结合多尺度多注意力的遥感图像超分辨率重构  

Combining multi-scale with multi-attention for super-resolution reconstruction of remote sensing image

作　　者：熊承义[1,2] 郑瑞华 高志荣[3] 何缘 完颜静萱 XIONG Chengyi;ZHENG Ruihua;GAO Zhirong;HE Yuan;WANYAN Jingxuan(College of Electronic and Information Engineering,South-Central Minzu University,Wuhan 430074,China;Hubei Key Lab of Intelligent Wireless Communication,South-Central Minzu University,Wuhan 430074,China;College of Computer Science,South-Central Minzu University,Wuhan 430074,China)

机构地区：[1]中南民族大学电子信息工程学院,武汉430074 [2]中南民族大学智能无线通信湖北省重点实验室,武汉430074 [3]中南民族大学计算机科学学院,武汉430074

出　　处：《中南民族大学学报（自然科学版）》2024年第5期692-700,共9页Journal of South-Central University for Nationalities：Natural Science Edition

基　　金：多谱信息处理技术国家重点实验室开放基金资助项目(6142113210303);中央高校基本科研业务费专项资金资助项目(CZY21013)。

摘　　要：视觉Transformer在改进图像超分辨率性能方面展现了良好的潜能.然而,遥感图像中不同目标表现的尺度多样性限制了其超分辨率的图像质量.为此,研究了一种结合多尺度多注意力的Transformer遥感图像超分辨率网络,旨在增强其特征学习能力,从而有效提升遥感图像的超分辨率性能.具体来说,输入特征首先通过多级下采样,得到多个尺度的特征;然后,逐级将低维特征通过一种交替密集注意力与稀疏注意力的Transformer网络进行变换,并将输出结果升维后与高维特征融合.密集注意力与稀疏注意力的结合可同时兼顾对局部相关性和全局相关性的有效提取,而多通路多尺度变换能够增强对图像小目标的建模能力.基于两个开源的遥感数据集的大量实验结果,验证了该方法的有效性.The Vision Transformer(ViT)shows promise in enhancing image super-resolution performance.However,the diverse scale of objects inherent in remote sensing images significantly constrains the quality of their super-resolution.To address this,a method for remote sensing image super-resolution using a Transformer network combining multi-scale and multi-attention is introduced,with the goal of enhancing its feature learning capability and effectively improving the superresolution performance of remote sensing images.Specifically,the input features are continuously downsampled to obtain multiple features at different scales.Subsequently,the low-dimensional features undergo a stepwise transformation through a Transformer network,utilizing alternating dense attention and sparse attention,and the resulting output is upscaled for fusion with the high-dimensional features.The combination of dense attention and sparse attention enables the simultaneous extraction of local and global dependencies,while the multi-path,multi-scale transformation enhances the modeling capability for small objects within the images.Extensive experimental results on two public remote sensing datasets validate the effectiveness of the proposed method.

关 键 词：视觉Transformer 遥感图像超分辨率 多尺度 密集注意力 稀疏注意力 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]