基于多尺度残差注意力网络的全色锐化方法  

The panchromatic sharpening method based on multi-scaleresidual attention network

作　　者：吴燕燕 王亚杰 谢延延 WU Yanyan;WANG Yajie;XIE Yanyan(Engineering Training Center,Shenyang Aerospace University,Shenyang 110136,China)

机构地区：[1]沈阳航空航天大学工程训练中心,沈阳110136

出　　处：《重庆理工大学学报（自然科学）》2023年第8期222-230,共9页Journal of Chongqing University of Technology：Natural Science

基　　金：辽宁省兴辽英才计划项目(XLYC1906003);辽宁省教育厅系列项目(LJKMZ20220582)。

摘　　要：针对传统深度学习的遥感图像全色锐化方法未考虑源图像多尺度方向信息和通道间的关联紧密性,导致融合图像出现空间信息丢失和光谱失真问题,提出了一种基于多尺度残差注意力网络的遥感图像全色锐化方法。将低空间分辨率多光谱(low-resolution multi-spectral, LRMS)图像经过双三次插值上采样与高空间分辨率全色(panchromatic, PAN)图像级联,得到一个5通道图像作为输入;设计3个并行的多尺度残差注意力网络从空间和通道两方面提取源图像不同频度的特征,每个子网络包含浅层特征提取、深层特征提取、特征融合、特征重建过程,将三者的输出进行跳连接获得最终融合图像,将光谱角映射、平均绝对误差和几何梯度权重相加作为一种新的损失函数来训练参数,以提高网络的信息表征能力;选择在WorldView-2和WorldView-3数据集上与其他7种融合方法进行对比实验,并在WorldView-3数据集上对模型的网络结构及损失函数性能进行验证,实验结果表明:该方法在光谱信息和空间信息保持方面具有明显优势。According to the panchromatic sharpening method for remote sensing image fails to consider the multi-scale directional information and the correlation between channels of the source image,which will lead to loss of spatial information and spectral distortion.A new panchromatic sharpening method of remote sensing image based on multi-scale residual attention network is proposed.Firstly,low-resolutionmulti-spectralimages after bicubic interpolation upsampling were cascaded with high-resolution panchromatic images,then the 5-channel image obtained serves as the input.Secondly,three parallel multi-scale residual attention networks are designed to different extract features of source images from spatial and channel.Each network includes shallow feature extraction,deep feature extraction,feature fusion and feature reconstruction processes,andoutputs of three networks are jump-connected to obtain the final fusion image.Spectral angle mappper,mean absolute error and geometric gradient are added with weights as a new loss function to train parameters,improving the information representation ability of the network.At last,WorldView-2 and WorldView-3 are selected to perform comparative experiments with other 7 fusion methods,and the performance of network structure and loss function of the model are verified on WorldView-3 data set.The experimental results show that this method has significant advantages in the preservation of spectral information and spatial information.

关 键 词：多光谱图像 全色图像 全色锐化 多尺度残差注意力网络 损失函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]