Robust Face Recognition with Occlusion by Fusing Image Gradient Orientations with Markov Random Fields

Abstract

Partially occluded faces are very common in automatic face recognition (FR) in the real world. We explore the problem of FR with occlusion in the domain of Image Gradient Orientations (IGO) and center on the probabilistic generative model of occluded images. The existing works usually put stress on the error distribution in the non-occluded region but neglect the distribution in the occluded region for the unpredictability of occlusions. However, in the IGO domain, this distribution can be built simply and elegantly as a uniform distribution in the interval $\left[ -\pi ,\pi \right)$. We fully use this distribution to build the probabilistic error model conditioned on the occlusion support and construct a new error metric, which fully harnesses the spatial and statistical local information of two compared images and plays a very important role in initializing the occlusion support. In addition, we extend the definition of occlusions to other variations, such as highlight illumination changes, and suggest these occlusion-like variations should also be detected and excluded from further recognition. To detect the occlusion support accurately, the contiguous structure of occlusions is modeled using a Markov random field (MRF). By fusing IGO with MRF, we propose a new error coding model, called Double Weighted Error Coding (DWEC), for robust FR with occlusion. Experiments demonstrate the effectiveness and robustness of DWEC in dealing with occlusion and occlusion-like variations.