Context-based Image Inpainting with Improved Generative Adversarial Networks

Abstract

Recent deep learning based approaches especially with generative adversarial networks and their variants have shown promising results for the challenging task of inpainting large missing regions in an image. These methods can generate visually plausible image structures and textures, but often create distorted structures or blurry textures inconsistent with surrounding areas. For this problem, we proposed a novel approach which not only can repair the defective image with global and local consistency, but also is extremely lightweight. This model is an improved version of Generative Adversarial Network, which has a global and local context discriminator to guarantee the global and local consistency of generated images, utilizes Wasserstein loss to assist the training of its generator and replaces a part of standard convolution with some dilated convolution to expand the receptive field. After experimentation, we found our approach can generate satisfactory image content on the restoration problem of specific scene images such as human faces and street views.