive feature descriptor named as anisotropic Weber synergy gradient descriptor (AWSGD). The proposed algorithm contains two parts:differential excitation component and gradient direction component. Firstly, by introducing the differential synergy excitation pattern (DSEP) and anisotropic LOG operator with variable scales and angles, we propose the anisotropic differential synergy excitation pattern (ADSEP) as the differential excitation component. Next, focused on the shortage that local gradient pattern (LGP) lacks detailed description of local features with single-layer model, we propose weighted local synergy gradient pattern (WLSGP) as the gradient direction component based on two-layer structure model and weight coefficient distribution model. Finally, ADSEP and WLSGP are fused to form AWSGD histogram. Meanwhile, we adopt XGBoost classifier to conduct related experiments on face databases CMUPIE, Yale B and texture databases PhoTex, RawFooT. The experimental results indicate that the proposed algorithm has stronger robustness to illumination variation and achieves the best performance compared with state-of-the-art methods, which has a certain theoretical significance and practical value in image recognition field under complex illumination condition.