A Progressive Image Inpainting Algorithm with a Mask Auto-update Branch

Recently, learning-based image inpainting methods have made inspiring progress with squared or irregular holes. The generative adversarial networks (GANs) have been able to produce visually realistic and semantically correct results. However, most existing methods generate the results by one stage. They may have a slight advantage in computation time, but more information is lost during the inpainting process. Due to the lack of sufficient context information, these inpainting approaches cannot inpaint large holes in natural images very well. This paper proposes a progressive image inpainting algorithm for solving the above problem. This algorithm synthesizes different image components in a parallel manner within one stage. Moreover, this paper design a branch, which transmits the image features to the generative model iteratively. In each iteration, we adopt a mask auto-updating mechanism to shrink the boundary of a hole. Finally, the generative component can shrink the large corrupted regions in natural images and yield promising inpainting results.