ESGAN: Edge Loss and Spatial Convolution Generative Adversarial Network for Image Inpainting

Deep learning-based algorithms have a wide application in image inpainting because of their special capabilities in synthesizing new texture structures. However, the generated images are usually blurry with artificial textures. This paper, to solve this problem, this paper introduces spatial convolution to expand the receptive field of the convolution kernel under the same computational complexity. It also proposes edge loss to emphasize the role of object shape in weight optimization. Experiments are conducted in Place2 with irregular masks, and the results show that comparing with most other existing algorithms, our proposed new network model has achieved a better performance in model scale and PSNR, SSIM, etc.