Region Attention Network For Single Image Super-resolution

Abstract

The task of single image super-resolution (SISR) is a highly inverse problem because it is very challenging to reconstruct rich details from blurry images. Most previous super-resolution (SR) methods based on Convolution Neural Network (CNN) tend to design a more complex network structure to directly learn the mapping between low-resolution images and high-resolution images. Nevertheless, it is not the best choice for blindly increasing the depth of the network since the performance improvement may not increase but the computing cost. In order to tackle this problem, we propose an effective method, which integrates the image prior to the model to enhance image reconstruction. In fact, the role of categorical prior as an important image feature has been widespreadly used in several high challenge computer vision tasks. In this work, we propose a region attention network (RAN) to recovery clear SR images with the assistance of categorical prior. The proposed RAN can be divided into two branches: the rough image reconstruction branch and the subtle image reconstruction branch. Meanwhile, we also propose a coupling group to make full use of the feature of two branches. Extensive experiments demonstrate that our RAN obtains satisfactory performance with the help of image categorical prior.