Multi-degradation image super-resolution using texture-transfer

Abstract

Most leading Image Super-Resolution (SR) methods assume that input low-resolution (LR) images are bicubically downsampled from their high-resolution (HR) counterparts. This causes state-of-the-art models to not perform as well when evaluated with non-bicubic LR images. This research presents a non-blind reference-based SR (RefSR) using multi-degradation method that aims to be more well-rounded compared to leading SR methods thus far. It uses Texture Transformer Network for Image Super-Resolution (TTSR) combined with Super-Resolution Network for Multiple Degradations (SRMD) as the base model. The approach transforms blur (degradation) kernel information that is applied on the LR to a degradation map that is fed to the network. Not only does the model perform better in non-bicubic LRs, but it also caters to LRs where the blur kernel information is not known, commonly known as real LRs. KernelGAN is used to estimate the otherwise unknown blur kernel. Performance improvements were achieved by training the model with augmented LRs and feeding the degradation map information in multiple scales in the network.