【作者】 刘华林；

【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， 2008， 博士

【摘要】 现代高分辨雷达的兴起为目标识别提供了新的途径。高分辨距离像反映了目标沿雷达径向的几何结构分布,较之于二维或三维成像,不仅获取要容易得多,而且避免了成像过程中复杂的运动补偿问题。因此,近年来高分辨雷达目标识别受到了业界的广泛关注。本文在前人工作的基础上,着重对高分辨距离像雷达目标识别系统中特征提取与分类识别两个环节作了较深入的研究,并提出了一些创新的算法。这些算法在多组仿真与实测目标距离像数据的基础上进行了验证。归纳起来,本文的主要内容包括以下几个方面:1.结合扰动法和零空间法,分别提出了基于QR分解的线性辨别分析和直接线性辨别分析雷达目标距离像识别方法,并利用核机器学习理论分别对其进行了非线性推广。实验结果表明,基于QR分解的辨别分析在实时性能上具有明显的优势,而直接辨别分析则具有良好的识别性能。2.针对传统Gram-Schmidt正交化算法敏感于舍入误差的不足,首先提出了核修正Gram- Schmidt正交化算法,然后以此为基础发展了批处理式和类增量式两种核辨别分析雷达目标距离像识别方法。新方法充分利用了类内散布矩阵最具辨别力的零空间信息,具有良好的识别性能。尤其类增量式的核辨别分析在有新目标数据嵌入训练样本集时可以动态刷新特征矢量,有效地避免了将所有目标数据同时调入内存,造成计算负担过重的问题,具有明显的实时性能优势。3.在模式识别理论中,特征提取的一般原则是希望所提取的目标特征之间统计相关性越小越好,最好是不相关的。依据这一理论,提出了一种基于核不相关辨别分析的雷达目标距离像识别框架,其不相关最优辨别矢量集可以通过联合对角化或广义奇异值分解方式求解。由于去除了模式样本特征之间的冗余信息,新方法体现了良好的识别性能。4.针对经典辨别分析中可能存在的矩阵奇异问题,首先依据Fisher准则导出了距离像总散布矩阵的零空间中不含有有用辨别信息的结论。利用这一结论,可以对各散布矩阵进行预降维,以减小后续运算的计算复杂度。然后从全局角度出发,提出了一种双辨别子空间雷达目标距离像识别方法。该方法充分利用了类内散布矩阵零空间和非零空间中所包含的有用辨别信息,获得了良好的识别性能。5.在经典最近特征线和最近特征平面分类器的基础上,利用核机器学习理论分别将其推广为核非线性分类器,使两者无需经过特征提取即可以直接对原始距离像样本进行分类。同时,针对这些分类器在大数据样本量与高维数时计算量大,且有可能失效的问题,基于局部最近邻准则提出了改进的分类方法,使其在保持较高识别率的同时,显著提高了分类的实时性能。更多还原

【Abstract】 The increasing availability of high resolution range (HRR) radars provides a new way for radar target recognition. High resolution range profile (HRRP) shows the target’s scatterers distribution along the radar line-of-sight, which contains potentially discriminative information about the target geometry. Furthermore, the HRRP can be easily captured and avoids the complex motion compensation processing, relative to two-dimensional or three-dimensional imagery. Therefore, HRR radar target recognition has received extensive attention from the radar technique community in recent years.Based on the previous work, this dissertation is focused on the feature extraction and classification of a radar target recognition system using HRRP. Some new methods are presented, and all of them are evaluated on both simulated and measured data of aircrafts.The main content is summarized as follows:1. According to the perturbation theory and null-space method, two feature extraction methods for radar HRRP recognition are proposed respectively. One is QR decomposition based linear discriminant analysis (LDA), the other is direct LDA. Meanwhile, both methods are generalized to nonlinear versions via kernel trick. The experimental comparisons show that QR decomposition based methods have great advantage in terms of real-time performance, while another two achieve excellent recognition performance.2. The classical Gram-Schmidt (GS) orthogonalization procedure is very sensitive to round-off-errors. Thereby, a modified GS orthogonalization procedure using kernel function operator (KMGS) is first proposed. Then two nonlinear algorithms, batch and class-incremental kernel discriminant analysis (KDA), are put forward for radar HRRP recognition. Compared with other kernel-based methods, batch KDA and class-incremental KDA both achieve good recognition performance for making use of the significant discriminative information in the null space of within-class scatter matrix. Moreover, class-incremental KDA introduces an incremental approach to update the discriminant vectors when new target data sets are inserted into the training set, which is very desirable for designing a dynamic recognition system. Therefore, it has apparent advantage in real-time performance.3. In pattern analysis, the common principle of feature extraction is desirable to extract feature vectors with uncorrelated attributes. Motivated by this principle, a new formulation for KDA is proposed for radar HRRP recognition, which can solve the uncorrelated discriminant vectors by joint diagnonalization and GSVD respectively. The methods both achieve good recognition performance for removing the redundancy among feature vectors extracted.4. It is well known that classical Fisher discriminant analysis algorithms suffer from singularity problem and lose some significant discriminative information. To address this problem, one conclusion that there exists no useful discriminative information in the null space of the population scatter matrix is first derived, which can be used to reduce the dimensionality of original scatter matrices as well as the computation complexity of the following operation. Then a double discriminant subspaces algorithm for radar HRRP recognition is proposed. The new method considers the separability from a global viewpoint to some extent, which can make full use of the discriminative information in both null space and non-null space of within-class scatter matrix. Therefore, it makes the new method a more powerful discriminator.5. Kernel nonlinear classifiers from classical nearest feature line and nearest feature plane are proposed for radar HRRP classification, which can directly classify original range profiles and need no feature extraction beforehand. Meanwhile, these classifiers are modified based on locally nearest neighborhood rule. Compared with those original ones, the modified classifiers achieve competitive performance and take much lower computation cost, while the probability of failure is reduced to some extent.更多还原