Single image super-resolution via multi-scale fusion convolutional neural network

An important practical issue in building Convolutional Neural Network (CNN) is a trade-off between the number of parameters and the performance. This paper proposes multiscale fusion convolutional neural network for single image superresolution. The network has the following two advantages: 1) the multi-scale convolutional layer provides the multi-context for image reconstruction; and 2) the fusion of cross-channel features reduces the dimensionality of the output of the intermediate layer. Thus the experimental results on image super-resolution demonstrate that our network achieves better performance over the state-of-art approaches.