【作者】 付建胜；

【导师】 杨万麟；

【作者基本信息】 电子科技大学 ， 信息获取与探测技术， 2012， 博士

【摘要】 现代雷达技术的兴起为雷达目标识别提供了强有力的技术支撑。通过宽带雷达获取的高分辨距离像反映了目标散射点沿雷达视线的径向距离分布细节，不仅提供了比低分辨雷达目标回波更多的特征信息，而且还避免了二维或三维成像过程中复杂的运动补偿和过多的成像耗时，具有易于获取和处理的独特优势。因此，近年来基于高分辨距离像的目标识别受到了雷达科技工作者的广泛关注。本文主要集中于特征提取和分类识别这两个环节，层层深入地展开对雷达多目标距离像识别理论及技术的相关研究。主要工作和创新概括如下：（1）针对经典核方法中大散布矩阵特征分解的运算压力问题，提出了两种基于扩展二分类辨别单元的核方法。这两种核方法均采用了“一对一”策略，通过拆分和重组，将大散布矩阵划分成若干小散布矩阵，并分别按串联和并联方式进行组合。实验结果表明，这两种核方法均能有效降低训练耗时，并能改善识别性能，非常适用于多目标识别。（2）针对多目标识别中辨别信息的质量和数量孰重孰轻问题，设计出三种辨别信息提取模型，即：被动识别和总体挑选类中绝对辨别信息（Passive recognitionand General selection for Among-class absolute discriminant information, PGA）模型、被动识别和个体挑选类间相对辨别信息（Passive recognition and Individual selectionfor Between-class relative discriminant information, PIB）模型以及主动识别和个体挑选类间相对辨别信息（Active recognition and Individual selection for Between-classdiscriminant information, AIB）模型。理论分析表明：PGA模型侧重于辨别信息的质量，而PIB和AIB模型侧重于辨别信息的数量。（3）将广义辨别分析（Generlized DiscriminantAnalysis, GDA）应用于PIB和AIB模型中，由此产生了两种针对辨别信息数量的核方法，即：基于PIB模型的GDA（PIB-based GDA, PIB-GDA）和基于AIB模型的GDA（AIB-based GDA,AIB-GDA）。与GDA比较，这两种核方法不仅能大幅降低训练耗时，还能提取更多辨别信息，具有良好的识别性能。（4）对多智能技术进行了归纳和总结，并设计出一种多智能体识别模型。将GDA用于模型实现，由此产生了一种PIB-GDA与GDA并联组合的新方法，即综合GDA（Synthetic-GDA, S-GDA）算法。实验结果表明，S-GDA能实现PIB-GDA与GDA在识别性能上的优势互补。（5）依据三种基本组合结构，本文提出了四种组合核方法。它们是，1.并联结构：基于核Fisher辨别（Kernel Fisher Discriminant, KFD）的多类综合辨别分析（KFD-based Multiclass Synthetical DiscriminantAnalysis, KFD-MSDA）和全局分布式KFD（Global Distributed KFD, G-DKFD）；2.串联结构：基于多KFD的LDA（Multi KFD-based LDA, MKFD-LDA）；3.混合结构：核混合辨别分析（KernelMixed Discriminant Analysis, KMDA）。实验结果表明，这四种组合核方法的识别性能从高到低依次为KFD-MSDA、G-DKFD、KMDA、MKFD-LDA，其中KMDA与GDA的识别性能相当。更多还原

【Abstract】 The rise of modern radar technology provides a strong technical support for radartarget recognition. The high resolution range profile (HRRP) obtained by widebandradar shows the radial distance distribution details of target scattering centers along theradar line of sight, and contains more structure information than that of the target echoobtained by low resolution radar. Furthermore, the HRRP can be easily captured andprocessed, while potentially avoiding the complex motion compensation processing andtoo much imaging time-consuming, relative to the two or three dimensional imagery.Therefore, the target recognition based on HRRP has received extensive attention fromthe radar technology workers in recent years.Focused on the feature extraction and classification subprocesses, this dissertationprogressively deepens the research upon the theory and technology of radar multi-targetrecognition using HRRP. The main content and innovation are summarized as follows:(1) To reduce the eigen-decomposition operation burden of big scatter matrixes inclassical kernel methods, two new kernel-based methods are proposed by expanding thetwo-class discriminant units. By splitting and restructuring under the so called “oneagainst one” strategy, the two new methods both can divide a big scatter matrix into aseries of small ones, and then arrange the small ones by series or parallel. Theexperimental results show that the two new methods both can reduce the trainingtime-consuming effectively, improve the recognition performance, and are very suitablefor radar multi-target recognition.(2) In multi-target recognition, the quality and quantity of discriminant information(DI), which one is more important? Accompanied with this issue, three DI extractionmodels, i.e., passive recognition and general selection for among-class absolutediscriminant information (PGA) model, passive recognition and individual selection forbetween-class relative discriminant information (PIB) model, and active recognitionand individual selection for between-class discriminant information (AIB) model, aredesigned. Theoretical analyses indicate that the PGA model prefers to the DI qualitywhile the PIB and AIB models both prefer to the DI quantity. (3) Generlized discriminant analysis (GDA) is applied in the PIB and AIB models,and then two kernel-based methods, i.e., PIB-based GDA (PIB-GDA) and AIB-basedGDA (AIB-GDA), come forth for the DI quantity. Compared with GDA, PIB-GDA andAIB-GDA can not only reduce the training time-consuming greatly, but also improvethe recognition performance perfectly.(4) A summary and conclusion is given to the multi-agent technology, and then amulti-agent model is designed for radar HRRP target recognition. Also GDA is appliedfor model actualization, and then a new method, synthetic-GDA (S-GDA) algorithm,comes forth, which can be considered as the parallel combination of PIB-GDA andGDA. The experimental results indicate that S-GDA can realize the advantagecomplementary of PIB-GDA and GDA with respect to the recognition performance.(5) According to the three fundamental composite constructions, this dissertationproposes four composite kernel methods, i.e., parallel construction: kernel Fisherdiscriminant (KFD)-based multiclass synthetical discriminant analysis (KFD-MSDA)and gobal distributed KFD (G-DKFD), series construction: multi-KFD-based lineardiscriminant analysis (MKFD-LDA), mixed construction: kernel mixed discriminantanalysis (KMDA). The experimental results show that the recognition performance ofthe four composite kernel methods is labeled from high to low by KFD-MSDA、G-DKFD、KMDA and MKFD-LDA, among which the the recognition performance ofKMDA is very similar to that of GDA.更多还原