【作者】 万朋；

【导师】 黄顺吉；

【作者基本信息】 电子科技大学 ， 通信与信息系统， 2000， 博士

【摘要】 近年来合成孔径雷达（SAN）在许多方面得到了应用，其技术能够用来检测 杂波背景中的感兴趣目标。一些学者和研究机构研究了SAR目标检测和识别， 在这方面比较有影响的研究机构是MIT的Lincoln实验室。该实验室在DARP （国防高级研究计划署）的支持下展开了自动目标识别（ATR ）研究，该课题己 在发展和完善中。 本文在国防科工委支持下研究了SAR目标检测技术，以及与之有关的如噪 声抑制，边缘提取等方法。针对这些技术和方法的不足之处，提出了一些新的方 法和算法。其内容如下： 1．研究了相干斑噪声抑制技术，提出了门）增强小波软阈值相干斑噪声抑 制方法。考虑SAN杂波复杂情况，不同类型的区域滤波要求不同，本文将小波 软阈值滤波方法与 SAR杂波特点结合。（2）增强小波维纳相干斑噪声抑制方法。 基于合成孔径霄达图像杂波结构，结合小波变换和自适应维纳滤波提出了新的抑 制 SAR图像相干斑噪声方法，该方法能够较好保留杂波边缘和点目标。方法（2） 同方法门）比较，利用自适应维纳滤波的优点，避兔了小波阈值的选取，克服 了方法（1）的不足之处。实验表明这两种相干斑噪声抑制方法能够获得比较满 意的效果，并且增强小波维纳方法略优于增强小波软阈值方法效果。 2．分析了 SAR目标检测技术，提出了u）根据 SAR统计分布特性的鲁棒恒 虚警目标检测方法，在Gaona分布条件下研究了SAR目标检测，得到了阈值系数 与恒虚警关系，提出了阈值选择理论及其简单实现方法，优化了杂波均值估计并 给出了其鲁棒算法。在方法（3）中，相干斑噪声的存在影响了杂波均值的估计， 中值具有较好的抗噪特性，但是中值往往是均值的有偏估计，对于Ganun分布而 言，中值与均值具有比例关系。（4）抑制相干斑后的增强目标检测方法，分析了 抑制SAR图像相于斑噪声后的多分布特性，研究了相应的SAR目标检测，提出了 一种新的 SAR图像目标检测方法及其实现。在方法（4）中，去噪后的 SAR图像 中有三种不同的统计分布，即：高斯分布、Gamma分布、高阶Gamma分布，因此 要针对不同的统计分布分别作出相应的检测方法，最后采用区域掩膜得到检测后 的SAR目标。实验表明鲁棒恒虚警目标检测方法和增强目标检测方法均具有较好 检测性能。 3．研究了SAR边缘提取，提出了鲁棒比例边缘提取方法o人在经典比例边 缘提取方法中，需要进行均值估计，这要受到相干斑的影响，相干斑噪声是边缘 提取的障碍，采用传统的抑制方法会损害边缘，而增强相于斑抑制技术事先对边 缘作了某种假定。是否有这样一种边缘检测方法，它不考虑噪声抑制问题，而又 IV要避兔噪声的干扰？这里利用了第四章目标检测的研究成果，提出了中值比例边缘提取的鲁棒方法。实验表明鲁棒比例边缘法是一种效果较好的边缘提取方法。 4．研究了SAR数据的地物和目标分类。分类也是一种描述SAR中目标、地物和地貌的常用手段，本文叙述了常用的适合SAn的一阶和二阶纹理特征，用大量的实验结果来描述这些特征的相对贡献，从而获得有用特征集：分析了常用的分类器，采用最小距离分类法，最大似然分类法，C均值聚类法，学习矢量量化神经网络聚类法来对SAR地物分类。同时本文研究了利用将小波结合恒虚警技术来表示SAR目标方法。小波变换适合信号多分辨分析，小波滤波后其方差特征能有效地表示信号。将SAR图像经过恒虚警处理后通过小波滤波器，用不同频段的方差矢量能够有效地表示目标特征。实际SAR数据的测试结果也表明了该研究方法的有效性。 本文在目标检测，相干斑抑制，边缘提取方面完成了大量的新算法验证，其结果表明它们都优于常用的经典方法。更多还原

【Abstract】 In recent years, Synthetic Aperture Radar (SAR) has been used in many fields. SAR technology can be used to detect radar targets of interesting, which embedded in strong ground clutter. Some researches and institutes have studied target detection and recognition for SAR. Lincoln Lab of MIT has produced much important influence in the field, which is support by DARPA (Defense Advanced Research Projects Agency) to study automatic target recognition (ATR). The project has been processing and finishing.This thesis has studied methods of target detection and its related technologies such as speckle suppressing and edge extracting for SAR by support of institute of science and technology of national defense. Some technologies and algorithms are proposed according to shortcoming of recent methods.1. Technology of speckle noise suppressing is studied, first new method of enhanced wavelet soft-threshold for speckle noise suppressing is proposed (1). This method combines wavelet soft-threshold and scene heterogeneity of SAR, because the different scene needs different filter method. Second new technology of enhanced wavelet Wiener for speckle noise suppressing is proposed (2), which combines wavelet transform and adaptive Wiener filter according to SAR image scene heterogeneity. It can better preserve clutter edge and point target. Method (2) has advantage because it adopts adaptive Wiener filter so the method don’ t select wavelet threshold. Real SAR image testing satisfies the validities of these two methods and method (2) is better than method (1).2. Technology of Target detection for SAR is studied. A first new method of robust constant false alarm rate for target detection according to statistic of SAR image is proposed. The relationship between threshold coefficient and constant false alarm probability (CFAR) is obtained after SAR target detection is analyzed according to Gamma distribution. The theory and simple realizing approach of selecting threshold coefficient are proposed. The robust way of optimizing clutter’ s mean is given, which has de-noising ability without processing of de-noising in target detection. The mean value is influenced due to speckle, the middle value has better performance in the field of anti-noise, but middle value is error estimate to mean, which is corrected by a coefficient. Another new method of target detection is proposed (4), in which enhanced speckle noise suppressing is applied, so multi-statistical has appeared, it’ sGaussian distribution, Gamma distribution and high Gamma distribution. Different methods of target detection are applied for different statistical. The final result is obtained by area combining. Experiments show the two methods are valid and they have better behavior.3. A new robust edge extracting (5) is proposed after the edge extracting for SAR image is studied. Edge extracting is influenced by speckle noise, because mean need to estimate in the traditional ratio edge method. Speckle noise is harmful to edge extracting, traditional ratio edge fltethod isn’ t better, and enhanced speckle suppressing has an edge suppose. Is a method for edge extracting existed, in which speckle is no harmful without suppressing speckle noise? A new middle value robust appeared according to result of chapter four. The validity of the methods is proved by experiments for real SAR image.4. Classification of Ground scene and target for SAR image is studied in the thesis. Method of classification is always a technology for SAR. Some one scale and two scale texture features are described and useful features are selected by testing relatively contribution according to much experiment result. Some classifier such as minimum distance classifier, maximum likelihood classifier (ML), and learning vector quantify (LVQ) classifier and c-mean classifier. Target’ s describer by combining wavelet transform and constant false alarm rate is studied, wavelet transform is suitable to multi-distinguish analysis and its variance of wavelet field can represent signal. So that the更多还原