【作者】 李禹；

【导师】 粟毅；

【作者基本信息】 国防科学技术大学 ， 信息与通信工程， 2007， 博士

【摘要】 复杂背景中战术机动目标的提取与分类识别是SAR战场情报侦察的重要任务,也是SAR图像解译技术关注的难点。这一过程可大致分为检测、鉴别和分类识别三个阶段,其中机动目标的检测将在观测数据中快速提取出目标的ROI,经鉴别后形成待识别的ROI。由于现阶段包括目标电磁散射在内的多项关键技术还远远没有达到支持自动目标识别的程度,所以深入研究机动目标的检测和鉴别技术,实现对大幅观测场景中感兴趣机动目标的快速提取和定位具有重要现实意义。针对SAR图像机动目标的检测和鉴别问题,本文主要研究了机动目标的CFAR检测、区域分割、特征提取和目标鉴别等方面的关键技术,并以此构建机动目标ROI快速提取技术框架,实现在大幅、复杂观测场景中对机动目标的快速发现和定位。论文第二章针对非均匀区域的SAR图像数据,提出了基于ODVI-AC的机动目标CFAR检测算法。该算法首先从理论上推导了ODVI自动筛选阈值的计算公式,从而可以快速计算自适应于局部背景数据的自动筛选阈值,以消除待测单元背景参考像素集合中的强杂波像素和人造目标的干扰像素,并结合双参数CFAR检测和高斯分布模型来计算检验统计量和对应的判决阈值,完成SAR图像机动目标的自适应检测。论文第三章在对SAR图像分割技术进行综述的基础上,针对机动目标ROI的区域分割问题,提出了三种分割算法,包括基于OS-CFAR的分割算法、基于统计模型组的MRF分割算法和基于优化P-M模型的MAP分割算法,以获取机动目标的边缘、轮廓和区域等结构信息。同时,对于机动目标区域分割过程产生的多个子区域,文中采用形态学的连通、填充和消除运算,以及空间聚类算法进行滤波处理,形成较为完整的目标区域,并以此计算整个目标区域中心的像素坐标,完成ROI的定位。论文第四章总结了SAR图像目标鉴别的常用特征,包括几何特征、散射特征和多分辨率鉴别特征,给出上述特征的计算方法。文中针对单极化SAR图像的机动目标ROI,引入最小外切矩形算法和Radon变换方法来分别估计机动目标的长、宽尺寸,并对两者计算的结果进行比较和分析。同时,考虑机动目标和自然地物具有不同的后向散射特性,提出了机动目标的间隙度特征,分析了该特征对噪声包括加性噪声和乘性相干斑噪声的不敏感特性。论文第五章针对鉴别特征空间的降维处理,提出了机动目标鉴别特征的优选方法,包括对特征的冗余性、稳定性和可分离程度的定量分析,综合利用分析的结果来选择最优的鉴别特征。同时,在上述特征优选的基础上,提出了基于多特征联合的序贯鉴别算法,并与二项式距离鉴别算法进行了比较。另外,在应用本文多项关键技术的基础上,构建了高分辨率SAR图像机动目标ROI提取技术的整个框架。更多还原

【Abstract】 Detection and recognition of vehicles in a SAR image is an important task for acquirement of information from the battlefield, and is also a concerned difficult problem in SAR image interpretation. The problem is usually divided into three stages: detection, discrimination and classification. In the first stage, the observed data are scanned quickly to acquire the vehicles’ ROIs, and these ROIs are distinguished in the second stage to remove the false ROIs by the features of vehicle. And then, the only target-like ROIs are sent to the computationally expensive classification stage. However, it is almost improbable to recognize automatically those vehicles based on the existing technology, such as the theory of electromagnetic scattering. Therefore, it is significant to research further the detection and discrimination of vehicles, so that we can find quickly the targets and calculate their orientations from a large scene.In order to detect and discriminate the vehicles from the large complex observed scene, some key technologies including the CFAR detection, region segmentation, feature extraction and target discrimination are studied systematically in this thesis. And we can use these technologies to form a framework to find quickly the vehicle from the test SAR image and calculate its orientation.The CFAR detecting algorithm for vehicle based on ODVI-AC in a nonhomogeneous SAR image is developed in chapter 2. Firstly, the method of counting the threshold of the ODVI-AC is proposed. Using this method, we can calculate the adaptive thresholds of automatic censoring process. After removing the pixels of strong clutter and interfering targets in the reference window of a test cell by an ODVI-AC algorithm, the remaining pixels are used to estimate the parameters of statistical model. Adopting a two-parameter CFAR detector and Gauss distribution, we calculate the testing statistic and its threshold in CFAR to fulfill an adaptive CFAR detection of the vehicle.The region segmentation of vehicle ROI and clustering method are discussed in chapter 3. The methods of SAR image segmentation are summarized. And three algorithms, including OS-CFAR, MAP based on MRF and MAP based on the modified P-M model, are used to carry out the region segmentation of vehicle, respectively. Because a vehicle may be separated into several isolated regions, it is necessary to use the space clustering algorithm and morphologic operators, such as connection, filling, and eliminating, to filter the regions. So, we can have the whole region of the vehicle and calculate its center coordinate.Features for vehicle discrimination, including geometry features, scattering features, multiresolution discriminant feature, are summed up in chapter 4, and the methods of extracting those features are also presented. At the same time, the method of minimum enclosing rectangle and the Radon transform are introduced to estimate the length and width of the vehicle in single polarimetric SAR image. According to the different back scattering characteristic between the vehicle and natural terrain, the lacunarity feature is developed. Moreover, it is concluded that lacunarity is robust to the additional noise and the speckle noise.For reducing the dimension of discriminating feature space, the method of selecting the optimal features for target discrimination is proposed in chapter 5. The redundancy, robustness and separability of features are quantitatively analyzed by this method. Based on the optimal features, the multi-features sequential discrimination is developed and compared with the quadratic distance discriminating method. In addition, useing those key technologies in this thesis, we form the framework of acquiring the vehicle’s ROI in a high resolution SAR image.更多还原