Study of Radar High Resolution Range Profile Target Recognition Based on Statistical Modeling
|School||Xi'an University of Electronic Science and Technology|
|Course||Signal and Information Processing|
|Keywords||Radar automatic target recognition (RATR) High resolution rangeprofile (HRRP) Noise robustness Small sample size Statistical modeling Bayesian Ying-Yang harmony learning Automatic model selection|
With the application of information and intelligence technologies in modernwarfare, there is an urgent need for Radar Automatic Target Recognition (RATR).Wideband radar can provide abundant target structure information, and thus hasreceived intensive attention from the RATR community. The general wideband signalsinclude High Resolution Range Profile (HRRP) and Synthesized Aperture Radar (SAR)image or Inverse SAR image. Since the former is easily obtained and computationallyefficient, HRRP based recognition is more attractive in many circumstances. Thisdissertation focuses on HRRP recognition from feature extraction, statistical modelingand noise robustness, etc. The main research efforts are summarized as follows.1. The first part analyzes the target-aspect sensitivity of HRRP and proposes a newframe partition method which is based on subspace modeling and Kullback-Leiblerdistance. Experimental results show that the new method can allocate those samplesfollowing the similar statistical distribution into a same frame, and hence decrease thetotal frame number of all targets and improve the final recognition performance.2. The second part focuses on noise robust recognition. In many literatures, it wasassumed that the training and testing samples were measured under high signal-to-noise(SNR) ratio conditions, and the effect of noise was ignored. However, in practicalapplications the testing samples are unavoidably contaminated by noise, andconsequently the SNR mismatches between training and testing samples will deterioratethe recognition performance. We firstly make a review of existing methods and thenpropose a new framework for noise robust recognition. In the training phase, we learndifferent models for different SNRs by adding noise into the clean samples, while in thetesting phase we choose appropriate models for recognition based on the estimated SNRvalue of testing sample. Experimental results show the new framework can considerablyimprove the recognition performance under low SNR conditions.3. In the third part, feature extraction from HRRP is discussed. HRRP ischaracterized by high-dimensionality and four sensitivities, which pose hurdles forHRRP based target recognition. After analyzing the statistical property of HRRP’sfrequency spectrum amplitude (FSA), we propose to model FSA by Autoregressive (AR)model and extract the AR coefficients and partial correlation coefficients as recognitionfeatures. Both the two features are low-dimensional and invariant to translation and amplitude-scale changes of HRRP, and more importantly, they preserve most of thestructure information of FSA. In addition, to tackle the remaining target-aspectsensitivity, we propose a Gaussian mixture model based frame partition method whichcan determine the frame number automatically and guarantee the distributionconsistency of samples in each frame.4. The fourth part manages to fulfill the recognition task under small trainingsample size condition. Both the high feature dimensionality and limited sample sizemake HRRP based recognition a typical small sample learning problem. To cope withthis, we adopt FSA as recognition feature and model it along the frequency dimensionby sequential models. In this way, we relax the heavy requirement of training samples inHRRP recognition. Firstly, assuming the FSA components are Gaussian distributed, weemploy Linear Dynamic model to describe FSA. This model can well capture thestationarity of FSA and thus obtain satisfactory recognition performance. Afterwards,experimental results based on measured data reveal that the Gaussian assumption isinappropriate, hence we further make a study on more accurate statistical modeling.Mixture Autoregressive model is thus introduced to describe the stationarity andmulti-modality of FSA simultaneously which show incremental improvements onrecognition accuracy. The models mentioned above have low degrees of freedom and alltheir parameters can be estimated via a single sample, which collectively make theirrecognition performance robust to the variation of training sample size.5. In the fifth part, we investigate accurate statistical modeling and model selectionproblem. HRRP samples are temporally dependent and jointly non-Gaussian distributed,the statistical modeling of which is a challenging task for HRRP based recognition. Ourmain work includes:1) adopting Local Factor Analysis model to describe thenon-Gaussian property and inter-dimensional dependency of HRRP;2) employingTemporal Factor Analysis model to describe the spatio-temporal structure of HRRP.Since both the two models can describe HRRP more accurately, they achieve betterperformance than traditional models. Furthermore, the conventional two-phaseapproach for model learning suffers from intensive computation burden and unreliableevaluation. To tackle these problems, we adopt Bayesian Ying-Yang (BYY) harmonylearning that has automatic model selection ability during parameter estimation.Experimental results illustrate that the BYY learning can significantly decrease thecomputation burden while improving the accuracy of parameter estimation and modelselection.6. The last part demonstrates the feasibility of recognition using complex HRRP. Based on the theoretical analysis, it is pointed out that the variation of initial-phase hasno effect on the statistical distribution of HRRP. So it is natural to generalize the FAmodel to the complex domain and model complex HRRP directly. By using additionalphase information in complex HRRP, the complex FA (CFA) model is superior to the FAmodel in recognition performance. Also, we develop a noise adaptive modificationalgorithm for the CFA model. Since there is no approximation made in the modificationstep, this algorithm can gain better performance under low SNR conditions.