Image Super-Resolution by Combining Non-Local Sparse Attention and Residual Channel Attention

Abstract

Single Image Super Resolution (SISR) is an illposed problem that aims to gener ate a high resolution (HR) image from a single low resolution (LR) image. The low resolution image and its associated features are very rich in low frequency information. The main objective of super resolution is to add relevant high frequency detail to complement the available low frequency information. Classical techniques such as nonlocal similarity and sparse representations both have shown promising results in SISR task in past decades. Nowadays, deep learning techniques such as convolutional neural networks (CNN) can extract deepeatures to improve the results of SISR task. However, CNN does not explicitly consider the similar information in the image. Hence, we employ non local sparse attention (NLSA) module in the CNN framework such that it can explore the non local similarity within an image. We consider sparsity in the non local operation by focusing on a particular group named attention bin among many groups of features. Non local Sparse Attention is intended to retain the longrange of non local operation modeling capacity while benefiting from the efficiency and robustness of sparse representation. Additionally, we try to rescale the channel specific features adaptively while taking into account channel interdependence by using residual channel attention. In this thesis work, we try to incorporate and combine the advantages of non local sparse attention (NLSA) and residual channel attention to produce results similar to state of the art methods.