基于统计判决的分类器设计及在雷达目标识别中的应用

发布时间：2018-11-18 15:19

【摘要】：随着科学技术的飞速发展和大规模高维数据的涌现,模式分类在越来越多的领域得到广泛的重视和应用。在国内外最新研究成果的基础上,本文着重从多类分类和一类分类两个方面展开研究。1.本文第二章介绍了现有的多类分类器和一类分类器。1)多类分类器方面,依次介绍了贝叶斯分类器、互信息准则、互信息与贝叶斯错误率之间的关系,并由此引出基于互信息准则的信息判别分析(IDA)和其它相关的多类分类器;2)一类分类器方面,介绍了支持向量域描述(SVDD)和一类支持向量机(一类SVM)两种一类分类器。最后介绍了几种公认的分类器性能评价指标。2.针对IDA中存在的高维估计误差问题,本文第三章提出了基于线性统计模型和互信息准则的多类分类器。该分类器采用线性统计模型描述观测数据的子空间统计结构,利用互信息准则对子空间可分性进行约束,通过联合优化对数似然函数和互信息函数,求取最优变换矩阵和噪声方差。该分类器可以直接求取最优变换矩阵,在使得子空间强可分性的同时,也能尽可能准确的描述观测数据。基于人工合成数据、加州大学欧文分校(UCI)公共数据和雷达实测数据的仿真实验共同验证了该分类器良好的分类性能和鲁棒性。3.针对现有一类分类器中普遍存在的模型选择问题,本文第四章提出了无限贝叶斯一类SVM分类器。首先利用归一化函数对现有的一类SVM进行改进,然后借助数据增强技术将改进的一类SVM用概率模型进行表达,得到贝叶斯一类SVM分类器,最后利用Dirichlet过程(DP)无限混合专家模型对贝叶斯一类SVM进行扩展,得到无限贝叶斯一类SVM分类器。该分类器不需要人工干预设置参数,能够自动适应数据的变化,自动学习模型参数实现模型选择。人工合成数据、UCI公共数据和雷达实测数据证明了该分类器具有良好的一类分类性能。
[Abstract]:With the rapid development of science and technology and the emergence of large-scale high-dimensional data, pattern classification has been widely paid attention to and applied in more and more fields. On the basis of the latest research results at home and abroad, this paper focuses on multi-class classification and class classification. 1. The second chapter introduces the existing multi-class classifiers and a class of classifiers. 1) the relation between Bayesian classifier, mutual information criterion, mutual information and Bayesian error rate is introduced in turn. From this, the information discriminant analysis based on mutual information criterion (IDA) and other related multi-class classifiers are elicited. 2) for a class of classifiers, two kinds of classifiers, support vector domain description (SVDD) and support vector machine (SVM), are introduced. Finally, several generally accepted performance evaluation indexes of classifier are introduced. 2. In order to solve the problem of high dimensional estimation error in IDA, a multi-class classifier based on linear statistical model and mutual information criterion is proposed in chapter 3. The classifier uses linear statistical model to describe the subspace statistical structure of observed data, uses mutual information criterion to constrain the separability of subspace, and optimizes logarithmic likelihood function and mutual information function by jointly optimizing logarithmic likelihood function and mutual information function. The optimal transformation matrix and noise variance are obtained. The classifier can directly obtain the optimal transformation matrix, which can describe the observation data as accurately as possible while making the subspace strongly separable. Based on synthetic data, the simulation results of (UCI) common data and radar measured data at UCLA Irvine verify the good classification performance and robustness of the classifier. In order to solve the problem of model selection in the existing class of classifiers, we propose an infinite Bayesian class of SVM classifiers in chapter 4. First, the existing SVM is improved by normalized function, and then the improved SVM is expressed by probability model with the help of data enhancement technique, and a Bayesian SVM classifier is obtained. Finally, an infinite Bayesian SVM classifier is obtained by extending the Bayesian class of SVM by using the (DP) infinite mixed expert model of the Dirichlet process. The classifier does not need manual intervention to set parameters, and can automatically adapt to the change of data, and automatically learn model parameters to realize model selection. Synthetic data, UCI common data and radar data show that the classifier has good classification performance.
【学位授予单位】：西安电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN957.52

【共引文献】