Study On Support Vector Machine Method And Its Application In Target Recognition For Airborne MMW Radar

The support vector machine(SVM) method and its application in target recognition for airborne MMW radar is studied in this paper.First, the research background of this paper is expatiated, the basic content and difficulty of radar target recognition are reviewed, and the potential advantages of applying SVM to resolve practicing problem are analyzed. Then the SVM theory is introduced, and the present research situation of SVM about pattern classification is summarized.In chapter 2, the problem of pre-extracting support vector(SV) is studied to improve SVM training algorithm. Firstly, the characteristic of the SV is analyzed, and the equivalence of the SV set and the training set is pointed out. Secondly, in allusion to the shortcoming of the existing pre-extracting algorithm, and combining with the characteristic of radar target recognition, a method based on density is proposed to pre-extracting SV. This method needn't confirm that the training samples is whether linear separable or not beforehand, and has strong ability of diminishing the effect of noises and outliers. This method is simple and quick realization. Experiments show the validity of this method.In chapter 3, SVM incremental learning which is an another approach of improving SVM training algorithm is studied. Firstly, the KKT condition for the optimal solution of SVM is presented, the relationship between the KKT condition and training samples, and the relationship between the KKT condition and newly-added samples are discussed, and the possible changes of SV set after including new samples into training set are analyzed. Secondly, in allusion to the shortcoming of the existing algorithm, a new removing algorithm based on density for incremental learning is proposed by analyzing the advantages of removing the non-boundary vectors with density. Lastly, the detailed learning process and removing rule based on density are presented. Compared with traditional SVM method, this algorithm can not only keep the testing accuracy, but also reduce storage cost and increase training speed.In chapter 4, the practical problem of transductive SVM(TSVM) is studied in the case that there are a large number of unlabeled samples in radar target recognition. Firstly, the TSVM model is introduced, and the questions will to be solved by transductive inference are pointed out. Secondly, the characteristic of transductive inference is analyzed, then the transductive support vector machine learning algorithm devised by Joachims and the progressive transductive support vector machine learning algorithm devised by Chen Yisong et al are presented. Next, in allusion to the shortcoming of the progressive transductive support vector machine learning algorithm, some measures are taken to improve its speed and performance. Lastly, in order to achieve higher performance, KNN is used to reduce the useless unlabeled samples before the learning by the progressive transductive support vector machine learning algorithm. Experiments with radar raw data show the validity of these algorithms.In chapter 5, in allusion to the shortcoming of the existing high resolution radar target detection algorithm, taking the problem of high resolution radar target detection as the problem of true-false target recognition, and borrowing ideas from the dealing with novelty problem, this paper introduces one-class SVM into high resolution radar true-false target recognition for the first time. That can provide a new idea for solving high resolution radar true-false target recognition problem. Firstly, the high resolution radar target detection algorithm based on the position correlation information of the target scattering points devised by Huang Deshuang et al is introduced, and the disadvantage of the algorithm designing is analyzed. Secondly, the data characteristic of true-false target recognition problem is discussed, and the difficulties of dealing with this problem by two-class SVM are pointed out. Next, the existing one-class SVM algorithm is introduced briefly, the algorithm theory and the influence of parameter changing to the hyperplane are particularly analyzed, and the main measure to realize one-class SVM by describing positive samples domain is pointed out. Lastly, in allusion to the incompleteness of the existing one-class SVM in describing data domain, and combing with the data distribution characteristics of the high resolution radar object, a cluster one-class SVM model is proposed. It conducts training positive kind based on clustering, uses several small spheres instead of previously one big sphere, and gives more accurate description of the data domain. At the same time, in allusion to the condition of the existing several kinds of true targets, a method that deals with every single kind of true target separately is proposed to satisfy with the need of the succeeding true target type identifying.In chapter 6, combing with the characteristic of radar target recognition technique research, a radar target recognition software platform with blocking design merits is designed to resolve the difficulties in developing radar target recognition software system. The algorithms in this paper are integrated in the platform, and the performance of these algorithms are validated by using raw data from practical engineering project.At last, summary of this paper is made and the problems needed further research are pointed out in chapter 7.