Two-stage Network For Single Image Super-Resolution

The task of single-image super-resolution (SISR) is a highly inverse problem because it is very challenging to reconstruct rich details from blurred images. Most previous super-resolution (SR) methods based on the convolutional neural networks (CNN) tend to design more complex network structures to directly learn the mapping between low-resolution images and high-resolution images. However, this is not the best choice to blindly increase the network depth, because the performance improvement may not increase, but it will increase the computational cost. To solve this problem, we propose an effective method that learns high frequency information in high-resolution images to enhance the image reconstruction. In this work, we propose a two-stage network (TSN) to recover clear SR images. The proposed TSN firstly learns the high frequency information in high-resolution images, then learns how to transform to high-resolution images. A large number of experiments show that our TSN achieves satisfactory performance.