Random Degradation Model for Deep Blind Image Super-Resolution

Abstract

Degradation models are widely explored in Super-resolution (SR) field. The traditional degradation model, which mainly involves blur and downsampling degradation, cannot well simulate the degradation in real-world scenarios. Although some degradation models have made good progress by using a series of complex degradation types, they ignore some simple cases in reality. They, therefore, cannot cover the various degradations of real images. To address this problem, we propose a novel random degradation model. Specifically, our degradation model consists of various degradation types, including blur, noise, multi-step sampling, rotation and JPEG compression. Moreover, a novel random and dropout strategy is employed to generate the degradation sequence of different lengths to simulate various real-world degradations better. Based on the random degradation model, we extend the powerful ESRGAN to a practical super-resolver (namely, RDGAN). To verify the effectiveness of the new degradation model, we have made extensive comparisons on synthetic and real datasets. The results have shown that the new degradation model helps to improve the ability of the super-resolver significantly, thus benefiting existing deep super-resolution networks in practical use.