Image Super-Resolution Reconstruction Method Based on Global and Local Residual Learning

In the image super-resolution (SR) algorithm, some more effective methods are obtained by convolutional neural networks (CNN). However, at present, there are two main problems in the SR algorithm using CNN: First, due to the degradation of the image, it is easy to cause partial loss of image details in a very deep network. Second, although the depth network is very powerful, when performing efficient nonlinear mapping from low-resolution (LR) input images to high-resolution (HR) target images, a large number of parameters are required, which easily causes learning difficulties. Therefore, in this paper, combined global residual learning (GRL) and local residual learning (LRL), we propose a new method to effectively obtain image details. Especially, using the stacked local residual block (LRB) structure for nonlinear mapping the parameters of CNN can be reduced and the image degradation problem can be overcome effectively. Due to the high correlation between LR images input and HR images output in the network, we can use the skip-connection method to reconstruct of HR images from most of the LR information. The experimental results show that the proposed image SR algorithm is effective.