基于编解码网络的SAR影像建筑物提取  

Building Extraction from SAR Images Based on Encoder-decoder Network

作　　者：苗国英[1] 王慧琴 张恩伟 MIAO Guoying;WANG Huiqin;ZHANG Enwei(School of Automation,Nanjing University of Information Science and Technology,Nanjing 210044,China)

机构地区：[1]南京信息工程大学自动化学院,南京210044

出　　处：《遥感信息》2024年第2期127-134,共8页Remote Sensing Information

基　　金：国家自然科学基金(62073169);江苏省“333工程”项目(BRA2020067)。

摘　　要：针对合成孔径雷达(synthetic aperture radar,SAR)遥感图像中出现的散斑噪声和几何畸变对建筑物提取过程造成干扰的问题,提出了一种多尺度特征注意力融合(multi-scale feature attention fusion,MSFAF)网络。首先,结合深度神经网络和SAR图像的优势,在深层处设计了一个空间注意力融合(spatial attention fusion,SAF)模块,来整合不同层次的特征以及关注重要的空间信息。然后,利用不同尺度的卷积核以及对通道信息的转换,提出了一个多尺度细节提取(multi-scale detail extraction,MSDE)模块用于提取不同尺度的特征信息和重新分配通道信息,有利于缓解散斑噪声的干扰问题。实验证明了所提方法在SAR图像建筑物提取中取得了比其他现存方法更加优秀的性能。There are always speckle noise and geometric distortion appearing in synthetic aperture radar(SAR)remote sensing images,which interferes with the building extraction process and results in unclear building boundaries.In view of the above problems,the paper proposes a multi-scale feature attention fusion(MSFAF)network.Firstly,combining the advantages of deep neural networks and SAR images,a spatial attention fusion(SAF)module is designed in the deep layer,where it integrates different levels of features and focuses on important spatial information.Moreover,by applying convolution kernels of different scales and the conversion of channel information,a multi-scale detail extraction(MSDE)module is given to extract feature information of different scales and redistribute channel information,which is beneficial to alleviate the interference problem of speckle noise.The experimental results show that the proposed method has better performance than other existing methods in SAR image building extraction.

关 键 词：SAR遥感图像 建筑物提取 多尺度特征 注意力融合 神经网络 

分 类 号：P236[天文地球—摄影测量与遥感]