Single image super resolution techniques based on deep learning: Status, applications and future directions

Abstract

Single Image Super Resolution (SISR) reconstruction aims to recover high-resolution images from corresponding Low-Resolution (LR) versions, which is essentially an ill-posed inverse problem. In recent years, learning-based methods have been frequently exploited to tackle this problem, which correspond to promising calculation efficiency and performance, especially in image sharpening processing based on deep neural networks. Learning-based methods can be generally categorized as conventional methods and deep learning-based methods. This survey aims to review deep learning-based image super-resolution methods, including Convolutional Neural Networks (CNN) and Generative Adversarial Networks (GAN) based on internal network structure. Furthermore, this paper describes the applications of single-frame image super resolution in various practical fields. In addition, a few future research directions of image super resolution techniques are identified.