融合跨阶段局部网络与注意力的遥感图像锐化方法  

Remote Sensing Image Pansharpening Algorithm Fusing Cross Stage Partial Network and Attention

作　　者：刘蕙绮 陈应霞[1] LIU Huiqi;CHEN Yingxia(School of Computer Science,Yangtze University,Jingzhou 434023,CHN)

机构地区：[1]长江大学计算机科学学院,湖北荆州434023

出　　处：《半导体光电》2024年第5期837-846,共10页Semiconductor Optoelectronics

基　　金：2022年中国高校产学研创新基金项目(2021RYC06002);2022年度湖北省自然资源厅自然资源科技项目(MB1644959/2021-70118);2022年湖北省教育厅科研计划项目(B2022040);上海市多维度信息处理重点实验室开放课题基金资助项目(MIP20222);中央高校基本科研业务费专项资金资助(MIP20222).

摘　　要：针对遥感图像全色锐化中存在光谱分布不均、空间细节缺失的问题,提出一种融合跨阶段局部网络(CSPNet)与无参数注意力(SimAM)的遥感图像锐化方法。首先,在主干结构中引入CSPNet,利用普通卷积加跳连的方式替代特征提取中的残差块,以缓解梯度冗余,提升模型学习力。其次,添加SimAM块,直接从特征中推导出三维权重,而后反向优化提取到的特征,使得模型能提取到更深层次的特征信息。最后,设计一种可学习作差参数来控制相减权重,以便突出融合图像的边缘信息。实验结果表明,所提方法不仅能改善模型的梯度冗余,还能进一步提升融合图像的空谱分辨率。A remote sensing image sharpening method is proposed that combines cross stage local network(CSPNet)and parameter free attention(SimAM)to address the issues of uneven spectral distribution and missing spatial details in pansharpening of remote sensing images.Firstly,CSPNet is introduced into the backbone structure to replace the residual blocks in feature extraction with ordinary convolution and skip connections,in order to alleviate gradient redundancy and improve model learning power.Secondly,add SimAM blocks to directly derive 3D weights from the features,and then reverse optimize the extracted features to enable the model to extract deeper level feature information.Finally,design a learnable subtraction parameter to control the subtraction weights,in order to highlight the edge information of the fused image.The experimental results show that the proposed method can not only improve the gradient redundancy of the model,but also further enhance the spatial spectral resolution of the fused image.

关 键 词：全色锐化 跨阶段局部网络 无参数注意力 可学习作差参数 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]