高分辨率SAR图像车辆目标分形特征提取技术

发布时间：2018-11-15 07:53

【摘要】：随着合成孔径雷达(SAR)技术的飞速发展和载装平台的多样化,战场环境的实时监测已成为可能。尤其是当SAR图像的分辨率达到了亚分米级,已具备了战场车辆目标(比如坦克、装甲车、卡车、导弹发射车等)的快速发现和分类识别能力。因此,如何基于高分辨率SAR图像实现车辆目标的判读处理已经成为战场信息处理中亟待解决的关键问题。其中,目标特征提取技术是SAR图像车辆目标判读处理的核心,所提取特征的可区分能力和普适性很大程度上影响SAR图像车辆目标鉴别虚警率,以及分类准确度。本文主要针对SAR图像车辆目标检测过程中普遍存在的高虚警概率问题,分析车辆目标和背景地物在SAR图像上的现象学差异,采用分形的理论实现上述差异性现象学特性的定量表征,并以此实现车辆目标和自然地物的辨别。论文综述了分形理论的基本概念、物理性质、研究现状和发展趋势。结合SAR成像过程中其成像对象具有复杂多变的后向散射特性,以及车辆目标像素强度在SAR图像上呈现出剧烈起伏的特性和较明显的间隙尺寸,指出欧氏几何在描述上述现象的局限性。并总结了基于分形理论定量表征车辆目标特征的物理量,如分形维数特征、扩展分形特征和间隙度特征。然后,论文重点研究了如何基于分形理论提取车辆目标的多种间隙度特征的问题。文中在对现有间隙度特征鉴别性能进行分析的基础之上提出了二重方差间隙度特征。该特征定义盒子质量为盒子内像素幅度值的方差,再计算盒子质量的标准方差,得到整个切片内像素集合的二重方差,以此作为一种新的间隙度特征量。多种实测数据的对比分析实验表明二重方差间隙度特征提升了对SAR图像车辆目标的鉴别性能,具有较好的稳定性。论文通过挖掘车辆目标和自然地物在高分辨率SAR图像上呈现的显著性差异,基于分形理论的度量方式,提出了一种高维分层间隙度特征矢量。该特征矢量能够在不同尺度上定量描述车辆目标边缘轮廓的间隙度和目标像素的不规则程度,用于消除目标提取过程中由自然地物产生的虚警。该特征矢量的计算过程中,首先对待测切片数据进行线性变换,统一其像素的灰度动态范围;然后对其由中心向四周扩散分层,以每一层的像素幅度方差作为高维分层间隙度特征矢量的一维分量;最后用模糊C均值聚类方法将高维分层间隙度特征矢量进行聚类处理,并利用聚类隶属度函数实现车辆目标与背景地物的鉴别。论文采用仿真SAR图像数据、MSTAR数据库和国内自主的机载SAR图像数据测试了文中提取的多种间隙度特征的鉴别性能,并对它们进行了对比分析。
[Abstract]:With the rapid development of synthetic Aperture Radar (SAR) technology and the diversification of loading platform, real-time monitoring of battlefield environment has become possible. Especially when the resolution of SAR image reaches sub-decimeter level, it has the ability of fast detection and recognition of battlefield vehicle targets (such as tanks, armored vehicles, trucks, missile launchers, etc.). Therefore, how to realize vehicle target interpretation based on high-resolution SAR images has become a key problem in battlefield information processing. Target feature extraction technology is the core of vehicle target interpretation processing in SAR images. The distinguishing ability and universality of the extracted features greatly affect the false alarm rate and classification accuracy of vehicle target identification in SAR images. Aiming at the problem of high false alarm probability in the process of vehicle target detection in SAR images, this paper analyzes the phenomenological differences between vehicle targets and background objects on SAR images. The fractal theory is used to realize the quantitative representation of the phenomenological characteristics of the above differences and to distinguish the vehicle objects from the natural features. The basic concept, physical properties, research status and development trend of fractal theory are reviewed in this paper. Combined with the complex and variable backscattering characteristics of the object in the SAR imaging process, the pixel intensity of the vehicle object presents the characteristics of sharp fluctuation and obvious gap size on the SAR image. The limitation of Euclidean geometry in describing the above phenomena is pointed out. Based on the fractal theory, the physical quantities of vehicle target features are summarized, such as fractal dimension feature, extended fractal feature and clearance feature. Then, the paper focuses on how to extract various clearance features of vehicle targets based on fractal theory. Based on the analysis of the existing gap degree features, a double variance gap degree feature is proposed in this paper. This feature defines the box mass as the variance of the pixel amplitude in the box, then calculates the standard variance of the box mass, and obtains the double variance of the pixel set in the whole slice. The comparison and analysis of various measured data show that the characteristics of double variance gap degree can improve the performance of vehicle target identification in SAR images and have good stability. By mining the significant differences between vehicle targets and natural objects on high resolution SAR images, a new feature vector of high dimensional stratified clearance degree is proposed based on fractal theory. The feature vector can quantitatively describe the clearance degree of vehicle target edge contour and the irregularity of target pixels at different scales, and can be used to eliminate false alarm generated by natural objects in the process of target extraction. In the calculation of the feature vector, the linear transformation of the measured slice data is firstly carried out to unify the dynamic range of the gray level of the pixel. Then, the pixel amplitude variance of each layer is taken as the one dimensional component of the feature vector of high dimensional stratification gap degree. Finally, a fuzzy C-means clustering method is used to deal with the feature vectors of high dimensional stratified clearance degree, and the membership function is used to identify the vehicle objects from the background objects. In this paper, simulated SAR image data, MSTAR database and domestic independent airborne SAR image data are used to test the discriminant performance of various gap features extracted in this paper, and to compare and analyze them.
【学位授予单位】：国防科学技术大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN957.52

【相似文献】