Study On Distributed Feature Extraction Of Scattering Centers And Key Technique Of Classifier Based On Kernel Method

All the work in the dissertation is supported by the project of"973"National Security Important and Foundational Research, which is focused on investigation of the new mechanism and methods in radar automatic target recognition. Aiming at the background of high-resolution radar target classification, the feature extraction of scattering center parameter estimation and the key technique of classifier based on kernel method are investigated. Three important issues are included, which are feature extraction method based on scattering centers parameter estimation, parameter optimazation of Support Vector Machine (SVM) in high resolution radar target classification and research on generalizing Support Vector Data Description (SVDD) to multi-class classification.In the aspect of feature extraction of scattering center parameter estimation, one new feature extraction method is proposed. The resolution of traditonal methods is low and the extracted position of scattering centers cannot be directly used in radar target classification. The scattering center number estimation is performed and the amplitude and position of scattering centers are estimated based on one super-resolution method. Then the distributed feature of relative distance between scattering centers is extracted, which is rotationally and translationally invariant compared with position feature itself in a certain aspect scope. Finally by identifying one-dimensional range profiles of three targets, the effectiveness and robustness of the proposed method are testified.In the aspect of SVM parameters optimazation in high-resolution radar target classification, one parameter optimization method based on improved Gentic algorithm (GA) is proposed.Aiming at the problems of loose relationship between fitness function selection of GA and research background, blind determination of parameter searching scope and nonautomatic adjustation of GA convergence, several problems have been investigated.The fitness function selection in high-resolution radar target classification is investigated firstly. Parameter optimization performances of several generalization error estimation methods are analyzed and the estimation precision, computing time cost and robustness are compared respectively. One estimation method fit for SVM parameters optimization in high range resolution radar target classification is chosen as the fitness function of GA.Then the relationship between parameters and performance of SVM classifier is studied using one-dimensional range profiles. The parameter seaching scope is determined by analyzing the relationship. The study here overcomes the blind determination of parameter searching scope. In the end, one parameter optimization and feature extraction algorithm based on improved GA is proposed. Parameters such as kernel function parameter, penalty factor, composite kernel coefficient, the crossover ratio and mutation ratio of GA are all optimized simultaneously. The improved GA can automatically adjust the crossover and mutation ratios according to the state of GA progress, which overcomes the fast-speed or low-speed convergence of GA because of inappropriate crossover and mutation ratio. At last by classifying radar range profiles of three targets, the effectiveness of the algorithm is testified.In the aspect of research on generalizing SVDD to multi-class classification, one SVDD multi-target classification method based on minimal distance classifier is brought forward.SVDD is a one-class classifier and it cannot be directly used in multi-class classification. To solve this problem, research has been performed as follows.The performances of SVDD and other data description method are compared and then the relationship between kernel function parameter and SVDD classifier is analyzed. Finally, based on the idea of minimum distance classifier and threshold strategy, one SVDD multi-class classification algorithm is propsed, which is nominated as SVDD_MDRTS. Based on simulation range profiles, the effects of kernel function parameter and noise on SVDD_MDRTS are analyzed. Based on high-frequency electronic and magnetic scattering data, the confusion matrix and Receiver Operating Characteristic (ROC) curve are investigated. Finally the classification performance and testing time of SVDD_MDRTS, SVM and NN are compared, experimental results indicate that SVDD_MDRTS has better perfomance and shorter training and testing time than SVM and NN.