SILP-autoencoder for face de-occlusion

Abstract

Recognizing faces with partial occlusion is a challenging problem in many real-world applications. Although various methods have been proposed to deal with the facial image de-occlusion tasks, most of them only concern the local features of occluded images, obviously ignoring the global facial expressions and structural prior information. In this paper, we propose a novel end-to-end SILP-Autoencoder to effectively restore partial occluded faces. To improve the recovery quality and occlusion removal robustness, our framework mainly consists of two components, Laplacian prior subnetwork, and left-and-right symmetric match module (LR-match module), which preserve the global facial expression features and fully make use of the symmetrical characteristics of facial regions and structures respectively. Based on the above characteristics, a composite loss function is designed to achieve end-to-end training of the entire network. Extensive experiments on the face expression datasets with various shaded areas suggest that our approach achieves superior performance against the state-of-the-art methods. In particular, our method is more useful for facial detail recovery and distortion expression suppression.