Using a Two-Stage GAN to Learn Image Degradation for Image Super-Resolution

Abstract

Recent super-resolution (SR) methods based on generative adversarial networks (GANs) almost assume that the degradation process is known. Most of these works are to use bicubic or bilinear down-sampling to obtain low-resolution (LR) images, but for real-world images, these methods often can not recover the details well. Thus affecting the performance. In this paper, we propose to first build a self-attention gradient degradation GAN, then build a self-attention gradient super-resolution GAN to alleviate the above problem. Specifically, first of all, we learn the down-sampling process by self-attention gradient degradation GAN to approximate the real-world degradation of high-resolution (HR) images, and use unpaired HR and LR images in the training process. Then, we get the LR images by self-attention gradient degradation GAN, and send them into the self-attention gradient super-resolution GAN together with the corresponding original HR images to get the SR images. The experimental results show that our method is superior to other state-of-the-art methods in terms of FID and we get competitive results on PSNR. It also potentially means that our method can be used for other categories of images.