·雷达系统与技术·DOI:10．16592/j．cnki．1004－7859．2022．12．016基于端到端网络机制跨域稀疏SAＲ与光学图像精准匹配方法黄柏圣(南京信息工程大学电子与信息工程学院，南京210044)摘要:由于合成孔径雷达(SAＲ)和光学图像两种模态之间存在显著的几何和辐射差异，传统方法难以实现SAＲ与光学图像的精准匹配。文中提出了一种基于端到端网络机制的跨域稀疏SAＲ与光学图像精准匹配方法。该方法首先预测每幅图像中最适合匹配的区域，然后通过多尺度特征空间互相关运算生成匹配热图，最后将匹配热图分为正匹配和负匹配来消除异常值，实现SAＲ与光学图像匹配的精确匹配。实验结果表明，所提方法性能指标优于已往SAＲ与光学图像匹配方法，可用于大规模场景的精确匹配，利于提升光学卫星图像的地理定位精度。关键词:图像配准;特征检测;深度学习;合成孔径雷达;光学图像中图分类号:TN957文献标志码:A文章编号:1004－7859(2022)12－00106－06引用格式:黄柏圣．基于端到端网络机制跨域稀疏SAＲ与光学图像精准匹配方法［J］．现代雷达，2022，44(12):106－111．HUANGBaisheng．Accuratematchingmethodofcross-domainsparseSAＲandopticalimagebasedonend-to-enddeeplearningframework［J］．ModernＲadar，2022，44(12):106－111．AccurateMatchingMethodofCross-domainSparseSAＲandOpticalImageBasedonEnd-to-endDeepLearningFrameworkHUANGBaisheng(SchoolofElectronicsandInformationEngineering，NanjingUniversityofInformationScienceandTechnology，Nanjing210044，China)Abstract:Duetothesignificantgeometricandradiationdifferencesbetweensyntheticapertureradar(SAＲ)andopticalimages，itisdifficultfortraditionalmethodstoachieveaccuratematchingbetweenSAＲandopticalimages．Inthispaper，anaccuratematchingmethodbetweencrossdomainsparseSAＲandopticalimagebasedonend-to-endnetworkframeworkisproposed．Themethodfirstpredictstheregionthatisconsideredtobethemostsuitableformatchingineachimage，andthengeneratesthematchingheatmapthroughmulti-scalefeaturespacecross-correlationoperator．Finally，thematchingheatmapisclassifiedintopositiveandnegativematchingtoeliminateoutliersandachieveaccuratematchingbetweenSAＲandopticalimages．Theexperimentalresultsshowthatthattheproposedapproachprovidesasubstantial...
