Research On The Nonlinear Features Based MMW HRRP Recognition Methods

Radar automatic target recognition(RATR)technique is one of the key technologies in the application of short-range detection and accurate guidance.As an important signal in RATR,High resolution range profile(HRRP)represents the projection of the target scattering centers on the radar line of sight(RLOS)and reflects the characteristics of the target structure,with the advantage of obtaining and processing easily.It is easier for a millimeter wave(MMW)radar to transmit a wide-band signal,which improves the range resolution and obtains more details of the target.Thus,the result of target recognition can be more accurate.However,the HRRP is influenced by the radar parameters,target azimuth,target posture,the background,the weather and other factors,which make the HRRP show highly nonlinear.The traditional linear feature extraction methods can't extract the discriminant features hiding in HRRP effectively and the recognition accuracy is not satisfactory.Kernel method uses a nonlinear mapping to map the linear inseparable samples from the sample space to a high-dimensional feature space in order to achieve linear separable,it has obvious advantage,when dealing with the nonlinear problem.Manifold learning is one kind of nonlinear dimensional reduction technologies which have been studied widely.The low dimensional feature structure can be obtained from the high dimensional nonlinear space through manifold learning.In this paper,we mainly focus on the problem of manifold learning and kernel method based three ground targets MMW HRRP recognition.Issue on sparse representation classification,feature extraction,design of the clustering objective function and removal of the sample redundancy have been studied.The main achievements of this work are listed as follows.To overcome the target-aspect sensitivity problem in HRRP recognition,we studied the dictionary learning algorithm and sparse representation classification algorithm,and then we proposed a novel segment sparse representation classification algorithm on the basis of fast dictionary learning.Based on the high correlation of HRRP samples in adjacent azimuth range,we imported the effectiveness parameter and correlation parameter to improve the dictionary learning.Since the target useful information only occupies a part of all the features in HRRP sample,we segmented the HRRP asymmetrically and calculated the sparse reconstruction error weight of each part based on the length of range resolution cell and energy.By performing sparse representation on each parts of sub-HRRP and searching the minimum weighted reconstruction error of the segment sparse representation,we can obtain the recognition result.We studied the problem of nonlinear feature extraction from the view of algorithm and proposed two feature extraction algorithms,called supervised discrimination sparse neighborhood preserving embedding(SDSNPE)and semi-supervised kernel adaptive marginal fisher analysis(SKAMFA).Based on the similarity among the objective functions of three manifold learning algorithms,we merge the three objective functions to propose the SDSNPE algorithm.SDSNPE not only has the same ability to deal with the nonlinear problem with the manifold learning algorithms,but also can preserve the local structure,global structure and the sparse reconstruction relationship among the HRRP samples effectively and simultaneously.Using the kernel function,SKAMFA gives the scatter weight to each sample adaptively based on the similarity measurement and constrains the within-class scatter and between-class scatter at the same time to deal with the nonlinear feature extraction of HRRP,which solves the contradiction of preserving geometrical structure and extracting the nonlinear features perfectly.In order to avoid the problem of hybrid overlap and the drop of clustering accuracy caused by the nonlinear and distribution of HRRP,we studied the HRRP clustering problems from the view of designing new objective function and proposed the linear discriminant kernel fuzzy c-means clustering(LDKFCM)algorithm.We used the kernel functions to map the HRRP samples to a high-dimensional feature space and used the objective function to constrain the within-cluster scatter and between-cluster scatter at the same time.It avoided the hybrid overlap clustering effectively while the nonlinear feature is extracted effectively.For the choice of the optimal parameters,we give out the parameter selection methods based on the division objective function optimization and the fuzzy decision theory.In practical engineering applications,as to remove the redundancy caused by the hardware oversampling and determine the start point of the target extraction,we proposed a peak searching and reverse threshold choosing-large algorithm.We first found the position of the peak through peak searching,following by reversely calculating the extraction starting point,and then acquired the HRRP without redundant via the classical choosing-large algorithm.