面向高分辨率多视角SAR图像的端到端配准算法  

End-to-end Registration Algorithm for High-resolution Multi-view SAR Images

作　　者：孙晓坤 贠泽楷 胡粲彬 项德良 SUN Xiaokun;YUN Zekai;HU Canbin;XIANG Deliang(College of Information Science and Technology,Beijing University of Chemical Technology,Beijing 100029,China)

机构地区：[1]北京化工大学信息科学与技术学院,北京100029

出　　处：《雷达学报(中英文)》2025年第2期389-404,共16页Journal of Radars

基　　金：中央高校基本科研业务费专项资金(buctrc202218);中央高校基本科研业务费专项资金自由探索项目(ZY2413)。

摘　　要：由于侧视和相干成像机制,当高分辨率合成孔径雷达(SAR)图像的成像视角变化较大时,图像间的特征差异会变大,使图像配准难度增加。针对高分辨率多视角SAR图像,传统的配准技术主要面临提取的关键点定位精度不足和匹配精度低的问题。基于上述难点,该文设计了一种端到端的高分辨率多视角SAR图像配准网络。文章主要贡献包括:提出基于局部像素偏移模型的高分辨率SAR图像特征提取方法,该方法提出多样性峰值损失监督训练关键点提取网络中响应权重分配部分,并通过检测像素偏移量来优化关键点坐标;提出基于自适应调整卷积核采样位置的描述符提取方法,利用稀疏交叉熵损失监督训练网络中描述符匹配。实验结果显示,相比于其他配准方法,该文提出的算法针对高分辨率多视角SAR图像配准效果显著,平均误差降低超过65%,正确匹配点对数提高了3~5倍,运行时间平均缩短50%以上。Due to the side-looking and coherent imaging mechanisms,feature differences between high-resolution Synthetic Aperture Radar(SAR)images increase when the imaging viewpoint changes considerably,making image registration highly challenging.Traditional registration techniques for high-resolution multi-view SAR images mainly face issues,such as insufficient keypoint localization accuracy and low matching precision.This work designs an end-to-end high-resolution multi-view SAR image registration network to address the above challenges.The main contributions of this study include the following:A high-resolution SAR image feature extraction method based on a local pixel offset model is proposed.This method introduces a diversity peak loss to guide response weight allocation in the keypoint extraction network and optimizes keypoint coordinates by detecting pixel offsets.A descriptor extraction method is developed based on adaptive adjustment of convolution kernel sampling positions that utilizes sparse cross-entropy loss to supervise descriptor matching in the network.Experimental results show that compared with other registration methods,the proposed algorithm achieves substantial improvements in the high-resolution adjustment of convolution kernel sampling positions,which utilize sparse cross-entropy loss to supervise descriptor matching in the network.Experimental results illustrate that compared with other registration methods,the proposed algorithm achieves remarkable improvements in high-resolution multi-view SAR image registration,with an average error reduction of over 65%,3~5-fold increases in the number of correctly matched point pairs,and an average reduction of over 50% in runtime the network.Experimental results show that compared with other registration methods,the proposed algorithm achieves substantial improvements in the high-resolution adjustment of convolution kernel sampling positions,which utilize sparse cross-entropy loss to supervise descriptor matching in the network.Experimental results illustrat

关 键 词：合成孔径雷达 遥感图像配准 特征描述符提取 旋转鲁棒性 像素偏移量 

分 类 号：TP75[自动化与计算机技术—检测技术与自动化装置]