Robust Discriminative Nonnegative Patch Alignment for Occluded Face Recognition

Face occlusion is one of the most challenging problems for robust face recognition. Nonnegative matrix factorization (NMF) has been widely used in local feature extraction for computer vision. However, standard NMF is not robust to occlusion. In this paper, we propose a robust discriminative representation learning method under nonnegative patch alignment, which can take account of the geometric structure and discriminative information simultaneously. Specifically, we utilize linear reconstruction coefficients to characterize local geometric structure and maximize the pairwise fisher distance to improve the separability of different classes. The reconstruction errors are measured with weighted distance, and the weights for each pixel are learned adaptively with our proposed update rule. Experimental results on two benchmark datasets demonstrate the learned representation is more discriminative and robust than most of the existing methods in occluded face recognition.