Study On Unknown Target Discrimination Based On SVDD Using HRRP

The prerequisite for the recognition of the High Resolution Range Profiles(HRRP)is to obtain the feature data of the target and establish the target feature database.In practical applications,the target to be identified may be an unknown target,that is,a target that does not participate in training.Due to the lack of a feature library template corresponding to the unknown target,the traditional identification method will result in an incorrect determination result.Therefore,before performing the conventional target recognition,it is necessary to perform the discrimination of the unknown target to determine the library attributes of the target to be identified.This dissertation conducts research on the method of discriminating unknown targets.Specifically include:1.Support Vector Domain Description(SVDD)is easy to over-fit or under-fit when constructing classification boundaries.Therefore,this dissertation studies two SVDD discriminating methods based on the distribution of training samples.By calculating the distribution parameters of the training data samples and using them as weighting factors to improve the classifier’s optimization function,the classification boundary can be adjusted in order to avoid overfitting or underfitting,particle swarm parameter optimization method is adopted to solve the problem that the parameters in the training process are difficult to determine.Simulation results show that the two discriminant methods based on sample distribution have better results than SVDD in terms of discriminant effect,anti-noise,and azimuth sensitivity.2.The SVDD requires the training data to satisfy the spherical distribution as much as possible,but most of the measured data cannot be satisfied.Therefore,this dissertation studies three clustering-based SVDD discriminant methods.This method first clusters the training dataset according to the corresponding clustering criteria,then performs SVDD training on each subclass,obtains corresponding boundary support vectors,and finally performs SVDD training on the support vector set again to obtain the final decision boundary.This kind of method considers the distribution of the sample and uses multiple SVDD trainings,which improves the accuracy of the discrimination.At the same time,Differential Evolution(DE)is used to solve the parameter setting problem of the discrimination method.The simulation results show that the above method has relatively good results in terms of discriminant effect,anti-noise,and azimuth sensitivity.3.Since the traditional SVDD can only use one type of sample data for training,when it is able to obtain part of the unknown target information,it cannot be used.However,the NSVDD does not consider the distribution of training data,and it easily leads to the decision boundary being tilted to a certain class.Therefore,a method based on the kernel density weights is studied in this dissertation.This method computes the density distribution of samples in the nuclear space around the NSVDD classification boundary,and uses them as weighting factors to improve the optimization function,so that the boundary of the structure is compact and smooth.At the same time,a new parameter optimization method is proposed,which solves the difficulty of parameter setting in the method and the problem that other optimization algorithms easily fall into local optimum.Simulation experiments verify the feasibility and effectiveness of the above method.