A Review of Deep Learning for Single Image Super-Resolution

Abstract

Super-Resolution (SR) refers to the reconstruction of corresponding high-resolution images from observed low-resolution images, which has important application value in monitoring equipment, satellite images and medical images. According to the number of input images, super-resolution can be divided into single image Super-Resolution (SISR) and multi-frame image super-resolution (MISR), in which single image super-resolution is better and more respected in efficiency and practical application. So far, mainstream algorithms of SISR are mainly divided into three categories: interpolation-based methods, reconstruction-based methods and learning-based methods. Because of the high performance of in-depth learning, Deep Learning for Single Image Super-Resolution has attracted much attention in the past five years. In view of the current SISR hotspot, i.e. the single image super-resolution method based on depth learning, this paper summarizes the development history of SISR, studies the advantages and disadvantages of each excellent algorithm, and discusses the development trend and challenges of the algorithm.