Abstract

Image inpainting represents a fundamental task in computer vision, primarily focusing on the generation of missing content within an image to restore its integrity and aesthetics. Existing GAN-based approaches often yield content with ambiguity and entail high training costs. They tend to concentrate narrowly on damaged regions, leading to distortions along edges, which consequently hampers generalization. To overcome these challenges and achieve high-fidelity image inpainting, we introduce an image editing algorithm to the image inpainting task by designing two distinct networks. The first network, Edge-e4e, uses pretrained StyleGAN2 for global image generation, mitigating edge distortions in damaged regions and reducing training costs. Simultaneously, we incorporate contour information in the damaged areas to ensure the correctness of the restoration content. The second network, the Appending network, includes two style-based encoders and a generator to refine the images restored by the Edge-e4e network. Specifically, we subtract the restored images from the input images in the channel dimension to obtain a distortion map, which serves as a prior to refine the restored images. The encoders extract features from the input images and distortion map, while the generator is employed to generate optimized images. To enhance the quality of refined images, we propose integrating plugin and modulate plugin modules into the Appending network for style extraction and fusion, leveraging the available information from input images and blending it into the generator. Experimental results demonstrate that our algorithm achieves high-fidelity restoration and excellent generalization, with optimal FID and Lpips metrics of 0.0631 and 0.875, respectively.