Dual branches network for image super-resolution

Abstract

Recent studies have shown that deep convolutional neural networks (CNNs) significantly boosted the performance of single-image superresolution (SISR). In this Letter, the authors present a novel dual branches network (DBN) for SISR. Different from traditional CNN, the authors' DBN utilises the benefits of the residual structure and the densely connected structure together. Their key strategy is to divide the input path of the network into dual branches: A residual branch and a dense branch. This dual branches structure reuses valuable features and explores new features effectively. Experimental comparisons demonstrated the high ability of their DBN over the state-of-the-art framework for SISR with alleviating blurs of output images.