Super-resolution via hierarchical attention and detail enhancement transformer network

Abstract

Optical imaging is a prospective technique in probing system owing to its ability for capturing scenario information. Recently, deep super-resolution techniques have significant advantages to enhance image resolution. Nonetheless, current algorithms treat all image pixels equally and fail to adequately consider the significance of edges and textures within the images. Consequently, the resulting super-resolution images often exhibit blurriness and lack clarity in the edge and texture regions. Furthermore, the existing super-resolution framework only utilizes a limited spatial range of input information for attribution analysis, disregarding the inter-pixel importance. To enhance the reconstruction quality by activating the pixels at the edges, we propose an innovative SISR transformer network. Our contributions can be summarized as follows: (1) A hierarchical attention module is introduced that functions as a plug-and-play component, enabling improved extraction of both global semantic information and local detailed information. (2) A detail enhancement module is implemented that facilitates high-fidelity image super-resolution. Numerous experiments have been conducted, demonstrating the competitive results achieved by our approach across five publicly available datasets. The reconstructed images showcase sharper edges and more intricate textures, accentuating the efficacy of our method. Our code has been released at https://github.com/DL-YHD/HADEM-SR/tree/main.