Semantic Face Completion Based on DCGAN with Dual-Discriminator

Abstract

Image completion refers to utilizing image residual information to repair the missing part. In order to solve the problem that the existing models cannot learn the deep features of the image, resulting in inaccurate artifacts in the repaired details, and the inconsistency between the repaired part and the adjacent pixels, we propose a semantic face completion method based on DCGAN with dual discriminator in this paper. We improve DCGAN with dual discriminator to ensure the consistency of local part and global image. In addition, we utilize the VGG16 network to help learn the deep features of the image to make the repaired result clearer and more realistic at the pixel level. Therefore, our method generates images by optimizing the generators with three kinds of loss functions, which are composed of dual image reconstruction loss, dual adversarial loss, and dual image feature reconstruction loss. Experiments on celebA datasets show that the model can get reasonable repairing results, and PSNR and SSIM are higher than baseline on most test datasets.