Inpainting larger missing regions via progressive guidance decoding network

Abstract

Abstract For images corrupted for various reasons, the size of the corrupted area is often arbitrary and it has been a challenge to inpainting the larger missing areas. Though popular multistage networks ease the inpainting difficulty by repairing damaged image from coarse to fine, their common drawback is that the result of each stage is easily misguided by the wrong content generated in the previous stage. To address this problem, we propose a novel progressive guidance decoding network. First, multiple parallel decoding branches fill and refine the missing regions by top–down passing the reconstructed priors. This inpainting way of progressive guidance avoids adverse effects of inappropriate premises, since the decoding branches can learn what priors can be utilized. And convolution layers of decoder with different locations would pass down the different priors. The joint guidance of features and gradient priors helps the inpainting result contains the correct structure and rich details. The second fold of progressive guidance is achieved by our fusing strategy, combining ghost convolution and the designed cascaded efficient channel attention (CECA) to fuse and reweight the features from different branches. CECA explores the dependencies among adjant and non-adjant channels more effectively than popular ones. Finally, we merges the different-scale feature maps reconstructed by the last decoding branch and mapping them to the image space, which further improves the semantic plausibility of the restoration results. Extensive experiments verify the effectiveness of our method in both subjective and objective evaluation.

Keywords

• Cascaded efficient channel attention (CECA)
• Efficient multiscale fusion (EMF)
• Image inpainting
• Progressive decoding network