地面目标多角度SAR数据集构建与目标识别方法  

SAR Database Construction for Ground Targets at Multiple Angles and Target Recognition Method

作　　者：朱岱寅[1] 耿哲 俞翔[2] 韩胜亮 杨卫星 吕吉明 叶铮 闫贺[1] ZHU Daiyin;GENG Zhe;YU Xiang;HAN Shengliang;YANG Weixing;LYU Jiming;YE Zheng;YAN He(College of Electronic and Information Engineering,Nanjing University of Aeronautics&Astronautics,Nanjing 211106,China;School of Computer Engineering,Nanjing Institute of Technology,Nanjing 211167,China)

机构地区：[1]南京航空航天大学电子信息工程学院,南京211106 [2]南京工程学院计算机工程学院,南京211167

出　　处：《南京航空航天大学学报》2022年第5期985-994,共10页Journal of Nanjing University of Aeronautics & Astronautics

基　　金：国家自然科学基金(62271252);江苏省自然科学基金青年基金(BK20200420)。

摘　　要：南京航空航天大学雷达探测与成像技术研究团队利用自主研制的无人机(Unmannedaerialvehick,UAV)机载高分辨率微小型合成孔径雷达(Minisyntheticapertureradar,MiniSAR)系统,针对多类具有代表性的地面目标进行全方位回波录取及成像处理,构建了拥有自主知识产权的复杂目标SAR数据集,并依托该数据集开展了基于人工智能的目标识别方法研究。针对无人机运动姿态不稳定、辅助传感器精度受限导致的图像散焦问题,本文提出了新型运动补偿及新型二维自聚焦算法。实验表明,虽然AlexNet、ResNet‑18、AConvNet和VGG等经典神经网络在MSTAR十类目标分类问题中取得了接近100%的分类准确率,但将其应用于南航MiniSAR数据集时分类准确率均明显低于90%。由于本文采取的实验方法与SAR目标识别技术的实际应用场景较为接近,该MiniSAR数据集对于面向工程应用的SAR目标识别算法研究将会具有重要参考价值。By using the self-developed unmanned aerial vehicle(UAV)borne mini synthetic aperture radar(MiniSAR)system,all-directional echoes for multiple types of representative ground targets are collected and used for image processing by the Radar Detection and Imaging Techniques Research Group of Nanjing University of Aeronautics&Astronantics(NUAA).A proprietary SAR database for complex targets is constructed,based on which artificial-intelligence-inspired target recognition approaches are studied.To solve the challenging issue of SAR image defocus caused by unstable movement of the UAV platform and the limited accuracy provided by the accessory sensors,novel movement compensation and auto-focus processing methods are proposed.The impact of the image defocus on the classification accuracies provided by the neural networks is revealed,and the limited generalizability of the existing neural networks is demonstrated.Simulation results show that although the classic neural networks,such as AlexNet,ResNet-18,AConvNet,and VGG offer a near-100%accuracy for the MSTAR 10-target classification problem,the 9-class target classification accuracies provided by these networks for the MiniSAR dataset are all much lower than 90%.Since the experiment method employed in this paper closely resembles the practical application scenario,the proposed database will be of great reference value to the development of SAR target recognition algorithms for engineering applications.

关 键 词：合成孔径雷达 目标识别 微小型SAR 目标数据集 人工智能 

分 类 号：V262[航空宇航科学与技术—航空宇航制造工程]