Image Inpainting Based on Generative Adversarial Networks

Abstract

Image inpainting aims to fill missing regions of a damaged image with plausibly synthesized content. Existing methods for image inpainting either fill the missing regions by borrowing information from surrounding areas or generating semantically coherent content from region context. They often produce ambiguous or semantically incoherent content when the missing region is large or with complex structures. In this paper, we present an approach for image inpainting. The completion model based on our proposed algorithm contains one generator, one global discriminator, and one local discriminator. The generator is responsible for inpainting the missing area, the global discriminator aims evaluating whether the repair result has global consistency, and the local discriminator is responsible for identifying whether the repair area is correct. The architecture of the generator is an auto-encoder. We use the skip-connection in the generator to improve the prediction power of the model. Also, we use Wasserstein GAN loss to ensure the stability of training. Experiments on CelebA dataset and LFW dataset demonstrate that our proposed model can deal with large-scale missing pixels and generate realistic completion results.