Image Super-Resolution Using Convolutional Neural Network

Abstract

A deep learning technique for the super-resolution of a single image. With our approach, spatial dependencies are captured and end-to-end mapping between the low/high-resolution images is learned. A deep convolutional neural network (CNN) is used which accepts the low-resolution images as input and produces the high-resolution ones used to represent the mapping. This model demonstrates a lightweight construction, high restoration quality, and quick performance for practical online usage. This paper investigates multiple network architectures and parameter settings to accomplish trade-offs between performance and speed. Furthermore, our model is built to handle three color channels at the same time and demonstrate improved overall reconstruction quality.