Classifier Design Based On Statistical Decision With Applications To Radar Target Recognition

With the rapid development of science technology and the emergence of large-scale high-dimensional data, pattern classification has received intensive attention and application in more and more fields. Based on the new research at home and abroad, this dissertation focuses on pattern classification and forms two aspects: multi-class classification and one-class classification.Chapter 2 introduces available multi-class classifiers and one-class classifiers. 1) In multi-class classification, we introduce Bayesian classifier, mutual information(MI) criterion, and thus leads to information discriminant analysis(IDA) and other related multi-class classifiers; 2) In one-class classification, we introduce two kinds classifiers: support vector data description(SVDD) and one-class support vector machine(OCSVM). Finally, several common evaluation criterions to the performance of classifiers are described.Aiming at the high-dimensional estimation error that exists in IDA, chapter 3 proposes a multi-class classifier based on linear statistical model and MI criterion. This classifier describes the subspace statistical structure of the observed data via linear statistical model, and utilizes MI criterion to further constrain the subspace separability. Finally, this classifier seeks the optimal transformation matrix and noise variance via the joint optimization of the log-likelihood function and the MI function, which can not only ensure the subspace separability, but also describe the observed data as accurate as possible. Simulation experiments for synthetic data, benchmark UCI data and measured radar data prove the validity of the proposed classifier.To deal with the model selection problem of the existing one-class classifiers, chapter 4 proposes an infinite Bayesian one-class support vector machine. Firstly, we utilize a nonmalized function to improve the OCSVM. Then, we express the improved OCSVM into Bayesian OCSVM based on data augmentation technique. Finally, the Bayesian OCSVM can be extended to infinite Bayesian OCSVM via Dirichlet process(DP) mixture model. This classifier, which does not need to manually set the model parameters, can adapt to changes in the data and learn the parameters automatically. Simulation experiments for synthetic data, benchmark UCI data and measured radar data show the advantage of the proposed classifier.