Image inpainting using Wasserstein Generative Adversarial Network

Abstract

Recent advances in convolution neural networks have shown promising results for the challenging task of filling large missing regions in an image with semantically plausible and context aware details. These learning-based methods are significantly more effective in capturing high-level features than prior techniques, but often create distorted structures or blurry textures inconsistent with existing areas. This is mainly due to ineffectiveness of convolutional neural networks in explicitly borrowing or copying information from distant locations. Motivated by these observations, we use a convolution neural networks architecture with Atrous Spatial Pyramid Pooling module, which can obtain multi-scale objection information, to be our inpainting network. We also use global and local Wasserstein discriminators that are jointly trained to distinguish real images from completed ones. We evaluate our approach on four datasets including faces (CelebA) and natural images (Paris Streetview, COCO, ImageNet) and achieved state-of-the-art inpainting accuracy.