基于高分辨距离像的特征提取与识别增强技术研究

发布时间：2018-05-01 05:27

  本文选题：高分辨距离像 + 特征提取　； 参考：《国防科学技术大学》2016年博士论文

【摘要】：高分辨距离像(HRRP)反应了目标散射中心沿雷达视线的几何分布,可以实现对目标散射结构的精细刻画。距离像包含了丰富的目标结构信息,基于HRRP的目标识别是实现雷达自动目标识别(RATR)的重要手段。然而,距离像仅能提供目标三维散射结构在雷达视线上的一维投影,携带的信息量有限,限制了HRRP目标识别性能的提高。本文围绕这一问题展开研究,主要考虑如何有效利用极化信息与多视角观测信息增强距离像目标识别性能,研究的主要工作概括如下:第一章阐述了课题研究的背景和意义,并对距离像目标识别相关领域的研究成果进行了简要总结,最后介绍了本文的主要工作及内容安排。第二章针对距离像特征提取展开研究。首先对HRRP识别特征进行了分类总结,指出传统的特征提取方法普遍没有考虑特征的空间分布,基于特征的分类器往往存在设计与实现困难、识别过程计算量大等问题。针对这一问题,提出了一种基于序列零分量分析(SVCA)的距离像特征提取方法。理论推导证明了不相交的数据集经SVCA变换后的特征是线性可分的,从而可以利用线性分类器实现对目标类别的有效判定,分类过程具有实现简单、计算量小的特点。基于电磁散射数据及实测数据的识别结果证明了所提方法的有效性。第三章考虑利用极化信息增强距离像目标识别性能。首先对极化HRRP进行分析,指出全极化距离像样本中各分量间具有一定的相关性,并且各分量均对应相同的目标姿态。为有效利用样本中包含的先验信息进行目标识别,提出了一种基于联合稀疏表示(JSR)的极化HRRP目标识别方法。该方法在合理构造过完备字典的基础上,利用联合稀疏性对样本中各分量对应的稀疏表示系数间的相互关系进行约束,可以同时利用多极化样本两个方面的先验信息增强距离像目标识别性能。基于电磁仿真数据的目标识别结果验证了所提算法的有效性。第四章研究利用多视角观测信息增强距离像目标识别性能。考虑两种应用情景下基于多视观测的目标识别问题:多视观测在目标小角度变化范围内连续获取以及多视观测从目标全方位角度随机获取。针对前一种应用情景,提出了一种基于多任务压缩感知(MtCS)的多视目标识别方法,该方法可以有效利用多观测间的具有强相关性的先验信息进行目标识别;针对第二种应用场景下的目标识别问题,引入动态联合稀疏表示(JDSR)来描述多视观测对应的稀疏表示系数之间的相互联系,可以对多视观测间的不同相关性进行灵活建模。最后,利用电磁仿真数据及实测数据对这两种方法进行了验证。第五章研究综合利用极化信息与多视角观测信息增强距离像目标识别性能。借鉴第四章的研究思路,本章同样考虑两种情景下的目标识别问题:多视多极化观测在目标小角度变化范围连续获取以及多视多极化观测从目标全方位角度随机获取。针对前一种应用,利用原子级稀疏性对多视多极化样本中各分量对应的稀疏表示系数之间的关系进行约束;针对后一种应用,提出了一种基于层次联合稀疏表示(HJSR)的多视多极化HRRP目标识别方法。该方法根据后一种应用情景下多视多极化样本中各分量间的不同联系,分别利用原子级稀疏约束与类别级稀疏约束对单视内各极化分量以及相同极化方式下多视观测间的相互关系进行描述,可以充分利用单视内各分量对应相同目标姿态以及相同极化方式下多视观测具有不同相关性这两个层次的先验信息进行目标识别。基于电磁散射数据的识别实验证明了所提方法的有效性。第六章对本文的主要工作及创新点进行了总结,并对需要进一步深入的研究方向给出了相应的建议。
[Abstract]:The high resolution range image (HRRP) reacts the geometric distribution of the target scattering center along the radar line of sight, and can realize the fine characterization of the target scattering structure. The distance image contains rich target structure information. The target recognition based on HRRP is an important means to realize the radar automatic target recognition (RATR). However, the range image can only provide the target 3D. The one dimension projection on the radar line of sight of the scattering structure is limited, which restricts the improvement of HRRP target recognition performance. This paper focuses on this problem, mainly considering how to effectively utilize polarization information and multi view observation information to enhance the performance of target recognition. The main work of this study is summarized as follows: the first chapter The background and significance of the research are summarized, and the research achievements in the field of distance image recognition are briefly summarized. Finally, the main work and content arrangement of this paper are introduced. The second chapter studies the feature extraction of distance image. First, it classifies the recognition features of HRRP, and points out that the traditional feature extraction methods are common. Without considering the spatial distribution of features, the classifier based on feature is often difficult to design and implement, and the computational complexity of the recognition process is large. In view of this problem, a distance image feature extraction method based on sequence zero component analysis (SVCA) is proposed. The theoretical deduction proves that the characteristic of the disjoint data set after SVCA transformation is the line. It can be divided by the linear classifier. The classification process has the characteristics of simple realization and small calculation. The recognition results based on the electromagnetic scattering data and the measured data prove the effectiveness of the proposed method. The third chapter considers the performance of the target recognition by using the polar information to enhance the distance image. The polarization HRRP is analyzed. It is pointed out that there is a certain correlation between the components of the full polarimetric range image sample, and each component corresponds to the same target attitude. In order to effectively use the prior information contained in the sample to recognize the target, a polarization HRRP target recognition method based on the joint sparse representation (JSR) is proposed. On the basis of an overcomplete dictionary, the joint sparsity is used to constrain the relationship between the sparse representation coefficients corresponding to each component in the sample. At the same time, the target recognition performance of the distance image can be enhanced by using the prior information of the two aspects of the multi polarization samples. The algorithm based on the target recognition of the electromagnetic simulation data has verified the proposed algorithm. The fourth chapter studies the performance of target recognition using multi view observation information to enhance distance image. Consider the target recognition problem based on multi view observation under two application scenarios: multi view observation in the range of small angle changes in the target and the random acquisition of multi view observation from the target omnidirectional angle. A multi view target recognition method based on multi task compression perception (MtCS) is proposed. This method can effectively utilize the prior information with strong correlation between multi observation and target recognition. In view of the target recognition problem in second application scenarios, the dynamic joint sparse representation (JDSR) is introduced to describe the sparse representation corresponding to the multi view observation. The correlation between the coefficients can be modeled flexibly for the different correlations between the multi view observation. Finally, the two methods are verified by using the electromagnetic simulation data and the measured data. The fifth chapter studies the recognition performance of the comprehensive use of polarization information and multi view observation information to enhance the recognition of distance images. The research ideas of the fourth chapters are used for reference. This chapter also considers the problem of target recognition under two scenarios: the continuous acquisition of multi view and multi polarization observation in the range of small angle of the target and the random acquisition of multi view and multi polarization observation from the target all angle. The relationship is constrained. For the latter application, a multi view multi polarization HRRP target recognition method based on hierarchical joint sparse representation (HJSR) is proposed. This method is based on the different connections among the components in the multi view multi polarization samples under the latter application situation, and uses the atomic level sparse constraint and the class level sparse constraint respectively to the single view each. The relationship between the polarization components and the multi view observation in the same polarization mode is described. It can make full use of the prior information of the two levels of the same target attitude and the different correlation in the same polarization mode. The recognition experiment based on the electromagnetic scattering data is proved. The effectiveness of the proposed method. The sixth chapter summarizes the main work and innovation points of this paper, and gives the corresponding suggestions for further research direction.

【学位授予单位】：国防科学技术大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TN957.52
，

本文编号：1827945