An improved Image Inpainting Algorithm Based on Attention Fusion Module

When we try to reproduce existing inpainting algorithms, our performances are often not stable, and the results are not up to expectations. In this paper, we devise an improved image inpainting algorithm mainly constructed by an attention fusion module (AFM) and a refined inpainting module. The AFM evaluates the existing methods' results pixelwise, similar to humans judging the quality of the inpainting results with naked eyes. The evaluations make a quantitative judgment on the inpainting quality and fused with pre-inpainted images. The refined inpainting module focuses on the areas with poor quality and improves them. This method address the shortcomings of other models like smoothness and texture sharpening. The experimental results on multiple datasets show significant improvements based on existing methods.