Image Inpainting Based on Edge Features and Attention Mechanism
Yuting Fu, Dan Xu, Kangjian He, Haipeng Li, Tingting Zhang · 2022
Image inpainting as a kind important application in our life and entertainment, it also is a popular task of computer vision. The latest deep learning-based approaches have shown promising results for the challenging task of inpainting damaged regions of an image. However, there are still structural differences between the restored images and the ground truth images. Aiming at this problem, we propose a model of image inpainting called ECF-Net, ECF-Net incorporates edge information into the process of image inpainting to help damaged images to obtain more structures similar to the ground truth images, which to guide the generation of the feature of the damaged area. At the same time, we introduce the knowledge consistency attention mechanism in ECF-Net, which can obtain more reasonable semantics to eliminate blurs for image inpainting. Extensive experiments on various datasets such as CelebA-HQ, Places2 and the Paris StreetView clearly demonstrate that our method gets a better performance in vision.