A-ESRGAN: Training Real-World Blind Super-Resolution with Attention U-Net Discriminators

Zihao Wei, Yidong Huang, Yuang Chen, Chenhao Zheng, Jingnan Gao

Abstract: Generative adversarial networks (GANs) have recently made great progress in blind image super-resolution (SR) with their superiority in learning mappings between manifolds, which benefits the reconstruction of image’s textural details. Recent works have largely focused on designing more realistic degradation models, or constructing a more powerful generator structure but neglected the ability of discriminators in improving visual performances. In this paper, we present A-ESRGAN, a GAN model for blind SR tasks featuring an attention U-Net based, multi-scale discriminator that can be seamlessly integrated with other generators. To our knowledge, this is the first work to introduce attention U-Net structure as the discriminator of GAN to solve blind SR problems. And the paper also gives an interpretation of the mechanism behind multi-scale attention U-Net that brings performance breakthrough to the model. Experimental results demonstrate the superiority of our A-ESRGAN over state-of-the-art level performance in terms of quantitative metrics and visual quality. The code can be find in https://github.com/stroking-fishes-ml-corp/A-ESRGAN.