I. Introduction
The goal of single image super-resolution is to construct a high resolution (HR) image from a low resolution (LR) image input. This problem is an classical and active topic in image processing, which is also a crucial step in many practical situations, e.g. image display, remote sensing, medical imaging and so on. However, image super-resolution problem is an inherently ill-posed problem, where many HR images may produce the same LR image when down-sampled. As a result, how to generate an HR image with good visual perception and as similar as its ground truth has become the goal of image super-resolution.