基于混合注意力的遥感图像超分辨率重建  

Super-resolution Reconstruction of Remote Sensing Images Based on Hybrid Attention

作　　者：姚善化[1] 潘品杨 王仲根[1] YAO Shanhua;PAN Pinyang;WANG Zhonggen(School of Electrical and Information Engineering,Anhui University of Science and Technology,Huainan Anhui 232001,China)

机构地区：[1]安徽理工大学电气与信息工程学院,安徽淮南232001

出　　处：《安徽理工大学学报(自然科学版)》2025年第1期64-73,98,共11页Journal of Anhui University of Science and Technology:Natural Science

基　　金：国家自然科学基金资助项目(62105004);安徽省高校自然科学研究项目(KJ2020A0308)。

摘　　要：目的为改善遥感图像局部区域模糊、部分细节信息重建丢失等问题。方法提出一种基于空洞卷积和混合注意力的遥感图像超分辨率重建算法。首先经过浅层特征提取模块得到浅层特征图,再利用卷积与空洞卷积以及非线性激活块相结合,扩大了整体感受野,提升了训练过程的稳定性,从而增强深层特征表达能力;其次,使用级联的空间注意力与通道注意力模块来改善高频信息缺失问题;最后,对所提取的特征进行上采样和重建获得高分辨率图像。结果在NWPU RESISC45和UCMerced-LandUse数据集上,仿真结果分析表明,该算法的峰值信噪比与结构相似性两项评价指标均优于所对比算法,在主观视觉效果上,重建图像也更能突出纹理细节信息。结论所提算法拥有更好的重建效果,提升了遥感图像的质量和可用性。Objective To solve the problems of local ambiguity inthe remote sensing images and loss of some detail informationinthe reconstruction.Methods A super-resolution reconstruction algorithm for remote sensing images based on dilated convolution and mixed attention was proposed.Firstly,the shallow feature map was obtained through the shallow feature extraction module,and then the convolution,dilated convolution and nonlinear activation block were combined to expand the overall receptive fieldand improve the stability of the training process,soasto enhance the ability to express deep features.Secondly,the cascaded spatial attention and channel attention modules were used to solve the problem of high-frequency information loss.Finally,the extracted features were upsampled and reconstructed to obtain high-resolution images.Results On the NWPU RESISC45 and UCMerced-LandUse datasets,the simulation results showed that the peak signal-to-noise ratio andthe structural similarity of the proposed algorithm were better than those of the compared algorithms,and the reconstructed images highlighted the texture details better in the subjective visual effect.Conclusion The proposed algorithm has better reconstruction effect and improves the quality and usability of remote sensing images.

关 键 词：超分辨率重建 遥感图像 空洞卷积 注意力机制 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]