基于双域增强Transformer的图像超分辨率重建  

Image super-resolution reconstruction based on double-domain enhanced Transformer

作　　者：傅慧滢 杨高明[1] 王瑜 FU Huiying;YANG Gaoming;WANG Yu(School of Computer Science and Engineering,Anhui University of Science and Technology,Huainan232001,China)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232001

出　　处：《哈尔滨商业大学学报(自然科学版)》2025年第2期143-151,共9页Journal of Harbin University of Commerce:Natural Sciences Edition

基　　金：国家自然科学基金资助项目(52374155);安徽省自然科学基金资助项目(2308085MF218).

摘　　要：针对现有图像超分辨率方法捕捉图像细节信息不充分导致生成图片质量不佳的问题,提出基于双域增强Transformer(DDET)的图像超分辨率重建方法.该算法从空间域信息学习和频域信息学习两个角度设计模型,通过交替连接空间域增强Transformer模块(SETB)和频域增强Transformer模块(FETB),在提取空间域信息的同时有效学习频域信息,增强网络信息提取能力.此外,为了使网络充分关注全局和局部信息,设计特殊的卷积结构与频域信息提取模块融合,进一步提高重建图像质量.相较于基于多尺度残差网络的图像超分辨率(MSRN),当放大倍数为3时,DDET在基准数据集Set5、Set14、Urban100、BSD100上,峰值信噪比(R PSN)指标分别提升0.37、0.22、0.69、0.18 dB;视觉对比上,DDET生成图片纹理更清晰.实验结果表明,DDET可以关注到更多细节信息,生成更高质量的图像,表现出更优越的性能.To address the limitation of existing image super-resolution methods in effectively capturing fine image details,which leads to the generation of low-quality images,a double-domain enhanced Transformer(DDET)was proposed for image super-resolution reconstruction.The algorithm designed the model from two perspectives:learning spatial and frequency domain information.By alternately connecting the spatial-domain enhanced Transformer block(SETB)and the frequency-domain enhanced Transformer block(FETB),The spatial domain information was extracted while the frequency domain information was effectively learned,which further enhances the network information extraction capability.In addition,To make the network pay more attention to global and local information,a special convolution structure was designed to be integrated with the frequency domain information extraction module to improve the quality of reconstructed images further.Compared with multi-scale residual network for image super-resolution(MSRN),the peak signal-to-noise ratio(R PSN)improved by 0.37 dB,0.22 dB,0.69 dB,and 0.18 dB on standard test sets,including Set5,Set14,Urban100,and BSD100,at the magnification of 3.Visual effects show that DDET generates clearer image textures.Experimental results demonstrated that DDET effectively captures finer details,produce higher-quality images,and achieved superior overall performance.

关 键 词：图像超分辨率 TRANSFORMER 频域 傅里叶变换 深度学习 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]