Research On Support Vector Data Description For HRRP-based Target Recognition

Modern warfare has developed into the high-tech war centered on electronic warfare and information warfare,setting higher requirement for the ability to analyze and process the battle-field situation.Following the new militart reform,Radar automatic target recognition(RATR)emerged for greatly supporting the defensive and offensive ability,the automated command and early warning system,which establishes the vital position of RATR in modern military high-tech fields.The high-resolution radar can achieve the highest level of target recognition,and the high Resolution Range Profile(HRRP),synthetic aperture radar(SAR)and inverse synthetic aperture radar(ISAR)can be extracted from the echo.Compared with SAR and ISAR,HRRP is easy to be obtained,requires less storage space and contains much structure information of target simultaneously,therefore this dissertation deeply study multi-target recognition technology based on HRRP.Firstly,aiming at the serious overlap problem of HRRP data,a novel model optimization algorithm based on support vector data description(SVDD)is proposed.Support vector machine is used to explore the mechanism mining the distribution of examples with least square(LS)method,that is,minimize the distance between examples and boundary.Then twin support vector data description(TSVDD)is developed to classify the HRRP-based targets,and there will be a remarkable improvement of performance after banlancing the true positive rate and false positive rate.Numerical experiments based on publicly UCI datasets and HRRPs of four aircrafts are taken to compared TSVDD with other available approaches,the results especially for multiple targets can demonstrate the feasibility and superiority of TSVDD.Secondly,aiming at the oversized boundary of SVDD model when used for classifying HRRP-based targets,a novel boundary optimization algorithm is proposed.After constructing the geometric model of SVDD in the kernel mapping space,the decision function is simplified as an angle decision.The pruning support vector data description(PSVDD)is proposed,which adopts the least square support vector machine(LSSVM)to prune the boundary of SVDD.The pruning depth is optimized by LSSVM because LSSVM is constructed by the distribution of data.Then genetic algorithm(GA)is used to optimize the PSVDD model further by adjusting the pruning direction dynamically,and the adaptive pruning support vector data description(APSVDD)is developed to enclose targets and exclude outliers as many as possible.Numerical experiments based on publicly UCI datasets and HRRPs of four aircrafts demonstrate the feasibility and superiority of APSVDD.Thirdly,aiming at the limitation of HRRP examples,a novel data mining algorithm based on SVDD model is proposed.The learning paradigm named learning using privileged information(LUPI)is introduced into SVDD model by the correcting function,which is obtained from the original training with privileged information,and the second order central moment is selected as the privileged information.And then extended support vector data description with negative examples(ENSVDD)is proposed after incorporating the LUPI paradigm into SVDD with negative examples(SVDD-neg)model for dealing with negative examples.Numerical experiments based on publicly UCI datasets and HRRPs of four aircrafts demonstrate the feasibility and superiority of APSVDD not only with small training dataset but also under the condition of low signal-to-noise ratio(SNR).Fourly,aiming at dealing with data from multiple sources,multiple kernel learning(MKL)algorithms are proposed.The large margin support vector data description(LMSVDD)algorithm is analyzed theoretically,which takes the advantages of both SVM and SVDD.Several MKL algorithms instead of a single kernel are constructed on the basis of fixed rules,kernel alignment rules,reduced gradient rule and semi-infinite linear programming rule,named FSMKL,APMKL,ASMKL,ACMKL,ADMKL,RGMKL and SPMKL,respectively.Numerical experiments based on publicly UCI datasets including Ionosphere,Iris and Wine demonstrate the feasibility and superiority of MKL algorithms,when compared with single kernel learning.Finally,aiming at dealing with incremental examples efficiently and effectively when learning with HRRP-based examples,a novel incremental learning(IL)algorithm is developed.The incremental support vector machine(IncSVM)is used for elaborating the idea of incremental learning,according to the constant of Karush-Kuhn-Tucher(KKT)condition.Then the incremental large margin support vector data description(IncLMSVDD)is proposed,which learns the close boundary with one by one of incremental examples,that is,updates the model once the property of examples migrate.Numerical experiments based on Toy dataset,publicly UCI datasets and HRRPs demonstrate the superiority of efficiency when compared with basic classifiers,especially for large scale datasets.