【作者】 胡利平；

【导师】 吴顺君；

【作者基本信息】 西安电子科技大学 ， 信号与信息处理， 2009， 博士

【摘要】 基于合成孔径雷达(Synthetic Aperture Radar,SAR)的自动目标识别(Automatic Target Recognition,ATR)技术在现代战场中的重要作用已经使得它成为了国内外研究的热点之一。近十几年来,SAR图像的目标识别研究在SAR图像的预处理、特征提取及识别等方面均取得了相当大的进展。本论文主要围绕教育部留学回国人员基金“基于SAR图像的雷达自动目标检测与识别技术”、“十五”国防预研项目“目标识别技术”和“十一五”国防预研项目“基于目标成像的识别技术”的研究任务,针对合成孔径雷达图像的目标识别,从SAR图像滤波、SAR图像分割以及SAR图像的特征提取与识别等方面展开了较为深入的研究。本论文的主要内容概括如下:1、针对SAR图像的相干斑特点,对比了几种常用的SAR图像滤波方法,分析了它们的优缺点,并给出了不同分辨率的SAR图像的滤波结果。2、针对相关文献中存在的问题:①没有考虑阴影,而阴影信息对识别是很有用的;②没有将目标及其阴影从杂波背景中提取出来,由于背景杂波具有多样性,不同的背景杂波特性会影响识别性能;③对于一个特定的识别问题,采用某一种分类算法或者某一种特征未必能获得很好的识别性能,基于多特征或多分类算法的分类器融合是必需的,我们提出一种基于多分类器融合的SAR图像目标识别方法,首先给出有效的SAR图像预处理方法,将目标及其阴影从杂波背景中提取出来,抑制了背景杂波对后续识别的影响,然后基于极化映射提取目标的强度分布特征、目标和阴影的形状特征等,最后基于平均准则融合多个分类器。3、主分量分析(Principal Component Analysis,PCA)是模式识别领域中一种经典的特征提取方法。然而,当PCA用于图像的特征提取时,要将2维图像矩阵(m×n)转换成1维向量(m ? n),这会带来两个方面的问题:①损失图像的2维空间结构信息;②特征提取要在高维向量空间中进行,但在高维空间中很难准确估计协方差矩阵,且维数很大(( m ? n )×( m ? n)),对其进行特征分解会大大增加计算负担。为了解决上述问题,二维主分量分析(Two-dimensional PCA,2DPCA)应运而出,它直接采用2维图像矩阵估计训练样本的协方差矩阵,估计得到的协方差矩阵更准确有效且维数仅为n×n或m×m,对其特征分解的效率也更高。但是,2DPCA仅去除了图像各行或各列像素间的冗余信息,因此得到的特征矩阵维数较大。本章首先根据投影形式的不同将2DPCA分为两种:右投影形式的2DPCA(Right 2DPCA,R-2DPCA)和左投影形式的2DPCA(Left 2DPCA,L-2DPCA)。为了降低特征维数、改善识别性能,给出相应的改进2DPCA方法。4、线性判决分析(Linear Discriminant Analysis,LDA)也是模式识别领域中一种有效的特征提取方法。与PCA类似,LDA在用于图像特征提取时,也需要将2维图像矩阵转化为1维图像向量,这会带来“维数灾难”和“奇异”等问题。为此,近年来提出二维线性判决分析(Two-dimensional LDA,2DLDA),它直接基于2维图像矩阵来构建散布矩阵。本章中根据投影形式的不同首先将2DLDA分为两种形式:右投影的二维线性判决分析(Right 2DLDA,R-2DLDA)和左投影的二维线性判决分析(Left 2DLDA,L-2DLDA)。然后针对它们特征维数过大的缺陷,提出三种改进算法:两级R-2DLDA(Two-stage R-2DLDA)、两级L-2DLDA(Two-stage L-2DLDA)和两向2DLDA(Two-directional 2DLDA)。5、LDA是一种被广泛应用的线性降维算法,但它要求各个类别的数据要满足单模分布结构,且在用于图像特征提取时通常会出现类内散布矩阵奇异的问题,还有LDA得到的特征维数仅为c? 1(c为类别数)。为了缓解上述局限性,近来提出了子类判别分析(Clustering-based Discriminant Analysis,CDA),它假设数据服从多模分布。然而,由于该方法采用的子类划分方法是k均值聚类算法,因而不能保证最终的识别结果是稳定的、最优的、且依赖聚类初始中心的选择。为此,我们给出一种改进的子类判决分析(Improved CDA,ICDA)方法,它首先采用快速全局k-均值聚类算法找到每类目标最优的子类划分,然后基于这些子类划分采用CDA准则找到最优的投影矢量,因此最终的识别性能不依赖聚类时初始聚类中心的选择,且能保证达到全局最优。6、针对SAR图像数据的多模分布特性,提出了以下几种图像特征提取方法:(1)提出二维子类判决分析(Two-dimensional CDA,2DCDA)及其改进的算法。它们直接基于2维图像矩阵构造子类类间和子类类内散布矩阵,可克服CDA的维数灾难、奇异等问题。本章先后给出2DCDA的两种投影形式:右投影形式的2DCDA(Right 2DCDA,R-2DCDA)和左投影形式的2DCDA(Left 2DCDA,L-2DCDA)。针对它们求得的特征矩阵维数过大的问题,相应地提出几种改进的2DCDA算法。(2)提出二维最大子类散度差鉴别分析及其改进算法。由于在2DCDA中,需要计算子类类内散布矩阵的逆矩阵,而在小样本问题中逆矩阵通常是不存在的。为了避免计算逆矩阵或逆矩阵不存在的问题,我们给出一种新的图像特征提取方法:二维最大子类散度差鉴别分析(Two-dimensional Maximum Clustering-based Scatter Difference,2DMCSD)。由于2DMCSD只沿行方向压缩图像,类似地,给出其另一种形式(称之为Alternative 2DMCSD),它只沿列方向压缩图像。为了克服2DMCSD和Alternative 2DMCSD的特征维数过大的问题,又提出两向二维最大子类散度差鉴别分析(Two-directional 2DMCSD ,(2D)2MCSD)。(3)提出对角子类判决分析算法。上述的2DPCA、2DFLD和2DCDA仅保留了图像行(或列)方向的相关性变化,而忽略了图像列(或行)方向的相关性变化。为了同时保留图像行和列像素间的相关性变化,提出对角子类判决分析(Diagonal CDA,DiaCDA)。它基于对角图像寻找最优的投影方向,且考虑了每类数据中存在多个子类的情况。为了缓解DiaCDA的特征维数过大的问题,将DiaCDA和2DCDA结合起来,提出DiaCDA+2DCDA的特征提取方法。更多还原

【Abstract】 Automatic target recognition (ATR) based on synthetic aperture radar (SAR) images is of great importance in the modern battlefield and has become a very hot research topic. In recent years, ATR based on SAR images has made great progress in related techniques including SAR images preprocessing, feature extraction, classifier design, and so on. This dissertation provides our researches for SAR target recognition, which are supported by Advanced Defense Research Programs of China and Natural Science Foundations of China.The main content of this dissertation is summarized as follows:1. In the first part, several SAR images filtering methods are first analyzed and compared according to the characteristics of the SAR speckles, and then the filtered results of some SAR images with different resolutions are presented.2. To solve the problems in many literatures, a SAR ATR method based classifier fusion is proposed where target and shadow images are first segmented via SAR image pre-processing, and then the shape information of target and its shadow and the intensity distributed information of a target are extracted based on polar mapping, and shape descriptors of a target and its shadow are also extracted, finally SAR targets are classified by the combined classifier based on the average rule.3. Principal component analysis (PCA) is a classical method in the pattern recognition area. However, when PCA is used to feature extraction for 2D images, 2D image matrices need to be transformed into 1D image vectors, this will bring on some problems such as a loss of 2D space structure information, disaster of dimensionality, etc. To solve these problems, 2DPCA is proposed recently where the projection directions are sought out from 2D image matrices, thus being more efficient. In this dissertation, 2DPCA is divided into right 2DPCA (R-2DPCA) and left 2DPCA (L-2DPCA) according to the ways of projection. To overcome the problem of more features of 2DPCA, we present four improved methods which not only can reduce feature dimensions but also can improve recognition performances.4. Linear discriminant analysis (LDA) is also a popular feature extraction method in the pattern recognition field. Similar to PCA, when LDA is used to 2D images recognition tasks, some problems (such as a loss of 2D space structure information, disaster of dimensionality, singularity, etc) will occur. Therefore, 2DLDA is presented to solve the above problems, which constructs the scatter matrices based on 2D image matrices. 2DLDA can also be divided into right 2DLDA (R-2DLDA) and left 2DLDA (L-2DLDA). However, a drawback of R-2DLDA and L-2DLDA is that they need more features. To overcome this problem, we propose three improved approaches, two-stage R-2DLDA, two-stage L-2DLDA and two-directional 2DLDA.5. LDA is a popular method for linear dimensionality reduction. LDA assumes that all the classes obey the unimodal distribution, singularity usually occurs when used to 2D image recognition tasks, and the dimensionality of features obtained by LDA is only c ? 1 ( c is the total number of classes). To alleviate the above-mentioned limitations, clustering-based discriminant analysis (CDA) is presented recently. In this method, k-means algorithm is employed for finding the clusters of each class, thus the final recognition results are unstable, not optimal and depend on the initial cluster centers. In this dissertation, we propose an improved CDA (ICDA) where the fast global k-means clustering algorithm is employed for finding the optimal cluster structures, thus the final results are stable and optimal.6. This section presents several image feature extraction methods, which aim at dealing with the multimodal distributions of SAR images. The main work concerns the following three aspects: (1) Two-dimensional CDA (2DCDA) and its improved algorithms are developed. 2DCDA constructs the cluster scatter matrices from 2D image matrices, thus overcoming the problems (such as disaster of dimensionality, singularity, etc.) of CDA, and hence 2DCDA combines the capability to model the multiple cluster structures embedded within a single class with the computational advantage that is characteristic of 2D subspace analysis methods, such as 2DPCA and 2DLDA. In this section, 2DCDA is also divided into right 2DCDA (R-2DCDA) and left 2DCDA (L-2DCDA). Moreover, in order to solve the problem of too much features of 2DCDA, we propose four improved algorithms, two-stage R-2DCDA, two-stage L-2DCDA, two-directional 2DCDA ((2D)2CDA), and generalized 2DCDA (G2DCDA). (2) Two-dimensional maximum clustering-based scatter difference discriminant analysis (2DMCSD) and its improved algorithms are proposed. In 2DCDA, the inverse matrix of the within-cluster scatter matrix has to be calculated, however, usually the inverse matrix does not exist in“small sample size”(SSS) problems. To solve this problem, we propose a novel image feature extraction method, 2DMCSD, which adopts the difference of between-cluster scatter and within-cluster scatter as the discriminant criterion for finding the projection vectors. So it not only can deal with the multimodal distribution problems but also is capable of avoiding the inverse matrix calculation and the“SSS”problems. Besides, an alternative 2DMCSD is presented. Two-directional 2DMCSD ((2D)2MCSD) discriminant analysis is also developed for further dimensionality reduction. (3) Diagonal CDA (DiaCDA) is proposed. In 2DPCA, 2DLDA and 2DCDA, the projection vectors only reflect the variations between rows (or columns) of images while the omitted variations between columns (or rows) of images may be also useful for recognition. To preserve the correlations between variations of both rows and columns of images, diagonal CDA (DiaCDA) is proposed, which seeks the optimal projection vectors from the diagonal images and takes into account the possibility of embedded multiple cluster structures within a single class. However, an inextricable problem of DiaCDA is that it requires vast memory for representation of images. To alleviate this problem, we combine DiaCDA with 2DCDA (DiaCDA+2DCDA).更多还原