Super-Resolution Network for General Static Degradation Model

Recent research on single image super-resolution (SISR) has made some progress. However, most previous SISR methods simply assume that a low-resolution (LR) image is bicubicly downsampled from a high-resolution (HR) image. when the LR images don’t follow this assumption, these previous methods will generate poor HR images that still retain the blur and noise information. To solve this problem, we propose the super-resolution network for general static degradation model (SR-GSD). Specifically, we propose degradation factors proposal Network (DFPN) which can automatically identify blur kernel and noise level, and furthermore, we utilize predicted degradation factors and the LR images to reconstruct the HR images in a high-resolution reconstruction network (HRN). Moreover, to simplify the training process, we unify the two-stages steps into a neural network and jointly optimize it through a multi-task loss function. Extensive experiments show that our SR-GSD can achieve satisfactory results on the general static degradation model.