Robust Matrix Regression for Illumination and Occlusion Tolerant Face Recognition

Jianchun Xie, Jian Yang, Jianjun Qian, Ying Tai; Proceedings of the IEEE International Conference on Computer Vision (ICCV) Workshops, 2015, pp. 46-53

Face recognition (FR) via regression analysis based classification has been widely applied in the past several years. In the existing regression methods, the testing image is represented as a linear combination of the training samples and the error image is converted into vector which is characterized by l1-norm or l2-norm. Therefore the two-dimensional structure of the error image is neglected in practice. In this paper, we operate on the two-dimensional image matrix directly, and propose a new face recognition method, namely Robust Matrix Regression (RMR). We perform the minimal weighted nuclear norm constraint on the representation error image as criterion to make full use of the low rank structural information. The proposed model is efficiently solved by an alternating direction method of multipliers (ADMM) and experimental results on public face databases demonstrate the effectiveness of our model in dealing with variations of occlusion and illumination.