Robust face recognition under varying illumination and occlusion considering structured sparsity

A large amount of work has been done over the past decades in face recognition (FR). Most of them deal with uncontrolled variations such as changes in illumination, pose, expression and occlusion individually. However, limited work focuses on simultaneously handling multiple variations. In real-world environment, uncontrolled variations usually coexist. FR approaches which are robust to one kind of variation may fail to deal with another. In this paper, we propose an approach considering structured sparsity to deal with the illumination changes and occlusion at the same time. Our approach represents a face image taking into account that the face images usually lie in the structured union of subspaces in a high dimensional feature space. Considering the spatial continuity of the occlusion, we propose a cluster occlusion dictionary for occlusion modelling. In addition, a discriminative feature is embedded in our model to correct the illumination effect. This enables our approach to handle images that lie outside the illumination subspace spanned by the training set. Experimental results on public face databases show that the proposed approach is very robust to large illumination changes and occlusion.