Generative Image Inpainting for Large-Scale Edge Area

Abstract

In recent years, applying deep learning to computer vision is a very popular research direction, and a number of models with amazing effects have appeared. Deep learning-based approaches for end-to-end image inpainting have shown promise results. Recent research has made great progress in repairing rectangular and free-form areas but there are still many problems and room for improvement. For example, artifacts, blur and color missing still exist among the completion results of the large-scale border area. In this paper, we propose an end-to-end GAN-based image inpainting method, which has a better effect on the large boundary area. Our model is a two-stage adversarial network. The first stage completes the corresponding edge image, and the second stage uses the edge image generated in the first stage as a prior to complete the color image. We added parallel residual blocks to the edge completion network, and for the image completion network we replace the original residual blocks with multi-scale dilated convolution fusion blocks. Besides, a content loss based on DenseNet is added to the second stage. Experiments on multiple publicly available datasets show that our results have better effects on larger edge areas and can increase the average PSNR (Peak Signal to Noise Ratio) and SSIM (Structural Similarity Index).