【作者】 王芳；

【导师】 盛卫星；

【作者基本信息】 南京理工大学 ， 电子科学与技术， 2017， 博士

【摘要】 由于逆合成孔径雷达(Inverse Synthetic Aperture Radar,ISAR)能够提供目标的形状和结构特征,ISAR图像对于姿态角的敏感度较小,同时又有许多现有的图像识别技术可以借鉴。因此,基于ISAR像的雷达目标识别是非常重要且具有发展前景的目标识别方法。本文主要针对ISAR像目标识别中的特征提取问题进行了深入细致的研究。论文的主要研究工作如下:1、研究了基于ISAR像变换的特征提取方法。为了解决目前ISAR像子空间特征提取方法不允许负分解量存在的问题,本文对非负矩阵分解方法进行了研究,提出了一种基于分块双向二维投影梯度非负矩阵分解的ISAR像目标识别方法。该方法在二维投影梯度非负矩阵分解的基础上,通过对二维图像矩阵构建的大非负矩阵进行分块,即同一类的训练样本形成一小矩阵,对每块小矩阵分别进行行投影梯度非负矩阵分解和列投影梯度非负矩阵分解,通过得到的基矩阵计算每类训练样本均值的系数矩阵作为标准特征样本,根据测试样本的系数矩阵与每类样本均值的系数矩阵的最小距离,确定目标分类。2、研究了基于ISAR像多尺度的特征提取方法。目前,ISAR像特征提取算法都是直接针对原图像在单一尺度上对目标特征进行分析,没有体现目标固有的多尺度特性。针对这一问题,本文提出了一种基于直方图统计量的ISAR像目标识别方法。该方法首先将ISAR像进行Gabor小波变换,提取不同尺度和方向的Gabor幅值图谱;然后把幅值图谱分成小的子区域,用多尺度局部二值模式提取空域增强的直方图作为特征,最后在χ2统计量作为不相似度量计算的特征空间里,采用最近邻分类器完成五类飞机目标模型的分类识别。3、研究了基于ISAR像Gabor小波变换幅值和相位融合的特征提取方法。基于Gabor特征的SAR、ISAR识别算法和目前大部分基于Gabor特征的其他领域的识别算法都只应用了 Gabor的幅值信息,而没有应用相位信息。针对这一问题,本文从融合的角度去研究ISAR像的Gabor特征,将Gabor的相位特征用于ISAR像识别,并和Gabor的幅值特征相融合,提出了一种基于Gabor幅值特征和相位特征相融合的ISAR像目标识别算法。该方法首先采用Gabor小波变换对ISAR像进行多尺度分析;然后对得到的Gaobr幅值图像划分为若干矩形不重叠的子块,分别对每个子块计算其直方图分布,将所有分块的直方图分布联合起来作为Gabor的幅值特征;再采用局部Gabor相位模式和分块直方图的方法提取Gabor的相位特征;最后,将Gabor的幅值特征和相位特征采用串联的方式相融合,作为最终的ISAR像Gabor特征,通过最近邻分类器对目标完成分类识别。4、研究了基于改进的脉冲耦合神经网络和视区划分的ISAR像目标识别算法。现有基于模板匹配的ISAR像识别技术,多通过姿态遍历来构建和存储模板库,不仅占用存储空间,而且匹配开销大,为此常用等间隔方位角范围内模板平均的方法提高识别速度。然而,方位角间隔取的过小,在精简模板数量方面起不到明显的改善作用;取的过大,又会影响识别性能。针对上述问题,结合计算机视觉分析中视区的概念,提出了一种基于改进的脉冲耦合神经网络和视区划分的ISAR像目标识别算法。该方法首先建立目标多姿态较完备的初始模板库,对于每一类目标,将脉冲耦合神经网络和形态学相结合共同提取特征表述参数,然后,基于特征参数的相似性分析,在方位范围内运用迭代自组织数据分析算法进行聚类,将每一类相近的图像划归为一个视区,从而将初始模板库划分为数个视区;再在每个视区中选取一幅能代表本视区的原型模板,组成原型模板库,最后,以多类目标的原型模板同输入的待识别图像进行匹配。5、研究了基于多分类器融合的ISAR像目标识别。对于本文提出的B(2D)2PGNMF、MB-LGBP、Gabor幅值特征和相位特征相融合、改进的PCNN四种特征提取算法,分析比较了它们在不同分类器中的特征提取性能,并从多特征多分类器融合的角度出发,结合B(2D)2PGNMF、Gabor幅值特征和相位特征相融合和改进的PCNN特征提取算法,选取最近邻、决策树以及余弦分类器构建子分类器,采用朴素贝叶斯融合方法和加权投票表决融合方法实现ISAR像的多特征多分类器融合,同时比较了朴素贝叶斯融合方法和加权投票表决融合方法的性能。更多还原

【Abstract】 Inverse Synthetic Aperture Radar(ISAR)can offer the shape and structure features of targets,ISAR image is not sensitive to attitude angle,and many image recognition methods can provide reference for ISAR recognition.So radar target recognition based on ISAR image is very important and has good development prospects.In this dissertation,some research works have been made on feature extraction of ISAR image target.The main contributions are illustrated as follows:1.The transformation feature extraction method of ISAR image is studied.The negative decomposition coefficient is not allowed in subspace feature extraction method.So,non-negative matrix factorization algorithm is studied to solve it.A novel ISAR image recognition algorithm based on block two-directional and two-dimensional non-negative matrix factorization with projected gradient(B(2D)2PGNMF)is proposed.Based on the two-dimensional non-negative matrix factorization with projected gradient,the same kind of training samples form a small matrix,row non-negative matrix factorization with projected gradient and column non-negative matrix factorization with projected gradient are used in each small matrix.The coefficient matrix of the sample mean of each type of training samples is calculated as standard feature samples by the obtained base matrix.According to the minimum distance of the coefficient matrix of the test sample and the sample mean of each class,five-type aircraft models are classified.2.The multi-scale feature extraction method of ISAR image is studied.The existing feature extraction methods analyse ISAR image on single scale,which can not reflect the intrinsic multi-scale characteristics of targets.A novel ISAR image recognition algorithm based on Multi-scale Block Local Binary Patterns(MB-LBGP)is proposed.Firstly,the corresponding Gabor magnitude maps(GMMs)are obtained by convolving the enhanced ISAR image with multi-scale and multi-orientation Gabor filters.Then,each GMM is divided into small regions from which multi-scale block local binary pattern is used to extract histogram features.At last,five-type aircraft models are classified by using a nearest neighbor classifier with Chi square as a dissimilarity measure in the computed feature space.3.The fusion feature extraction method of Gabor Magnitude and phase is studied.SAR,ISAR and the most recognition algorithms in other areas based on Gabor feature only use the amplitude information.Gabor phase information is not used.In order to solve it,a novel ISAR target recognition method based on the fusion feature of Gabor Magnitude and phase is proposed in this paper.Firstly,the corresponding Gabor magnitude maps(GMMs)and Gabor phase information are obtained by convolving the ISAR image with multi-scale and multi-orientation Gabor filters.Secondly,each GMM is divided into several nonoverlapping rectangular units,and the histogram of unit is computed and combined as the magnitude histogram features.Thirdly,the local Gabor phase pattern is used to extract phase histogram features.Then,the fusion features of the Gabor magnitude features and Gabor phase features is used as the features of ISAR image.At last,five-type aircraft models are classified by using a nearest neighbor classifier with Chi square as a dissimilarity measure in the computed feature space.4.ISAR image recognition algorithm based on improved neural network and view-aspect is studied.The template-based method in inverse synthetic aperture radar(ISAR)image recognition needs a set of target templates in different gestures.It results in the great consumption of storage and computation.So,a novel method is proposed to construct the simplified prototype template library to replace the conventional library,and the improved pulse coupled neural network(PCNN)is introduced to extract the features of radar targets.Firstly,the features of ISAR images are extracted to construct the template library of feature parameters by the improved PCNN which is a fusion algorithm based on PCNN and mathematical morphology.Secondly,depending on the concept of view-aspect in computer vision,the template library of feature parameters is partitioned into different view-aspects by iterative self-organizing data analysis algorithm(ISODATA).Thirdly,the parameters of view-aspect are extracted to construct the prototype template library.Finally,the testing image is classified by matching the prototype template library.5.The multiple classifier fusion of ISAR image is studied.B(2D)2PGNMF,MB-LGBP,the fusion of Gabor Magnitude and phase and the improved PCNN are compared in the different classifers.From the perspective of multiple classifier fusion,three kinds of features,which include B(2D)2PGNMF,the fusion of Gabor Magnitude and phase and the improved PCNN,are combined with three kinds of classifiers,which include the nearest neighbor classifier,decision tree classifier and cosine classifier,to realize the multiple features and multiple classifiers fusion.Naive bayes and weighted voting are adopted to carry out the fusion.更多还原