Image inpainting via progressive decoder and gradient guidance

Abstract

Very recently, with the widespread research of deep learning, its achievements are increasingly evident in image inpainting tasks. However, many existing multi-stage methods fail to effectively inpainting the larger missing areas, their common drawback is that the result of each stage is easily misguided by the wrong content generated in the previous stage. To solve this issue, in this paper, a novel one-stage generative adversarial network based on the progressive decoding architecture and gradient guidance. Firstly, gradient priors are extracted at the encoder stage to be passed to the decoding branch, and multiscale attention fusion group is used to help the network understand the image features. Secondly, multiple parallel decoding branches fill and refine the missing regions by top-down passing the reconstructed priors. This progressively guided repair avoids the detrimental effects of inappropriate priors. The joint guidance of features and gradient priors helps the restoration results contain the correct structure and rich details. And the progressive guidance is achieved by our fusion strategy, combining reimage convolution and design channel coordinate attention to fuse and reweight the features of different branches. Finally, we use the multiscale fusion to merge the feature maps at different scales reconstructed by the last decoding branch and map them to the image space, which further improves the semantic plausibility of the restoration results. Experiments on multiple datasets show that the qualitative and quantitative results of our computationally efficient model are competitive with those of state-of-the-art methods.