SwinFR: Combining SwinIR and fast fourier for super-resolution reconstruction of remote sensing images

作     者：Zhang, Jianfei Tu, Yanjie 

作者机构：Heilongjiang Univ Sci & Technol Sch Comp & Informat Engn Harbin Peoples R China 

出 版 物：《DIGITAL SIGNAL PROCESSING》 (Digital Signal Process Rev J)

年 卷 期：2025年第159卷

核心收录：

学科分类：0808[工学-电气工程] 0809[工学-电子科学与技术（可授工学、理学学位）] 08[工学] 

基　　金：National Natural Science Foundation of China 

主　　题：Remote sensing Super-resolution reconstruction Swin transformer Fast fourier convolution Multi-scale feature learning 

摘      要：In recent years, although transformer technology has been widely used in the field of image super-resolution, it still has some limitations in handling low-frequency information and local features of images, especially when facing the challenges posed by the complex backgrounds and diverse surface features of remote sensing images. In this paper, a new image super-resolution algorithm based on the Swin Transformer, called SwinFR, is proposed. The core innovation lies in the design of the Residual Swin Transformer Fourier Block (RSTFB), which combines the residual Swin Transformer layer with the fast Swin Transformer Fourier Block. This combination improves the model s ability to capture low-frequency information and preserve image structural details. The module also enhances deep feature extraction and inter-layer information flow by integrating convolution and residual concatenation, which improves the model s feature integration capability and its ability to handle complex background information in remote sensing images. In addition, this paper introduces the Multi-Scale Feature Learning (MSFL) module, which further enhances the processing of local and global information and enables high-quality image reconstruction. Experimental results show that SwinFR outperforms existing methods in key metrics such as visualization, PSNR, SSIM, and MOS on both the UCMLU and SIRI-WHU datasets, effectively demonstrating its superiority and practicality.