基于稀疏学习的雷达目标识别方法研究

发布时间：2018-04-18 07:46

本文选题：雷达目标识别 + 稀疏学习　；参考：《南京航空航天大学》2016年硕士论文

【摘要】：雷达目标识别技术是基于雷达回波信号,提取与目标特性相关的信息,实现目标属性或类别的判定。随着国际形势的发展,雷达目标识别越来越受到世界各国科研人员的青睐。随着高分辨率雷达体制的应用,使得获得更为细致的目标几何结构信息和细节信息成为可能,而高分辨率雷达(High Resolution Radar,HRR)和合成孔径雷达(Synthetic Aperture Radar,SAR)两种体制的雷达的回波信号HRRP和SAR图像,作为典型的高分辨率雷达信号,也成为当前各国雷达目标识别研究的热点。本文在稀疏学习理论基础上,研究基于HRRP目标和SAR图像目标的雷达目标识别方法,主要的研究工作如下:1.研究了稀疏学习理论。首先,对三种典型的稀疏建模方式、三类经典的稀疏求解方法、以及稀疏学习的应用进行了阐述;其次,分别研究了HRRP目标和SAR图像常用的稀疏表示方法,并对其稀疏性进行了分析。2.提出了一种基于贝叶斯模型的Shearlet域SAR图像去噪算法,所提出的算法既利用了稀疏系数间的空间相关性,又基于贝叶斯模型获取了动态的噪声阈值,在实现噪声滤波的同时可以有效的保持边缘信息。首先对对数变换后的SAR图像进行Shearlet稀疏表示,其次根据稀疏系数的统计特性利用贝叶斯模型进行噪声检测的建模,最后利用自适应加权收缩实现SAR图像噪声像素的平滑处理。在MSTAR数据库上的实验结果验证了所提方法的可行性和有效性。3.提出了一种基于动态稀疏K-SVD(DSK-SVD)的字典学习方法。该算法的突出优点在于能够动态的计算稀疏编码的稀疏度,并对字典原子进行并行更新。首先,利用字典的互相关来定义稀疏编码过程中的稀疏度,用来动态的控制稀疏系数的稀疏度。其次,利用并行原子更新准则实现字典更新过程中的字典原子和稀疏系数的更新。在MSTAR数据库和HRRP数据上的实验结果验证了所提方法的可行性和有效性。4.提出了一种基于D-S证据迭代折扣方法的雷达目标融合识别方法,该方法对利用DSK-SVD算法对训练样本的特征进行学习,并利用测试样本的重构误差来定义基本概率分配(BPA)函数。首先,利用混淆矩阵以及BPA函数,计算出各个证据对应的折扣因子;其次,利用每次迭代得到的折扣因子重复对证据源进行修正,直到冲突系数小于给定的阈值;最后,利用修正后的证据进行融合识别。与其它典型融合识别方法相比,本文提出的方法在小样本情况下能够保持较好的识别性能。
[Abstract]:Radar target recognition technology is based on radar echo signal to extract information related to target characteristics to achieve target attribute or category determination.With the development of international situation, radar target recognition is more and more favored by researchers all over the world.With the application of high resolution radar system, it is possible to obtain more detailed geometric structure information and detail information.As a typical high resolution radar signal, high resolution Resolution radar (HRR) and synthetic Aperture radar (SAR) echo signal HRRP and SAR image have also become the hot spot of radar target recognition in many countries.On the basis of sparse learning theory, the radar target recognition method based on HRRP target and SAR image target is studied in this paper. The main research work is as follows: 1.The sparse learning theory is studied.Firstly, three typical sparse modeling methods, three classical sparse solving methods and the application of sparse learning are described. Secondly, the sparse representation methods of HRRP targets and SAR images are studied, respectively.And its sparsity is analyzed. 2.A SAR image denoising algorithm in Shearlet domain based on Bayesian model is proposed. The proposed algorithm not only utilizes the spatial correlation between sparse coefficients, but also obtains the dynamic noise threshold based on Bayesian model.The edge information can be effectively maintained while noise filtering is realized.Firstly, the SAR image after logarithmic transformation is represented by Shearlet sparse representation; secondly, the Bayesian model is used to model the noise detection according to the statistical characteristics of the sparse coefficient; finally, the noise pixel smoothing of the SAR image is realized by adaptive weighted shrinkage.The experimental results on MSTAR database show that the proposed method is feasible and effective.This paper presents a dictionary learning method based on dynamic sparse K-SVD-DSK-SVD.The outstanding advantage of this algorithm is that it can dynamically calculate the sparse degree of sparse coding and update dictionary atoms in parallel.Firstly, the sparse degree in the process of sparse coding is defined by the cross-correlation of dictionaries, which is used to control the sparsity of sparse coefficients dynamically.Secondly, the parallel atomic update criterion is used to update the dictionary atoms and sparse coefficients in the process of dictionary updating.The experimental results on MSTAR database and HRRP data show that the proposed method is feasible and effective.A method of radar target fusion recognition based on D-S evidence iterative discount method is proposed. This method studies the features of training samples using DSK-SVD algorithm, and defines the basic probability allocation (BPA) function by using the reconstruction error of test samples.First, the discounted factors corresponding to each evidence are calculated by using the confusion matrix and the BPA function. Secondly, the discounted factors obtained by each iteration are used to modify the evidence source repeatedly until the conflict coefficient is less than the given threshold.The modified evidence is used for fusion recognition.Compared with other typical fusion recognition methods, the proposed method can maintain better recognition performance in the case of small samples.
【学位授予单位】：南京航空航天大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TN957.52

【参考文献】