Abstract: To restore images with large damaged areas, this paper proposes a new multi-stage inpainting network model. Firstly, the model reduces the accumulation of inpainting errors during the upsampling process by extending the feature inpainting procedure. Secondly, the multi-scale restoration  module(MSRM) is proposed, which can synthesize the information from different receptive fields to complete the feature map. Thirdly, an attention mechanism is employed to optimize the inpainting process, resolving the issue of color inconsistency in the output image caused by non-uniform restoration across different regions. Finally, the loss function is improved to make the model focus more on repair- ing the damaged regions. Experimental results show that the quality of the restored images of the model on both Places2 and CelebA datasets is improved to different degrees, and the improvement effect is more obvious as the proportion of missing pixels in the image increases.

Key words: image inpainting, deep learning, attention mechanism, encoder-decoder structure