Semantic segmentation guided face inpainting based on SN-PatchGAN

As a specific application of image inpainting, face inpainting based on generative adversarial network (GAN) has made great process in recent years. However, there are still many problems in the current face inpainting methods, such as asymmetric eyes, unsuitable size of nose and artificial expression. Considering the obvious structural feature of human face, this paper proposes a face image restoration method based on semantic segmentation guidance. In the base of the repair network Spectral-Normalized PatchGAN (SN-PatchGAN), the semantic segmentation network is used to guide the repair process, which can make the inpainting face image to be more realistic. Moreover, an asymmetry loss is designed to reduce the eye asymmetry. Experiments on public dataset show that our approach outperform existing methods quantitatively and qualitatively.