Research On High Resolution Range Profile Feature Extraction And Recognition Algorithm Based On PCA-LPP

Radar High Resolution Range Profile(HRRP)is the vector sum of the projection of the target scattered point echoes in the radar line of sight direction obtained from broadband radar signals,which can reflect the distribution of the equivalent scattering center in the radar line of sight direction.HRRP contains rich information about target physical structure,and has the relatively simple processing and small storage capacity for other broadband radar imaging,which has received wide attention from researchers.In this paper,we address the problems of existing HRRP feature extraction and target recognition methods,and focus on linear and nonlinear feature extraction,classification algorithm fusion,and classifier parameter optimization.The main research work of the paper is as follows.(1)To address the problems of nonlinear implicit features between HRRP range units and high feature dimensionality,we focus on the extraction of useful features and feature dimensionality reduction,and propose a PCA-LPP feature extraction method by fusing Principal Component Analysis(PCA)and Locality Preserving Projections(LPP).The method first constructs the overall objective function by integrating the global and local objective functions,and then applies the supervised similarity matrix to achieve the objective function optimization.The solution of the overall objective function provides a feature mapping matrix for feature space transformation to obtain low-dimensional feature data.By comparing experiments with other traditional feature extraction methods,it is verified that PCA-LPP makes the feature separability of three different targets stronger,even for very similar two types of targets,the recognition degree is increased,and the feature points of the reverse reconstruction profile match better with the original range profile.The results show that the HRRP features extracted by PCA-LPP are more beneficial to target recognition than other traditional methods.(2)For the problem that the training data of noncooperative targets are limited and the feature spaces are overlapped and indistinguishable,this paper introduces Support Vector Machines(SVM)combined with PCA-LPP for target recognition.The method first obtains highly separable features after PCA-LPP feature extraction,then combines SVM to map feature samples to Gaussian kernel space to improve the classification performance of nonlinear samples,and finally obtains the recognition results after combining PCA-LPP and SVM.The recognition comparison experiments with traditional classifiers verify that SVM and PCA-LPP have better fusion effect;the analysis experiments of training sample size verify that the recognition effect of the method is improved more significantly with relatively low bandwidth and relatively small number of training samples.This indicates that the proposed method still has better applicability in the case of small sample size and poor sample quality.(3)When the training samples change,how to set the kernel function parameters and penalty factors of SVM to achieve the best recognition accuracy is a problem that is difficult to specify in advance.To address this problem,this paper proposes to introduce the Bald Eagle Search(BES)optimization algorithm to SVM classification recognition.The method obtains the optimal SVM parameter values through three major stages: selection,search,and swoop.By setting up convergence time comparison experiments and recognition comparison experiments,it is verified that BES-SVM has shorter convergence speed and higher recognition accuracy than the traditional optimization algorithm.