Study On Radar Reconstructed High Resolution Range Profile Target Recognition

Radar Automatic Target Recognition(RATR)systems have ability to recognize the target attributes from the radar echo in all-time and all-weather conditions,and therefore widely applied in the military and civilian.The one-Dimensional High Resolution Range Profile(HRRP)has developed as a hotspot in the field of radar signal processing because of its simple imaging and low computing and storage costs.However,it always faces some difficulties in the HRRP target recognition in complex environment,which mainly display in the following.First,higher resolution provides abundant information of target.With the development of RATR systems,there is an urgent demand for improving the resolution of HRRP to enhance the target recognition performance.Unfortunately,the bandwidth restrict the improvement of resolution in the conventional imaging algorithm.Therefore it is necessary to develop the super-resolution techniques without additional extending the bandwidth.Furthermore,the complex electromagnetic scattering characteristics of target leads the statistical properties of HRRP exhibits obvious non-stationary.Hence,it is necessary to extract effective features to describe the local time-varying behaviors of HRRP.Finally,the design and selection of the c lassifier is also a problem to RATR systems that cannot be ignored.This dissertation revolves the above problem and is major in the following aspects.The surface weak target detection method based on one-Dimensional frequency-domain features is discussed.First,this dissertation analyses the PSD feature of sea clutter signal,and points out that the non-stationary characteristic of sea clutter cannot be simply summarized by PSD.Moreover,the DPS and NPS features,which originally proposed for the speech signal processing,are introduced on basis of PSD to the radar surface weak target detection community.In the detection phase,an optimal choice can be determined by the PSO-SVM.Simulations are demonstrated to evaluate the detection performance with DPS and NPS features.The simulation results show that both DPS and NPS are effective for the surface weak target detection,which can combine the magnitude spectrum and phase spectrum,then preserve more details of the spectral information and effectively suppress random fluctuations.PKA-ESPRIT parameter estimation algorithm is proposed on the basis of the ESPRIT andU-ESPRIT algorithm,and its applications in one-Dimensional scattering center extraction and HRRP reconstruction are analyzed.The main idea of PKA-ESPRIT is to introduce a scale factor,which could represent the influence of the cross-term on the estimation of autocorrelation matrix,to modify the autocorrelation matrix of the observation signal in cases of fewer samples.Finally,the HRRP reconstruction method for complex target is discussed,and the estimation performance of scattering center number by using Gerschgorin Disk Estimator in the situation of fewer snapshots is analyzed.A series of efficient features are extracted from HRRP signal by using time-frequency analysis techniques.This dissertation reviews a sample of representative time-frequency analysis algorithms.Their performance is studied from a qualitative and quantita tive point of view.For simplicity,we considered the mean-squared error and the effective frequency resolution as measures of performance in the quantitative trade-off studies.It is shown that AOK distribution considerably increases the detectability of signals while suppressing artifacts and noise.Unfortunately,the time-frequency distribution produces a great deal of redundant information.Hence,the normalized frequency marginal feature is extracted to reduce dimensions of the TF discriminant features,which is necessary to improve the efficiency of pattern classification.Finally,a multilayered feed-forward NN is selected as a classifier in the pattern classification process.The effectiveness of the proposed method is demonstrated by using experimental HRRP data and simulated data.To improve targets recognition performance of RATR systems,the dissertation put forward a novel time-frequency features fusion strategy assisted by population evolution algorithm.A Volterra-series-based weighted averaging model is utilised as the fusion rule to construct the fused time-frequency feature,which aims to enhance the performance of time-frequency analysis and suppress signal-dependent cross-term artefacts.Herein,the optimal fusion coefficient is estimated by culture-based population evolutionary algorithm without any prior information.Finally,a multi-layered feed-forward NN is utilized as a classifier in the pattern classification process.Experimental results demonstrate that the fused time-frequency feature constructed by the proposed scheme achieves better performance in sharpness and strength than any subset of time-frequency features or their combinations and,furthermore,increases the separability among various classes of target and improves the target recognition performance.The architecture of CNN and its applications in time-frequency features algorithmic fusion and ISAR image classification is researched.First,the trade-off studies on the CA-based optimal fusion coefficient estimation and CNN process is present.It is shown that CNN architecture considerably improves speed of fusion coefficient estimation process while preserving all the effective information.Besides,the CA evolutionary strategy is applied to the SARPROP algorithm to balance the local searching and the global searching at the initial stage,which largely improves the convergence properties of CNN training process.Finally,the ISAR image contour extraction method based on U-ESPRIT and convex hull detection algorithm is discussed.The CA-SARPROP trained CNN is employed to classify the target according to its contour feature.