Abstract:

Although difficult to deal with, face recognition with varying illumination and occlusion has been widely investigated in recent years. Motivated by the popular methods of robust principal component analysis (RPCA) and sparse representationbased classification (SRC), we present a novel algorithm which uses the sparse error of low rank representation (LRR) for face recognition with varying illumination and occlusion. As for each type of training samples, we first calculate their lowrank matrix using LRR and then construct low rank projection between the training face matrix and the obtained lowrank data matrix. With the constructed lowrank projection, any test face image can obtain a lowrank matrix and a sparse error matrix corresponding on each face category. In order to fully extract the discrimination information of the sparse error image, its smoothness and edge information are analyzed respectively. Furthermore, a set of concrete classification criteria is proposed, which fuses smoothness information with the edge information using weighted sum rules. Experiment results on face databases of AR and the Extended Yale B confirm that the proposed method is robust to varying illumination and occlusion, and has better recognition rate than many other methods.

Key words: low rank representation;low rank projection;sparse error image;face recognition