Robust LAD Based Occluded Face Recognition using a Support Map

Recently, face recognition has been widely studied. Many works use different fidelity terms to gain better robust performance in the case of occlusion. Most of them use least-squares-based estimation as the fidelity term, which is not far from the real situation especially under the occlusions. Moreover, these methods ignore the knowledge that occlusion has local spatial continuity. To get the more robust result, this paper proposes a novel method, namely robust least-sum of absolute deviations (RLAD) based occluded face recognition using a support map. Although LAD regression is more robust than LS regression, we propose a support map as the weight vector to gain robustness. The RLAD alternatively updates the support map and the representation code. Then, the classification stage is present based on the support map and the representation code. Extensive experiments on the Extended Yale B public face database demonstrate the effectiveness and robustness of the RLAD in face recognition against occlusion.