Research On Feature Extraction And Recognition Performance Enhancement Algorithms Based On High Range Resolution Profile

A high range resolution profile(HRRP)is the amplitude of the coherent summations of electromagnetic echoes from target scatterers.It provides rich geometric structure information of a target along the radar line-of-sight,and this information is useful for target recognition.HRRP-based radar automatic target recognition(RATR)is one of the most important ways for target identification.However,a HRRP is the one-dimensional projection of the three-dimensional scattering structure onto the radar line-of-sight.The projection may be not sufficient for confident decision in some situations.Focused on the issue,the polarization characteristic and multiple views of a target are utilized to boost recognition performance in the dissertation.The main research work is summarized as follows:In Chapter 1,the research background and significance of the dissertation are addressed.Then the state-of-the-art development about HRRP-based RATR is surveyed.Lastly,the main work and the organization of the dissertation are introduced.In Chapter 2,the feature extraction for HRRP-based RATR is studied.Firstly,the features used for target recognition are summarized and catalogued.It is pointed out that a commom drawback of the conventional methods is the spatial distribution of the extracted features is complicated and unpredictable,which makes the design and implementation of the feature-based classifier still a tough problem.To handle the issue,a novel feature extraction method based on sequential vanishing component analysis(SVCA)is proposed.The linear separability of the SVCA features has been proven by theoretic analysis,thus target identification can be efficiently achieved by linear classifiers.Experiments are carried out on simulated data and MSTAR database,and the results demonstrate the effciency of the proposed method.In Chapter 3,polarization information is used to improve recognition performance.Firstly,the characteristic of a full-polarization HRRP(FPHRRP)is analyzed.It is pointed that the HRRPs under different polarization combinations are correlated,and each entry is associated with the same target pose.Both the two aspects can serve as prior information to boost recognition performance.To efficiently utilize the information for target recognition,a joint sparse representation(JSR)based method for FPHRRP recognition is presented.The presented method assumes all the entries within a FPHRRP sample share a common sparsity pattern in their sparse representation vectors at atom-level,which has a significant advantage of exploiting the aforementioned information to enhance recognition performance.The effciency of the proposed method is demonstrated by the recognition results using simulated database.In Chapter 4,the utilization of multiple views for recognition performance improvement is investigated.Two applied situations are under consideration: multiple views are captured from a small aspect interval(Situation-I)and multiple views are obtained without pose constraint(Situation-II).For Situation-I,the multiple views are strongly correlated.To efficiently utilize the relations among the multiple views,a novel multitask compressive sensing(MtCS)-based method for multi-view RATR is presented.In Situation-II,the multiple views may be quite different.To accurately model the relationship among the multiple views,a joint dynamic sparse representation(JDSR)based method is introduced.The method supposes all the sparse representation vectors associated with each single view share a common sparsity pattern at class-level,which has the superiority of utilizing the different correlation among multiple views for target identification.The recognition results using simulated data and MSTAR database demonstrate that both of the methods are efficient.In Chapter 5,both polarization information and multiple views of a target are used for enhancing HRRP-based recognition performance.Motivated by the previous chapter,the recognition tasks under two particular situations are considered: multi-view FPHRRP samples are captured consecutively in a small aspect interval(Situation-III)and multi-view FPHRRP samples are obtained randomly from full aspect angles(Situation-IV).For the previous situation,the atom-level sparsity is used to model the relations among the sparse representation vectors associated with each entry within a multi-view FPHRRP sample.For the latter one,a novel hierarchical joint sparse representation(HJSR)model is proposed to formulate the recognition task.A joint sparsity is utilized to model the sparse representation vectors obtained from the entries within a single view and a joint dynamic sparsity is applied across the sparse representation vectors associated with the multiple views under the same polarization mode.The proposed method can not only utilize the prior information that the single-polarization HRRPs within a single view are associated with the same target pose,but also has the advantage of exploiting the relations among multiple views under the same polarization mode for target recognition.Experiment results using simulated database show the validity of the proposed methods.In Chapter 6,the main research work of this dissertation is concluded,and some suggestions are given for the future investigation.