Multi-scale Gated Inpainting Network with Patch-Wise Spacial Attention

Recently, deep-model-based image inpainting methods have achieved promising results in the realm of image processing. However, the existing methods produce fuzzy textures and distorted structures due to ignoring the semantic relevance and feature continuity of the holes region. To address this challenge, we propose a detailed depth generation model (GS-Net) equipped with a Multi-Scale Gated Holes Feature Inpainting module (MG) and a Patch-wise Spacial Attention module (PSA). Initially, the MG module fills the hole area globally and concatenates to the input feature map. Then, the module utilizes a multi-scale gated strategy to adaptively guide the information propagation at different scales. We further design the PSA module, which optimizes the local feature mapping relations step by step to clarify the image texture information. Not only preserving the semantic correlation among the features of the holes, the methods can also effectively predict the missing part of the holes while keeping the global style consistency. Finally, we extend the spatially discounted weight to the irregular holes and assign higher weights to the spatial points near the effective areas to strengthen the constraint on the hole center. The extensive experimental results on Places2 and CelebA have revealed the superiority of the proposed approaches.