Occluded Image Recognition with Extended Nonnegative Matrix Factorization

This paper addresses the challenge of recognizing face and facial expression under occlusion situations. We have introduced an extension of nonnegative matrix factorization called angle and graph constrained nonnegative matrix factorization (AGNMF). The proposed model is developed in term of minimizing angle of basic cone and preserving the geometrical structure of the projective data. The experimental results in the context of occluded images demonstrate that the technique of enforcing constraints on both basic and encoding matrices works well and the AGNMF method shows superior performance to other conventional NMF approaches.