基于局部特性的毫米波距离像识别方法研究
发布时间：2019-07-03 20:56
【摘要】：雷达自动目标识别技术是目标探测和精确制导等应用的关键技术之一。高分辨距离像作为一类重要的雷达目标识别信号,能够反映出目标在雷达视线上的强散射点分布情况。毫米波雷达容易实现大带宽的发射信号,可提高距离分辨能力,从而能够获得更多的目标细节特征,有利于实现精确的目标识别。然而,距离像受雷达参数、目标状态、背景环境以及天气等因素的影响,呈现出高度的非线性特点,使用传统的线性方法进行距离像识别并不能得到满意的结果。流形学习是一种被广泛研究的非线性维数约减方法,能够从高维的非线性特征空问中发现线性的低维特征结构。论文针对地面目标的毫米波距离像识别问题,基于流形学习方法,从特征选择、分类器设计、主动学习和非平衡学习等四个方面展开了研究工作,主要研究内容如下：从算法的角度研究了距离像特征选择问题,提出了基于局部重构误差排列的非监督特征选择算法、基于标签重构拉普拉斯得分的半监督特征选择算法和基于改进约束得分的半监督特征选择算法。基于局部重构误差排列的特征选择算法可以看作是特征选择版本的局部线性嵌入,通过最小化局部重构误差得到最优局部特征序列,再通过排列技术得到全局特征序列。基于标签重构拉普拉斯得分的特征选择算法利用标签重构技术将基于拉普拉斯得分的特征选择算法推广到半监督应用场合,同时利用测地距离代替欧氏距离来度量非线性特征空间中的样本相似度。在基于改进约束得分的特征选择算法中,假设对约束条件和样本的局部特性并非完全独立,而是存在一定联系,通过已知的对约束条件能够改进样本的局部特性,并利用改进后的局部特性和对约束条件进行特征选择。在设计分类器时,针对距离像的方位敏感性问题,提出了基于测地权重稀疏重构的分类算法。算法假设同一目标的距离像样本在归一化之后分布在一个单位超球面的子流形上,通过小方位角范围内的样本具有高相关性的特点,对这些子流形进行分类。首先使用改进的测地距离计算所有样本之间的相似度。然后计算测地权重样本,测地权重样本能够将超球面上的子流形展开,把非线性的样本结构变换成线性结构。最后将所有标签已知样本作为字典,利用标签重构技术估计标签未知样本的类别概率。在传统的距离像识别方法中,用于训练分类器的样本通过随机选择获得。针对同一种分类器模型,不同的训练样本可能会训练出不同的分类器参数,而这些参数不同的分类器的性能也可能相差很大。主动学习的目的是在给定的训练样本集中选择一个训练样本子集,当用这个子集训练分类器时,可以获得最优的分类器。论文针对距离像识别中的主动学习问题,研究了基于局部线性重构的主动学习算法,并在该算法的理论框架下,使用拉普拉斯矩阵代替局部线性重构矩阵来描述样本的局部结构,以最优重构的方法来选择训练样本,得到了基于拉普拉斯直推优化设计的主动学习算法,并比较了几种主动学习算法在距离像识别中的效果。非平衡学习是模式识别理论在实际应用中遇到的问题,它更关注小类样本的识别能力。当用于训练分类器的样本数量非平衡时,分类面会向小类样本移动,从而降低小类样本的识别率。论文针对非平衡数据条件下的距离像识别问题,提出基于代价敏感测地约束得分的半监督特征选择算法。算法引入代价敏感技术,将基于约束得分的特征选择算法推广到非平衡学习场合,然后通过约束重构技术,将其推广到半监督应用场合,使之适用于非平衡数据分布条件下的半监督分类问题,提高小类样本的识别率。
[Abstract]:Radar automatic target recognition technology is one of the key technologies of target detection and accurate guidance. The high-resolution distance image can be used as a kind of important radar target identification signal, which can reflect the distribution of the strong scattering point of the target in the radar's line of sight. The millimeter-wave radar is easy to realize large-bandwidth transmission signals and can improve the distance-resolution capability, so that more target detail characteristics can be obtained, and the accurate target recognition can be realized. However, the distance image is affected by the parameters of the radar, the target state, the background environment and the weather. Manifold learning is a widely studied nonlinear dimension reduction method, which can find a linear low-dimensional feature structure from a high-dimensional nonlinear feature space-to-question. Aiming at the problem of millimeter wave distance image recognition of the ground target, the research work is carried out from four aspects, such as feature selection, classifier design, active learning and non-equilibrium learning, based on the manifold learning method. The main contents of the study are as follows: A non-supervised feature selection algorithm based on the local reconstruction error arrangement, a semi-supervised feature selection algorithm based on the label reconstruction Laplacian score and a semi-supervised feature selection algorithm based on the improved constraint score are proposed. The feature selection algorithm based on the partial reconstruction error arrangement can be regarded as the local linear embedding of the feature selection version, the optimal local characteristic sequence is obtained by minimizing the local reconstruction error, and the global feature sequence is obtained through the arrangement technology. The feature selection algorithm based on the label reconstruction Laplacian score is used to extend the feature selection algorithm based on the Laplacian score to the semi-supervised application, and the similarity of the samples in the non-linear feature space is measured by the geodesic distance instead of the Euclidean distance. in the feature selection algorithm based on the improved constraint score, it is assumed that the local characteristic of the constraint condition and the sample is not completely independent, but there is a certain connection, and the local characteristic of the sample can be improved by the known constraint conditions, And feature selection is carried out using the improved local characteristics and the constraint conditions. In the design of the classifier, the classification algorithm based on the weight-sparse reconstruction of the geodesic is proposed for the orientation sensitivity of the distance image. The algorithm assumes that the distance image samples of the same object are distributed on a submanifold of a unit hypersphere after normalization, and the submanifolds are classified by the characteristics of high correlation in the samples in the small azimuth range. First, the similarity between all samples is calculated using an improved geodesic distance. Then, the weight sample of the geodesic is calculated, the submanifold on the hypersphere can be expanded, and the non-linear sample structure is transformed into a linear structure. And finally, all the labels are known as a dictionary, and the category probability of the tag unknown sample is estimated by the label reconstruction technique. In a conventional distance image recognition method, a sample for training a classifier is obtained by random selection. For the same classifier model, different training samples may train different classifier parameters, and the performance of the classifiers with different parameters may also differ greatly. The purpose of active learning is to select a subset of training samples in a given training sample set. When using this subset to train the classifier, the optimal classifier can be obtained. In this paper, the active learning algorithm based on local linear reconstruction is studied for distance image recognition, and the local structure of the sample is described by using the Laplacian matrix instead of the local linear reconstruction matrix under the theoretical framework of the algorithm. The optimal reconstruction method is used to select the training samples, and the active learning algorithm based on the Laplacian direct-push optimization design is obtained, and the effect of several active learning algorithms in distance image recognition is compared. Unbalanced learning is the problem of pattern recognition theory in practical application, and it is more concerned with the identification ability of small-class samples. When the number of samples used to train the classifier is not balanced, the classification surface moves to the small-class sample, thereby reducing the recognition rate of the small-class sample. In this paper, a semi-supervised feature selection algorithm based on cost-sensitive constraint score is proposed for distance image recognition under the condition of non-equilibrium data. In this paper, the cost-sensitive technique is introduced, the feature selection algorithm based on the constraint score is extended to the non-equilibrium learning situation, and then the constraint reconstruction technique is applied to the semi-supervised application, so that it can be applied to the semi-supervised classification problem under the non-equilibrium data distribution condition, And the recognition rate of the small sample is improved.
【学位授予单位】：南京理工大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TN957.52
本文编号：2509662
[Abstract]:Radar automatic target recognition technology is one of the key technologies of target detection and accurate guidance. The high-resolution distance image can be used as a kind of important radar target identification signal, which can reflect the distribution of the strong scattering point of the target in the radar's line of sight. The millimeter-wave radar is easy to realize large-bandwidth transmission signals and can improve the distance-resolution capability, so that more target detail characteristics can be obtained, and the accurate target recognition can be realized. However, the distance image is affected by the parameters of the radar, the target state, the background environment and the weather. Manifold learning is a widely studied nonlinear dimension reduction method, which can find a linear low-dimensional feature structure from a high-dimensional nonlinear feature space-to-question. Aiming at the problem of millimeter wave distance image recognition of the ground target, the research work is carried out from four aspects, such as feature selection, classifier design, active learning and non-equilibrium learning, based on the manifold learning method. The main contents of the study are as follows: A non-supervised feature selection algorithm based on the local reconstruction error arrangement, a semi-supervised feature selection algorithm based on the label reconstruction Laplacian score and a semi-supervised feature selection algorithm based on the improved constraint score are proposed. The feature selection algorithm based on the partial reconstruction error arrangement can be regarded as the local linear embedding of the feature selection version, the optimal local characteristic sequence is obtained by minimizing the local reconstruction error, and the global feature sequence is obtained through the arrangement technology. The feature selection algorithm based on the label reconstruction Laplacian score is used to extend the feature selection algorithm based on the Laplacian score to the semi-supervised application, and the similarity of the samples in the non-linear feature space is measured by the geodesic distance instead of the Euclidean distance. in the feature selection algorithm based on the improved constraint score, it is assumed that the local characteristic of the constraint condition and the sample is not completely independent, but there is a certain connection, and the local characteristic of the sample can be improved by the known constraint conditions, And feature selection is carried out using the improved local characteristics and the constraint conditions. In the design of the classifier, the classification algorithm based on the weight-sparse reconstruction of the geodesic is proposed for the orientation sensitivity of the distance image. The algorithm assumes that the distance image samples of the same object are distributed on a submanifold of a unit hypersphere after normalization, and the submanifolds are classified by the characteristics of high correlation in the samples in the small azimuth range. First, the similarity between all samples is calculated using an improved geodesic distance. Then, the weight sample of the geodesic is calculated, the submanifold on the hypersphere can be expanded, and the non-linear sample structure is transformed into a linear structure. And finally, all the labels are known as a dictionary, and the category probability of the tag unknown sample is estimated by the label reconstruction technique. In a conventional distance image recognition method, a sample for training a classifier is obtained by random selection. For the same classifier model, different training samples may train different classifier parameters, and the performance of the classifiers with different parameters may also differ greatly. The purpose of active learning is to select a subset of training samples in a given training sample set. When using this subset to train the classifier, the optimal classifier can be obtained. In this paper, the active learning algorithm based on local linear reconstruction is studied for distance image recognition, and the local structure of the sample is described by using the Laplacian matrix instead of the local linear reconstruction matrix under the theoretical framework of the algorithm. The optimal reconstruction method is used to select the training samples, and the active learning algorithm based on the Laplacian direct-push optimization design is obtained, and the effect of several active learning algorithms in distance image recognition is compared. Unbalanced learning is the problem of pattern recognition theory in practical application, and it is more concerned with the identification ability of small-class samples. When the number of samples used to train the classifier is not balanced, the classification surface moves to the small-class sample, thereby reducing the recognition rate of the small-class sample. In this paper, a semi-supervised feature selection algorithm based on cost-sensitive constraint score is proposed for distance image recognition under the condition of non-equilibrium data. In this paper, the cost-sensitive technique is introduced, the feature selection algorithm based on the constraint score is extended to the non-equilibrium learning situation, and then the constraint reconstruction technique is applied to the semi-supervised application, so that it can be applied to the semi-supervised classification problem under the non-equilibrium data distribution condition, And the recognition rate of the small sample is improved.
【学位授予单位】：南京理工大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TN957.52
【参考文献】
相关期刊论文 前2条
1 袁莉,刘宏伟,保铮;基于中心矩特征的雷达HRRP自动目标识别[J];电子学报;2004年12期
2 何松华,郭桂蓉,郭修煌;FMCW毫米波雷达高分辨率目标距离像及其处理[J];系统工程与电子技术;1991年10期
相关博士学位论文 前3条
1 王桂丽;毫米波精确探测系统的信号检测及处理[D];南京理工大学;2009年
2 王瑞;针对类别不平衡和代价敏感分类问题的特征选择和分类算法[D];中国科学技术大学;2013年
3 尹华;面向高维和不平衡数据分类的集成学习研究[D];武汉大学;2012年
，本文编号：2509662
本文链接：https://www.wllwen.com/kejilunwen/wltx/2509662.html
