Study Of Radar Target Recognition And Outlier Rejection Based On High Range Resolution Profiles

With the tendency of modern battle to become more and more information-based and intelligent,radar automatic target recognition(RATR) techniques have received intensive attentions. Radar high-resolution range profile(HRRP) denotes the amplitude of coherent summation of complex time returns from target scatterers in each range cell. HRRP contains lots of informative target structure information i.e. target size, scatterers’ distribution, etc. Also it has such advantages as easy acquisition, small storage needed, etc. Thus, HRRP based target recognition has received intensive attention in the RATR community. This dissertation focuses on HRRP recognition from target recognition and outlier rejection, feature extraction with max-margin classifiers and Bayesian nonparametrics techniques, etc. All the researches are supported by Advanced Defense Research Programs of China and National Science Foundation of China. The main four aspects of this dissertation can be summarized as follows.1. Due to the incompleteness of template library, it is important to consider both the recognition and rejection performance when evaluating an RATR system. We propose to utilize the output of a common classifier as the input of a nearest neighbor(NN) classifier to identify outliers, and refer to such a system as a ‘classifer-NN’ system. To improve the system performance, several ‘classifier-NN’ systems are combined and a cost function is defined to measure the recognition and rejection performance of RATR system. Based on the cost function, two algorithms are developed to select the optimal work point of each “classifier-NN” system. This radar target recognition and rejection system can take the advantages of the diversity information from different fatures and classifiers to improve the recogniton and rejection perfomance of the whole system.2. The high-dimensional HRRP data often contains redundant information, which is disadvantageous for RATR. Here we propose two max-margin feature extraction models---max-margin factor analysis(MMFA) model and max-margin Beta process factor analysis(MMBPFA) model, and build an integrated target recognition and rejection scheme. Factor analysis(FA) is a multivariate correlation model which reveals the low-dimensional latent space structure and has been used extensively in data reduction, classification and description. As a generative model, FA only focus on the observations without utilization of any label information, which may lead to the low-dimensional latent features not suitable for the following prediction task. Thus, we propose the MMFA model, which takes advantages of the latent variable support vector machine(LVSVM) to learn a discriminative subspace with max-margin constraint. MMFA jointly learns the feature space and the max-margin classifier in a unified framework to improve the prediction performance. Furthermore, to tackle the model selection problem, Bernoulli-Beta prior is introduced into MMFA and the MMBPFA model is proposed. MMBPFA can capture the latent feature of data, determine the number of factors automatically, and utilize label information simultaneously. Moreover, it is natural for MMFA and MMBPFA to handle outlier rejection problem, which benefits from the data description ability of FA.3. When dealing with large-scale and complexly distributed data, training a classifier using all input data will be very expensive and the underlying structure of the data will be ignored. To overcome these limitations, the mixture-of-experts(ME) system is proposed, which partitions the input data into several clusters and learns a classifier for each cluster. However, in traditional ME systems, the number of experts should be fixed in advance, and the clustering procedure and classification task are de-coupled. To deal with these problems, dp LVSVM, a Dirichlet process mixture of latent variable SVM, is proposed. In the dp LVSVM model, the number of cluster is chosen automatically by Dirichlet process mixture model,and linear latent variable SVMs(LVSVM) are employed in each clusters. Different from previous algorithms, the clustering procedure and LVSVM are jointly learned in dp LVSVM to gain infinite discriminative clusters. And the parameters can be inferred simply and effectively via Gibbs sampling technique.4. The performance of a classification model depends on the assumptions to estimate the parameters in the model. On one hand, in MMFA, the two assumptions, all dataset follow a single Gaussian distribution and the classifier is linear, cannot deal with multimodally distributed data. On the other hand,dp LVSVM works in original space, which is not practical to handle high dimensional data without feature extraction procedure. Therefore we combine MMFA and dp LVSVM model and propose an infinite max-margin factor analysis(i MMFA) model. With the idea of ME, the i MMFA model divides data into ‘infinite’ clusters via Dirichlet process(DP) mixture model in the low-dimensional latent space and meanwhile learns a linear max-margin classifier on each cluster to construct a highly nonlinear classifier. Furthermore, we introduce nonparametric Bayesian technique into the i MMFA model to deal with model selection problem and propose the i MMBPFA model. These two models can unify clustering, model learning and max-margin classifier designing under Bayesian framework, and exhibit superior performance in both data description and discrimination.