Degradation Reconstruction Loss: A Perceptual-Oriented Super-Resolution Framework for Multi-downsampling Degradations

Zongyao He, Zhi Jin, Xiao Xu, Lei Luo

Abstract: Recent years have witnessed the great success of deep learning-based single image super-resolution (SISR) methods. However, most of the existing SR methods assume that low-resolution (LR) images are purely bicubic downsampled from high-resolution (HR) images. Once the actual degradation is not bicubic, their outstanding performance is hard to maintain. Although several SR methods have super-resolved LR images with multiple blur kernels and noise levels, they still follow the bicubic downsampling assumption. To address this issue, we propose a novel degradation reconstruction loss (DRL) to capture the degradation-wise differences between HR images and SR images based on a degradation simulator. By involving the proposed degradation simulator and the loss, a perceptual-oriented SR framework for multi-downsampled images is formed. Extensive experimental results demonstrate that our method outperforms the state-of-the-art perceptual-oriented SR methods on both multi-downsampled datasets and bicubic downsampled datasets.