Research On Radar HRRP Target Recognition Based On Hybrid Generative Discriminative Models

Radar automatic target recognition(RATR)is a useful technology which combins radar technology and pattern recognition.RATR has very important application in civilian and military fields,thus RATR has received intensive attentions.Since radar high resolution range profile(HRRP)contains abundant target structure information and is easily obtained and stored,it has become the focus of research in RATR field.The models of RATR can be roughly divided into two categories,namely,generative models and discriminative models.Generative models mainly focus on the distrbution of data,while discriminative models usually pay more attention to the differences and the hyperplane between different classes data.The hybrid model which combins generative and discriminative models can enhance the recognition accuracy and have all benefits of both these models.This thesis focuses on HRRP recognition from Bayesian statistic learning,max-margin regularized model and Bayesian non-parameter technology.etc.The main research efforts are summarized as follow.1.As an unsupervised probabilistic generative model,factor analysis(FA)has a great ability of data description,while FA model doesn't consider the difference between different classes samples.To improve the recognition performance of FA model,max-margin factor analysis(MMFA)model proposed to combine FA and latent variable support vector machine(LVSVM),and build the max-margin constraint in latent subspace.In MMFA model,the latent variable is seen as the input of LVSVM,which will generate some information loss.Meanwhile,the model projects all training data into a same lower-dimensional subspace for classification.This may be unsuitable for the situation of HRRP related task.To solve the above problems,we propose max-margin regularized factor analysis(MMRFA)model.MMRFA model combines the FA model and LVSVM in the original space and establishes max-margin constraint for the reconstructed feature from FA.MMRFA model can improve the recognition accuracy because it contains almost all information for classification and guarantee the separability of reconstructed feature.Furthermore,to tackle the model selection problem,we propose max-margin regularized Beta process factor analysis(MMRBPFA)which combines Beta process factor analysis(BPFA)and LVSVM.The data description will be more precise by using MMRBPFA,and the reliability and extension of MMRBPFA are improved.2.The radar HRRP data is nonlinear and complexly distributed in many practical cases.Here we propose Bayesian kernel support vector machine(BKSVM)model and Dirichlet process Bayesian kernel support vector machine(DPBKSVM),and the complete recognition system based on the models are shown in dissertation.BKSVM generalize kernel support vector machine into Bayesian framework.We transformed the optimization problem of kernel SVM to calculate the posterior of BKSVM by assuming the probability distribution of the lagrange multipliers and employing the augmentation latent variable technology.BKSVM provides the possibility of combing kernel SVM with other generative models.Moreover,to enhance the recognition accuracy based on nonlinear and complexly distributed data sets,we proposed Dirichlet process Bayesian support vector machine(DPBKSVM)which combined Dirichlet mixture model and BKSVM.DPBKSVM establishes a BKSVM classifier on every cluster subset,and the cluster results are supervised by the classifier results.DPBKSVM not only can describe the data distribution construction very well but also has a great ability of nonlinear classification,thus DPBKSVM perform inspiringly in target recognition.