Structure and Texture Preserving Network for Real-World Image Super-Resolution

Abstract

Real-world image super-resolution (Real-SR) is a challenging task due to the unknown complex image degradation. Recent research on Real-SR has achieved remarkable progress by degradation process modeling; however, there are still undesired structural distortions and over-smoothed textures in the recovered images. In this letter, we propose a structure and texture preserving network, towards reducing the structural distortions while refining the perceptual-pleasant textures. Specifically, we propose a structure tensor (ST) branch to guide the restoration of high-resolution images by extracting channel-aggregated structural information. To further adaptively optimize different local texture, we replace the global discriminator with a global-local discriminator. By “local,” we mean that the discriminator loss, imposed on local areas randomly selected from the generated SR image, is minimized to generate textures with great visual perception in the selected local areas. Experimental results on five real-world datasets demonstrate the superiority of our methods in restoring structures, generating visually realistic SR images, as well as handling images of different degradation levels.