UR-SRGAN : a generative adversarial network for real-world super-resolution with a U-Net-based discriminator

Abstract

Despite several improvements in Super-Resolution deep learning techniques, these proposed methods tend to fail in many real-world scenarios since their models are usually trained using a pre-defined degradation process from high-resolution (HR) ground truth images to low-resolution (LR) ones. In this work, we propose a supervised Generative Adversarial Network (GAN) model for Image Super-Resolution which has as the first stage to estimate blur kernels and noise estimation from real-world images to generate LR images for the training phase. Furthermore, the proposal includes implementing a novel U-Net-based discriminator, to consider an input image’s global and local context, and it allows employing a CutMix data augmentation for consistency regularization in the two-dimensional output space of the decoder. The proposed model was applied to three main datasets that are ordinarily used in super-resolution official competitions. The commonly-used evaluation metrics for image restoration were used for this evaluation: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), Learned Perceptual Image Patch Similarity (LPIPS) and Natural Image Quality Evaluator (NIQE). After implementing this new architecture, three other prominent models in the state-of-the-art GAN proposals for super-resolution were trained with the same parameters and databases to perform a global comparison between all of them. Finally, the results of the experimentation in training and evaluation tasks between all the models suggest an improvement in the performance of the presented work compared to the other architectures based on the established metrics.