基于残差伪孪生卷积互相关网络的异源遥感图像匹配算法  

Heterogeneous Remote Sensing Image Matching Algorithm Based on Residual Pseudo-Siamese Convolution Cross-Correlation Network

作　　者：邹融平 朱斌 王晨阳[1] 朱耀轩 胡洋頔 Zou Rongping;Zhu Bin;Wang Chenyang;Zhu Yaoxuan;Hu Yangdi(College of Electronic Engineering,National University of Defense Technology,Hefei 230037,Anhui,China;Key Laboratory of Infrared and Low Temperature Plasma of Anhui Province,Hefei 230037,Anhui,China;Army 32256 of PLA,Guiling 541000,Guangxi,China)

机构地区：[1]国防科技大学电子对抗学院,安徽合肥230037 [2]红外与低温等离子体安徽省重点实验室,安徽合肥230037 [3]中国人民解放军32256部队,广西桂林541000

出　　处：《激光与光电子学进展》2022年第12期448-456,共9页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(61307025)。

摘　　要：合成孔径雷达(SAR)成像和可见光成像是遥感卫星的常用成像方式。由于两者在成像信息上具有高度互补性,SAR和可见光的图像数据融合已成为了遥感的一个重要研究领域。异源数据匹配算法的性能直接影响获取地面控制点的精度,匹配算法分为二阶段法和一阶段法,现有的二阶段法难以适应地形复杂的遥感图像,且在速度上无法满足实际的工程需求,而速度满足要求的一阶段法在精度上仍有所欠缺。为解决这个问题,提出了一个可端到端的高精度的基于残差伪孪生卷积互相关网络的异源遥感图像匹配算法。所提算法通过构建基于残差层的伪孪生网络,对提取的SAR和可见光图像的特征进行卷积互相关操作从而实现异源遥感图像匹配。实验结果表明,该算法在保持较高的速度下,较大提升了SAR与可见光图像的匹配精度,为深度学习方法在大规模异源遥感图像匹配任务中的工程应用奠定了基础。Remote sensing satellites commonly use synthetic aperture radar(SAR) and visible light imaging. SAR and visible image data fusion have become an important research field of remote sensing owing to their high complementarity in imaging information. The accuracy of obtaining ground control points is directly influenced by the performance of a heterogeneous data matching algorithm. There are two methods of matching algorithms: twostage and one-stage. The existing two-stage method is difficult to adapt to remote sensing images with complex terrain and it cannot meet the actual engineering needs in terms of speed, while the one-stage method meets the requirements in terms of speed but lacks in accuracy. To solve this problem, an end-to-end high-precision heterologous remote sensing image matching algorithm based on a residual pseudo-twin convolution cross-correlation network has been proposed. By constructing a pseudo twin network based on residual layer, the proposed algorithm performs convolution cross-correlation operation on the extracted features of SAR and visible images, so as to realize heterogeneous remote sensing image matching. The results show that this algorithm considerably improves the matching accuracy between SAR and visible images, maintaining a high speed and laying the foundation for the engineering applications of depth learning methods in large-scale heterogeneous remote sensing image matching tasks.

关 键 词：遥感 合成孔径雷达 可见光 异源匹配 端到端 

分 类 号：P407.8[天文地球—大气科学及气象学]