知识辅助的SAR目标索引及特征提取技术研究

发布时间：2018-04-02 23:17

本文选题：SAR　切入点：自动目标识别　出处：《国防科学技术大学》2014年博士论文

【摘要】：基于SAR图像的自动目标识别(Automatic Target Recognition,ATR)是SAR应用领域一个极为重要的课题。目前SAR ATR面临的主要挑战是目标及目标所处环境复杂多变,仅依靠有限的目标SAR图像进行目标识别存在信息量的不足,需要引入其他外部知识和辅助信息。论文以此为切入点展开相关研究。论文首先梳理了SAR ATR的现状以及面临的主要问题,指出知识辅助的SAR ATR是克服SAR ATR困境的一条可行思路,并提出了相应的系统处理流程。知识辅助的SAR ATR引入目标、环境、传感器以及领域专家等多方面的知识,采用层次化、自适应的信息处理流程。目标索引是其中的关键环节,它的作用是在高维假设空间中确定需要重点关注的区域,为后续精细化处理提供初始假设。论文从知识辅助和扩展工作条件处理的角度对SAR目标索引及其特征提取展开研究。第二章研究知识辅助的SAR目标分割。对地面目标SAR图像,阴影也是感兴趣的区域。现有的分割方法主要考虑目标与背景在图像灰度上的差异以及像素级的局部区域连通性,普遍存在较多的目标和阴影的缺失。论文引入目标和阴影相连通的空间关系约束先验,并利用空间关系势能函数(Spatial Relation Potential Function,SRPF)进行定量描述,在Markov Random Field(MRF)分割的基础上提出SRPF-MRF分割方法,以得到更加完整的目标和阴影。而且,该方法在MRF分割的基础上仅对每个像素施加额外的先验概率,算法的计算量几乎没有增加,能够较好地满足目标索引对效率的需求。第三章从目标索引的角度对SAR图像散射中心特征提取展开分析和研究。首先分析指出属性散射中心、复图像域高分辨点散射中心提取难以满足目标索引对时效性、稳健性、自动化等方面的需求。峰值提取方法能够满足这些需求,但峰值提取由于未能考虑展布式散射中心而存在散射中心“漏检”,影响散射中心特征的目标区分能力。分析推导了属性散射中心在实图像域的表现及其与理想点散射中心的近似等价关系,构建了一种新的散射中心模型知识。据此提出一种基于CLEAN策略的SAR实图像域散射中心提取。该方法能较好地解决峰值提取存在的散射中心漏检,而且由于基于CLEAN策略,算法稳健、自动化程度高,增加的计算量不大,能够满足目标索引的需求。第四章针对地面车辆目标,从车辆目标的简化草图模型—长方体模型入手,分析指出其在SAR图像中具有矩形轮廓。矩形轮廓本身构成了一种模型知识,可以为车辆目标的部件、子结构(如坦克炮管)的描述提供很好的上下文,有利于扩展工作条件下的目标索引。矩形轮廓提取的难点主要在于目标分割存在的虚假目标、缺失以及可能的附加结构,如坦克炮管。为了克服这些难点,论文引入阴影边界、目标长宽比等约束知识,提出了一种启发式的车辆目标SAR图像矩形轮廓提取方法。MSTAR数据实验结果说明了该方法的有效性。第五章研究了基于特征子结构的目标粗分类(即索引)。首先讨论了特征子结构的内涵、特征子结构的描述、提取和预测及其在SAR目标索引中的应用。从知识辅助的角度,特征子结构综合了目标模型、观测实例、传感器知识、专家知识等。在矩形轮廓的上下文约束下,根据炮管特征子结构在SAR图像中的特点,分别从分割结果和散射中心特征检测和提取炮管,并将之用于基于显著特征子结构检测的目标索引。提取稳定散射中心作为特征子结构,考虑矩形轮廓的上下文约束,通过稳定散射中心匹配进行SAR目标索引。最后MSTAR数据实验说明了特征子结构用于SAR目标索引的有效性。论文最后(第六章)总结了论文的研究工作和成果,并对后续的研究工作进行了展望。
[Abstract]:Automatic target recognition based on SAR images (Automatic Target Recognition, ATR SAR) is one of the most important applications of the subject. The main challenge facing ATR is SAR target and target environment is complex and changeable, lack of recognition of target deposit in the amount of information on only limited SAR image, we need to introduce other external knowledge and auxiliary information. The research paper as a starting point. The main problem of this paper firstly analyzes the status of SAR ATR and SAR ATR are pointed out, knowledge aided is a practical way to overcome the difficulties of SAR ATR, and put forward the corresponding process system. SAR ATR introduced knowledge aided target, environment. Experts in the field of sensors and other aspects of knowledge, using hierarchical, adaptive information processing process. The target index is one of the key links, it is in the high dimensional space hypothesis Determine the need to focus on the region, provide the initial hypothesis for subsequent fine processing. The research on SAR index and target feature extraction from knowledge support and extension work in view of the conditions. The second chapter studies the knowledge aided SAR target segmentation. SAR image target on the ground, the shadow of the region of interest. The existing segmentation methods mainly consider the difference between target and background in the image and the local pixel connectivity, lack generally have many targets and shadows. The introduction of target and shadow connected spatial constraints and a priori, using the spatial relationship potential function (Spatial Relation Potential Function, SRPF) quantitative description in Markov Random Field (MRF) segmentation is proposed based on SRPF-MRF segmentation method, to obtain a more complete target and shadow. Moreover, the segmentation method in MRF On the basis of the prior probability of each pixel only applying additional computation, almost no increase, can better meet the needs of the target index of efficiency. In the third chapter, from the perspective of the SAR image index target scattering center feature extraction is analyzed and researched. Firstly, point out the attributed scattering center, complex image domain high resolution point it is difficult to meet the target scattering center extraction index of the effectiveness, robustness, automation and other needs. The peak extraction method can meet the demand, but the peak extraction due to the failure to consider distributed scattering center and scattering center "missing", to distinguish the impact of scattering center feature target. Deduce the approximate equivalence relation in attributed scattering center the real image domain performance and ideal point scattering center, build a new knowledge. According to the scattering center model based on CLEA N strategy SAR image domain scattering center extraction. This method can solve the scattering center extraction peak detection, and the algorithm based on the CLEAN strategy, robust, high degree of automation, small amount of computation increases, can meet the demand of the target index. The fourth chapter for ground vehicles, starting from the simplified sketch model - cuboid model of vehicle target, it is pointed out that the analysis of rectangular profile in the SAR image. The rectangular outline itself constitutes a model of knowledge, for the target vehicle components, sub structure (such as gun) description provides a good context, a target index can extend the working conditions of the difficulties. Rectangular contour extraction mainly lies in the segmentation of false targets exist, the lack of additional structure as well as possible, such as gun. In order to overcome these difficulties, this paper introduced the target shadow boundary, the ratio of length to width The constraint knowledge, we propose a vehicle SAR image contour and a heuristic method of.MSTAR data extraction experimental results illustrate the effectiveness of the method. The fifth chapter studies the rough classification of target feature based on substructure (i.e. indexing). First discusses the connotation characteristics of sub structure, sub structure feature description, extraction and prediction and its application in the SAR target index. From the angle of knowledge aided feature, substructure target model, observation example, sensor knowledge, expert knowledge and so on. In the context constraints with rectangular profile, according to the characteristics of gun tube feature sub structure in the SAR images, respectively from the segmentation results and scattering centers feature detection and extraction of the barrel, and be used to target based on feature index substructure detection. Extracting stable scattering centers as feature structure, context constraints rectangular outline, through stable Scatter center matching is used to index SAR targets. Finally, MSTAR data experiments show the effectiveness of the feature substructure for SAR target index. The last chapter (the sixth chapter) summarizes the research works and achievements of the paper, and makes a prospect for the following research work.

【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】：2014
【分类号】：TN957.52

【相似文献】