Study On The Noise Robustness And Small Training Size Problems In Radar HRRP Target Recognition

Radar high resolution range profile(HRRP)is the amplitude of the coherent summation of projection vectors of the complex echoes from target scatterers along the radar line of sight(LOS),which contains the information of target's size and geometry structure.Compared with the synthetic aperture radar(SAR)and inverse synthetic aperture radar(ISAR),HRRP is more easily acquired and processed,which makes it a research hotspot in radar target recognition.This dissertation focuses on the robust problem in radar HRRP recognition,which includes the noise robust problem and the recognition robustness with small training size.Following summarizes two main work of this dissertation :The first part is about the noise robust problem in radar HRRP recognition.To solve noise robust problem,in second chapter we propose a denoising method based on sparse representation of HRRP data and develop a scatterer matching recognition algorithm for the denoised data.Based on the scattering center model,the dominant scattering centers sparsely spread over the Fourier dictionary,in which the echoes can approximate the target signal;meanwhile the weak scattering centers spread uniformly over the Fourier dictionary,which correspond to the noise components.According to the sparse representation theory,such a target signal can be stably recovered from the noisy measurements.Here the scattering coefficients and locations of the dominant scattering centers are first estimated from a noisy test sample by solving the sparse optimization problem with the noise level constraint,then the scatterer matching algorithm based on Hausdorff distances(HDs)between its dominant scatterers and those from the training templates is used to distinguish the unknown test sample.Another method to solve the noise robust problem is to construct the model in complex domain and then to update the parameters of the statistical model according to the noise power in test sample.To solve this issue,a statistical model based on multi-task leaning complex factor analysis and similarity preserving(MTL-CFA-SP)is proposed in third chapter.The model is constructed in complex domain,where the noise problem can be easily processed.The second work in this dissertation is the robust recognition with small training size.To solve the robust recognition with small training size,a statistical model based on multi-task leaning(MTL)and complex factor analysis(CFA),referred to as MTL-CFA,is proposed in third chapter.The statistical modeling of each training aspect-frame is considered as a single task,and all tasks share a common loading matrix.The factor number of each task is automatically determined via the Beta-Bernoulli sparse prior.To further improve the recognition performance with small training size,a statistical model based on multi-task leaning complex factor analysis and similarity preserving(MTL-CFA-SP)is proposed.Using a similarity preserving constraint term we enhance the separability of the statistical models for different classes.Experimental results based on measured data show that the proposed model MTL-CFA can not only describe the observed data with lower order of model complexity,but also obtain satisfactory recognition accuracy with small training size data.Compared with the traditional single-task learning(STL)based on FA model,the proposed model,MTL-CFA-SP,has better recognition performance with small training size data.To dig deeper about the recognition with small training size,a hybrid model based on the Multi-Task Learning Hidden Markov Model and Deep Neural Network(MTL-HMM-DNN)is proposed in the fourth chapter.Instead using the Gaussian distribution,the Deep Neural Network is used to describe the statistical distribution of the amplitude of the range profile.The experiments based on the toy data show that MTL-HMM-DNN is more suitable to fit the data which has the non-Gaussian observation probability in Hidden Markov Model.Results based on the measurement data show that MTL-HMM-DNN enjoys superior recognition performance with small training size.