A Multiscale and Densely Connected Convolutional Neural Network for Image Super-Resolution

Abstract

The aims of image super-resolution are to get high-resolution images from poor-resolution ones by software technologies. Inspired by the remarkable performance of deep neural networks for nonlinear relationships, densely connected convolutional neural network with multiscale was proposed and applied to achieve image super-resolution. A parallel module with multiscale convolutional layer has been designed to collect diverse features of the low-resolution images. Additionally, twodense blocks are added to another branch of the convolutional neural network. As a result, the closely connected convolutional neural network can avoid the vanishing-gradient problem during model training. Not only the quantitative assessment results, but also the visual assessment enable to verify that the designed network yielded high-resolution images, which are finer to the ones in comparison with the state-of-the-art approaches.