Using Generative Adversarial Networks for Single Image Super-Resolution

Abstract

In digital image processing, retrieving a high-resolution image from its low-resolution version is considered to be a major topic. Super-resolution SR is a problem that has direct applications in numerous disciplines like medical diagnosis, satellite imagery, face recognition and surveillance. The choice of the optimization function has been a major factor in previous super-resolution approaches. Optimizing metrics that are determined based primarily on pixel-level variance is the most common objective for supervised super-resolution algorithms. Nevertheless, these approaches do not output perceptually satisfactory images. This paper adopts an idea in which depending on only pixel-space similarity is avoided. Instead, the major goal is to utilize a content loss based on perceptual resemblance using feature maps of the VGG network in conjunction with Generative Adversarial Networks GAN. This depends on training two networks: a generator and a discriminator. In an adversarial game, they compete to outperform each other with an ultimate objective of producing super-resolution images that are identical to the real high-resolution images that already exist in the dataset. This paper’s main contribution is a comparison of the effects of taking the VGG-19 content loss from various layers. On public benchmarks, super-resolution GAN was successful in recovering detailed textures from highly downsampled images. SRGAN reveals large gains in perceptual quality in a mean opinion score MOS test.