Bi-weighted robust matrix regression for face recognition

Regression analysis based classification methods have attracted much interest in the face recognition area. However, dealing with partial occlusion or illumination is still one of the most challenging problems. In most of the current methods, the image needs to be stretched into a vector and each pixel is assumed to be generated independently, which ignores the dependence between pixels of the error image. That is, these methods do not consider the structure information of the image with continuous occlusion or disguise in modeling. In this paper, it is found that the non-convex function of the singular values can well describe the low rank structure of the image data. By virtue of this fact, we propose a bi-weighted robust matrix regression (BWMR) model for face recognition with structural noise, in which the non-convex function of the singular values is used as regularization. The alternating direction method of multipliers (ADMM) is applied to solving the proposed model. Experimental results demonstrate that the proposed method is more robust and effective than the state-of-the-art methods when handling the structural errors.