Progressive Feature Fusing for High-Resolution Image Inpainting

The purpose of image inpainting is to generate clear and realistic images. Despite the rapid development of image inpainting technology, there remain issues such as unclear texture of images and a lack of interaction between structure inpainting and texture inpainting. As a result, when dealing with large missing regions, the generated images are generally distorted. To address these problems, we propose a two-stage image inpainting network, which aims to achieve appropriate filling content through a coarse inpainting network, and subsequently enhance the content image details with a fine inpainting network. Moreover, we introduce a global and local gated convolution network (GLGC) module that utilizes long-range context information and thus performs better than the widely-used vanilla convolutions in balancing the structure and texture of images. We conduct experiments on two public datasets: CelebA-HQ and Paris StreetView. The results demonstrate that the proposed two-stage image inpainting network using our GLGC module enhances the quality of generated images on datasets with different scenes.