Study Of Radar Automatic Target Recognition Based On Sparse Bayesian Learning

Radar has the ability of all-time, all-weather, and long range detection, which gains substantial development and application in various civilian and military fields. With the evolution of the modern radar technique, we have more demands for radar, in which radar automatic target recognition (RATR) is the very important one. Usually, RATR is non-cooperative target recognition, which extracts the feature information of the target echo, and discriminates the classes and attributes of targets by utilizing some judgement criterions. RATR consists of five procedures:data acquisition, data preprocessing, feature extraction, feature selection and classification. Therefore, the related theory and techniques for RATR are researched from four aspects, i.e. data preprocessing, feature selection, classifier design and the classification with small training data size by utilizing the Bayesian analysis and inference methods. Based on the Bayesian theorem, the Bayesian methods can be used to elaborete and solve the statistical problems systematically. In addition, the Bayesian inference methods can obtain the posterior information by combining the prior information and sample information, and the posterior information can be used to infer the unknown parameters. The main four aspects of this dissertation is summarized as follows:1. This part focuses on realizing the parameter estimation (super resolution) of the complex radar echo rapidly and accurately. Most of the existing sparse representation methods estimate signal parameters via a dictionary with atoms defined on the discrete grid in parameter space, if the grid is coarse, there may be mismatch between the setting parameters and the true parameters. If the grid is fine, the computational complexity would increase significantly. Motivated by this, a sparse Bayesian representation with refined-dictionary, is developed for parameter estimation of complex signal in this dissertation. In the Bayesian model, the parameter estimation is achieved via selecting the parameterized atoms in a given dictionary where a Bernoulli-Beta prior is used to promote sparsity in the utilization of atoms. Then the dictionary grids can be refined by utilizing of the weighted clustering and zooming to learn the parameters more precisely, which are used to generate the dictionary. Finally the sparse Bayesian representation model is utilized to selected the parameters again. Operating the two procedures iteratively until convergence, and the repaid and accurate parameter estimation can be realized.2. This part discusses realizing the reconstruction of incomplete frequency data by utilizing the full-band frequency data. It usually suffers from long observing time and interference sensitivity when a radar transmits the high-range-resolution stepped-frequency chirp signal. Motivated by this, only partial pulses of the stepped-frequency chirp are utilized. For the obtained incomplete frequency data, a Bayesian model based on transfer learning is proposed to reconstruct the corresponding full-band frequency data. In the training phase, a complex beta process factor analysis (CBPFA) model is utilized to statistically model each aspect-frame from a set of given full-band frequency data, whose probability density function (pdf) can be learned from this CBPFA model. Importantly, the numbers of factors and dictionaries are automatically learned from the data. The inference of CBPFA can be performed via the variational Bayesian (VB) method. In the reconstruction phase, for the incomplete frequency data that "related" to the training samples, whose corresponding full-band frequency data can be analytically reconstructed via the compressive sensing (CS) method and Bayesian criterion based on the transfer knowledge of the previous pdfs learned from the training phase. About the "relatedness" between each training frame and the incomplete test frequency data, we utilize the frame condition distribution of incomplete test frequency data to represent.3. A Bayesian classifier for sparsity-promoting feature selection is developed in this part, where a set of nonlinear mappings for the original data is performed as a pre-processing step. The linear classification model with such mappings from the original input space to a nonlinear transformation space can not only construct the nonlinear classification boundary, but also realize the nonlinear feature selection for the original data. A zero-mean Gaussian prior with Gamma precision and a finite approximation of Beta process prior are used to promote sparsity in the utilization of features and nonlinear mappings in our model, respectively. We derive the variational Bayesian (VB) inference algorithm for the proposed linear classifier. Experimental results based on the synthetic dataset, measured radar dataset, high-dimensional gene expression dataset, and several benchmark datasets demonstrate the aggressive and robust feature selection capability and comparable classification accuracy of our method comparing with some other existing classifiers.4. Two recognition methods based on multitask sparse learning is proposed to realized the recognition of high resolution range profile (HRRP) with small training data size. Usually, two distinct decision approaches can be used to recognize the HRRP:generative models and discriminative models. Generative models are the statistically modeling methods based on Bayesian theory, while discriminative models find a discriminant function in the training stage, which maps each input sample directly onto a class label. Both the decision approaches adopt the multitask sparse learning, which are discussed respectively as follows. (1) the multitask learning of the generative models. The statistical modeling of each training aspect-frame is considered as a single task in this method. Since the training aspect-frames are not independent but inter-related, they can share a compact dictionary to make full use of the information. However, with the different targets and the aspect sensitivity of the same target, it is usually hard to assess the task relatedness and joint learning with unrelated tasks may degrade the recognition performance, therefore, we adopt the Bernoulli-Beta prior to learn the needed atoms of each aspect-frame automatically with the given training data. Then the relatedness between frames is determined by the number of shared atoms, and multitask learning can be realized adaptively. (2) the multitask learning of the discriminative models. This method is suitable for the multi-category classification, and the classification of one class versus others is considered as a single task. Firstly, we assume the tasks are related, i.e. the task predictors (classification weights) are related. Then, the task predictors may belong to a low dimensional sunspace to realize the information sharing. Since the task predictors usually have varying degree of relatedness, the subspace bases of each task predictor can be leaned by utilizing the Bernoulli-Beta prior from the given data. And the number of shared bases between different tasks can be utilized to determine their relatedness, and the adaptive multitask learning is realized.