结合注意力机制的卫星遥感影像立体匹配  

Stereo matching of satellite remote sensing images based on attention mechanism

作　　者：张泽瑞 范大昭[1] 纪松[1] 董杨 李东子 刘杰 ZHANG Zerui;FAN Dazhao;JI Song;DONG Yang;LI Dongzi;LIU Jie(Institute of Geospatial Space Information,Information Engineering University,Zhengzhou 450002,China;Powerchina Beijing Engineering Corporation Limtied,Bejing 100024,China)

机构地区：[1]信息工程大学地理空间信息学院,郑州450002 [2]中国电建集团北京勘测设计研究院有限公司,北京100024

出　　处：《时空信息学报》2024年第1期41-49,共9页JOURNAL OF SPATIO-TEMPORAL INFORMATION

基　　金：国家自然科学基金(42371459);嵩山实验室项目(221100211000-4);高分遥感测绘应用示范系统(二期)(42-Y30B04-9001-19/21)。

摘　　要：采用立体匹配技术对多视卫星遥感影像进行三维场景重建一直是摄影测量与遥感领域的核心问题。基于卷积神经网络的深度学习方法极大地促进了立体匹配技术的发展,然而其中涉及匹配困难和误匹配问题的相关研究仍然不足。为了提升卫星遥感影像不适定区域中视差估计的精度,本研究提出了一种结合注意力机制的立体匹配深度学习网络,在特征提取模块中加入注意力机制,分别从通道和空间两个维度捕获全局信息,对特征进行优化;在代价体的构建模块中构建新的代价体积,并重新设置视差的回归范围。为了验证本文方法的有效性,在US3D、WHU-Stereo两个数据集上分别与已有方法Stereo-Net、PSM-Net进行了比较分析。结果表明,本文方法在EPE(endpointerror)和D_1两个指标上均能达到最优,取得了较好的性能,提高了立体匹配的精度,尤其在无纹理、重复纹理、遮挡及视差不连续区域表现出良好的鲁棒性。In recent years,deep learning techniques based on convolutional neural networks have greatly promoted the development of disparity estimation.These methods have received widespread attention,and stereo matching,a crucial step in 3D reconstruction from satellite image,still encounters considerable challenges due to its high resolution and complex structure.Especially in areas with no texture,repeated textures,discontinuous disparity,and occlusion,stereo matching becomes challenging.Stereo matching is a technique that searches for the corresponding relationship between pixels from images captured from different perspectives.Parallax refers to the horizontal shift of corresponding points(left image(x,y),right image(x–d,y))calculated from a pair of corrected stereo image pairs.The three-dimensional scene reconstruction of multi-view aerospace remote sensing images through stereo matching remains a core issue in photogrammetry and remote sensing.With the successful launch and networking of China’s Resource 3 series of stereo satellites,stereo matching of high-resolution satellite images and the generation of digital terrain models(DSM)have become a research hotspot.For stereo matching of satellite images,mainstream network models rely on a large amount of training data from ground truth labels used for parameter learning.However,the scarcity of satellite image stereo matching datasets significantly impedes research and the application of deep learning techniques in this field.Although the resolution of domestic satellite images has been greatly improved,it is still difficult to restore the structural information of the images due to the fact that satellite images are collected at different times and have a wide range of perspectives.Secondly,the presence of large buildings and towering structures can lead to occlusion and discontinuous parallax,as well as the presence of rivers,water bodies,and vegetation.There are also situations where there are no textures or repetitive patterns,and few network models can fully

关 键 词：立体匹配 卷积神经网络 卫星遥感影像 注意力机制 深度学习 视差估计 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]