Generative Adversarial Networks for Inpainting Occluded Face Images

Abstract

Convolutional neural network (CNN) recognizers have made substantial progress in face recognition. Existing recognizers have a lot of power over un-occluded faces, but their performance suffers when it comes to recognizing occluded faces directly. As occlusions cause a lack of visual and recognition signals. The face inpainting task is complicated as it requires generating new pixels for the missing regions of the face image. Generative adversarial networks (GAN) are more suitable for this task when we have to reconstruct visually plausible occlusions in face inpainting. The GAN model is able to generate and inpaint the missing regions of the image. In this paper, we have developed a methodology that makes use of GAN and contextual attention to inpaint images. This image inpainting has applications in the area of face recognition, face animation, and generating synthetic data.