Occlusion recovery face recognition based on information reconstruction

The face data in the wild have a face occlusion, complex background, and inestimable posture, leading to more sample classification errors. To alleviate the problem of face occlusion, we apply facial information reconstruction to occluded face recognition. Our model processes face images with variable poses and reconstructs face information to compensate for information loss due to occlusion. In this context, we propose an unconstrained face recognition method based on information reconstruction and occlusion dictionary learning. One is that 3D facial information reconstruction makes up for the lost part of self-occluded face images. The other is that the Gabor-based occlusion dictionary learning method increases feature diversity and better represents occluded faces. We conduct experiments on the occluded AR dataset to verify the robustness of occluded dictionary learning. Compared with the classical algorithm, the accuracy of the CAS-PEAL dataset improves by 40.3%. The experimental results of the LFW dataset are 5.57-12.21% higher than those of the state-of-the-art algorithm, which indicates that the proposed method realizes effective occlusion face recognition.