【作者】 郭雷；

【导师】 郭桂蓉； 付强；

【作者基本信息】 国防科学技术大学 ， 信息与通信工程， 2009， 博士

【摘要】 雷达目标特征提取与分类器设计是雷达目标识别系统中的两个关键问题。本文基于宽带高分辨全极化雷达体制,以飞机、坦克、舰船等军事目标为识别对象,围绕目标极化特征提取、优选、核方法分类器模型参数优化、核方法分类器设计这几个方面展开深入研究,以求提取反映目标本质属性的特征、设计目标识别最优分类器,提高雷达目标识别性能。研究内容主要分为三大部分:1.雷达目标HRRP极化特征提取与优选(1)从三个不同角度研究目标高分辨距离像(High Resolution Range Profile,HRRP)极化特征提取:①将极化合成孔径雷达(Synthetic Aperture Radar,SAR)中常用的H /α分解方法引入到全极化目标HRRP,丰富了反映目标HRRP散射随机性的特征;②根据Sinclair散射矩阵相似性参数理论,定义了目标HRRP与6种标准体散射矩阵相似性参数的概率形式,以其为基础构建的特征可准确反映目标的物理结构特性;③提出了反映目标散射能量特性的Mueller矩阵相似性参数特征,并且证明了该参数特征的旋转不变性。(2)基于全极化与双极化体制,分别详细推导了这些特征在两种极化体制下的表达式。结果表明,两种体制下不仅这些特征的表达式不同,而且特征的个数也不同,有些特征之间存在一定的线性关系。实验部分采用飞机目标实测数据和舰船目标电磁特性软件计算数据从特征可分性以及识别性能两个方面验证以上特征的有效性。通过特征优选,分别为飞机、舰船目标提取有利于识别的极化特征。2.雷达目标识别中SVM可分性研究与模型优化选择(1)从理论上研究了支持矢量机(Support Vector Machine,SVM)线性可分性的本质,以及SVM线性不可分时引入惩罚因子c的物理意义,并且证明过程更简单,物理意义更明确。通过实验定量分析了核函数及其参数对识别性能的影响。(2)从理论上解释了SVM模型最优选择的本质,分析了SVM模型单参数优化选择的方法,指出了单参数模型对于识别的局限性,提出进行模型多参数选择的必要性,并且提出了非均衡数据识别中SVM模型多参数最优选择方法。实验部分采用国际通用的标准数据库和雷达目标极化特征作为输入,验证了优选后多参数模型能够有效提高目标识别性能。3.宽带雷达目标识别核方法分类器设计与核矩阵构造(1)在目标可分性比较差的情况下,标准的支持矢量数据描述(Support Vector Data Description,SVDD)判决方法会使雷达目标识别出现拒分与错分问题。针对这个问题,提出了SVDD多目标模糊识别方法。多个极化状态下雷达目标HRRP识别实验结果表明,该方法简单实用,能够有效克服标准判决方法带来的错分与拒分问题。(2)目标可分性比较差同样会使得SVM识别方法带来错分与拒分问题,针对这个问题提出了模糊SVM(FSVM)方法的判决改进策略,给出了简单实用的新判决策略。实验结果验证了该判决策略能够有效的提高目标识别性能。(3)针对目标样本在高维特征空间中不能线性分类这一问题,提出了特征空间中数据核矩阵收缩新方法。主要思路是改变数据在特征空间的结构,使得线性不可分的数据变得线性可分,以此提高目标识别率。该项研究内容首先从理论上证明了收缩后核矩阵性能更优,然后推导了收缩后核矩阵的表示形式。采用飞机、坦克HRRP数据的识别实验验证了新方法的有效性。更多还原

【Abstract】 Radar target feature extraction and classifier design are key problems in radar target recognition system. In order to extract the features reflecting target’s characteristic, construct the optimal classifier and improve radar target recognition performance, this dissertation regards the plane, tank and ship as the recognizing objects, focuses on polarimetric feature extraction, optimal selection, kernel methods classifiers model optimization and design based on wideband high-resolution fully polarimetric radar. The whole work contains three parts:1. Radar Target HRRP Polarimetric Feature Extraction and Optimal Selection(1) Polarimetric features extraction and optimal selection in radar target HRRP are mainly researched in three different aspects:①the entropy is introduced into fully polarimetric HRRP, which is usually used in polarimetric SAR, then the entropy feature based on HRRP is proposed, these parameters reflect the HRRP’s degree of scattering randomness;②according to similarity theory between two Sinclair scattering matrix, six probability description features are defined between HRRPs and some standard objects, the parameters reflect the target’s physics structure;③the similarity feature based on Mueller matrix is proposed, which reflects the target power characteristic. It is also proved that this feature does not vary with the orientation angle.(2) The formulas of the three kinds of polarimetric feature are deduced in fully and dual polarimetric radar. The results revel that every feature formula is different in fully and dual polarimetric radar, some of them are correlative linearly. To verify the features’utility, some experiments had been done in feature separability and recognition performance by using some experimental data on planes and ships. The experimental results present that the features of HRRP proposed in dissertation are robust and well separability.2. The Study on the Separability of SVM and Model Optimal Selection Used in Radar Target Recognition(1) The sufficient necessary condition of linear separability is proved using a brief and clear method, and also proved and explained the new essential of penalty factor in SVM. The effects on the recognition performance of kernel function and its parameters are quantitatively analyzed.(2) The essential of model optimal selection is explained and the method of model single parameter selection is analyzed. Some limitations of single parameter model are presented, and then the connotation and necessity of SVM model multi-parameter selection are theoretically analyzed. SVM model optimal multi-parameter selection method for imbalanced data recognition is proposed. Some experiments on Benchmarks and radar target polarimetric features verify that the new method can improve the recognition performance effectively.3. Kernel Method Classifier Design for Radar Target Recognition and Kernel Matrix Construction(1) Some misclassification and reject classification problems caused by the standard SVDD decision strategy always occur especially in lack of separability of radar target, so radar target fuzzy recognition method is proposed based on SVDD. Experimental results on the fully polarimetric HRRP data present that the fuzzy recognition method can achieve better accuracy than standard SVDD.(2) Lack of separability of radar target can also cause misclassification and reject classification problems using SVM. To deal with these problems, the improved decision strategy is proposed based on FSVM. Simulations verify the new method.(3) To deal with the linear non-separability, the new conception and method of kernel matrix contraction are proposed. The main idea of the new method is to make the non-separable data separability by changing the construction of data in feature space. The kernel matrix of contracted pattern is deduced, the separability of contracted pattern is also proved more excellent than before. The experiments conducted to evaluate the performance of the new method are mainly in two aspects: the measurement and classifier performance of kernel matrix. The results on planes and ships HRRP data reveal that the performance of contracted kernel matrix is more excellent than before.更多还原