【作者】 李飞；

【导师】 刘宏伟；

【作者基本信息】 西安电子科技大学 ， 信号与信息处理， 2014， 博士

【摘要】 雷达成像技术具有全天候、全天时、远距离观测能力，有效提高了雷达的信息获取能力，具有重要的军用和民用应用价值。随着雷达成像技术的高速发展，雷达图像收集能力越来越强。从大量雷达图像中获取目标信息进行目标检测、分类与识别则是雷达成像的目的，其中基于雷达图像的目标识别受到越来越广泛的关注。传统的基于数据驱动的目标识别方法依赖于数据本身所反映出来的目标信息，而在实际环境中数据反映的目标信息随环境变化而变化，这给基于数据驱动的目标识别方法带来挑战。目标的一些物理特征（如目标几何尺寸，物理结构等）受环境因素影响较小，而且雷达成像机理在一定程度上可以反映目标物理特征。本文以此为出发点，研究了基于目标参数化回波模型雷达图像中目标物理特征提取技术。论文分别从干涉ISAR（InISAR）横向定标、目标微动特征提取以及目标电磁特征（属性散射中心特征与极化特征）三个方面探讨和研究目标物理特征提取技术，主要工作概括如下：1.针对干涉ISAR（InISAR）横向定标中相位缠绕问题，提出一种基于随机霍夫变换（Randomized Hough Transform, RHT）的InISAR横向定标算法。该算法利用ISAR图像中的特显点的模糊横距（干涉相位主值得到的横距）与多普勒频率的线性关系，估计真实横距与多普勒频率之间的尺度因子，实现对目标ISAR图像横向定标，从而避免了繁琐的相位解缠绕过程。仿真实验结果表明该算法可以实现对ISAR图像的定标，并具有一定的抗噪声性能，由定标后的目标ISAR图像可以提取目标的几何尺寸特征。2.研究了目标微动特征提取方法，主要包含两部分：（1）在分析微动目标窄带回波信号基础上，指出窄带目标的微动特征提取等价于多分量非平稳信号的瞬时频率估计问题，我们提出一种基于曲线跟踪（Curve track, CT）算法的目标微动特征提取方法。该算法在时频域通过最近邻数据关联（Nearest Neighbor DataAssociation, NNDA）算法对各分量信号的时频曲线进行关联与分离，然后采用扩展卡尔曼（Kalman）滤波器对分离的时频曲线进行平滑滤波，并基于平滑后的时频曲线估计目标微动参数。（2）通过分析微动目标的宽带回波，指出目标回波在距离包络域具有微距特征，以及时频域具有微多普勒特征。对于微距特征提取，利用幅度相位估计方法（Amplitude and phase estimation, APES）得到目标距离包络信号的超分辨估计，在此基础上由CT算法实现目标微距特征曲线的分离与提取；针对微多普勒特征提取，利用多距离单元信号进行联合时频分析得到目标完整的多普勒谱，并由CT算法提取目标微多普勒特征。最后基于电磁计算数据的仿真结果验证所提算法的有效性。3.研究了属性散射中心提取方法，包含两部分：（1）考虑目标频率-方位二维观测数据在属性散射中心模型参数空间上的稀疏性，我们提出一种基于字典缩放的属性散射中心提取与参数估计方法。由于模型参数维数较高，构造的高维联合字典将消耗较多系统资源。针对该问题，所提算法采用交替优化与字典缩放实现了对参数化字典的降维。为了减小邻近散射中心之间的相互干扰，采用正交匹配追踪（OMP）-RELAX联合算法求解稀疏信号恢复问题，实现在频率-方位角域提取属性散射中心。（2）我们提出一种基于距离特性与方位特性解耦合的属性散射中心提取算法，进一步降低对系统资源的要求。该方法通过分别构建包含位置信息与方位特性信息的两个低维字典代替高维的联合字典实现距离特性与方位特性的解耦合，并得到散射中心参数估值。根据提取的属性散射中心可以有效地估计目标或目标重要部件的几何尺寸。基于电磁计算数据和实测数据的实验结果验证了上述算法的有效性。4.研究了全极化属性散射中心提取方法：（1）考虑目标全极化观测数据在属性散射中心模型参数空间上的联合稀疏特性，利用联合稀疏表示技术提取属性散射中心，并对估计的极化散射矩阵进行极化分解提取目标极化特征，联合干涉测高可以得到目标三维姿态信息。该方法采用基于字典缩放属性散射中心提取算法思想实现参数化字典的降维，对稀疏系数矩阵施加行稀疏约束，通过SOMP（Simultaneous Orthogonal Matching Pursuit）算法求解联合稀疏优化问题并提取属性散射中心。（2）针对散射中心重叠或者同一分辨单元内包含不止一种散射体的情况，依据目标全极化观测在属性特征域（属性参数以及散射类型）的稀疏特性，对目标极化分解系数矩阵施加行稀疏约束与矩阵稀疏约束，该算法利用坐标轮回下降法估计目标极化分解系数矩阵与极化散射机理字典，同时提取目标全极化属性散射中心及其极化特征。基于电磁计算数据的实验结果验证了上述算法的有效性。更多还原

【Abstract】 Radar imaging technique has the ability of all-weather, day/night and long rangeapplications, which enhances the radar capability of information acquisitiondramatically. Therefore, radar imaging technique plays an important role in manymilitary and civilian fields. As radar imaging technique has developed rapidly, thecapability of image collection has grown much stronger. The aim of radar imaging is toacquire information of target for target detection, classification and recognition, andradar image based target recognition has attracted more and more attention especially.The performance of data-driven recognition method depends on the information thatmeasurements reflect, while the information is influenced by the environmental factorsgreatly in practical application, and this is a challenge for the data-driven recoginitonmethod. The environmental factors have minor impact on physical feature of target,such as geometrical dimension and physical structure, which can be reflected by themechanism of radar imaging. This dissertation studies on the physical feature extractionfrom radar image based on the parametric signal model, and focuses on the technique ofphysical mechanism feature extraction in three aspects: cross-range scaling ofinterferometric ISAR, micro-motion feature extraction of target and electromagneticfeature extraction of target (attributed scattering center and polarimetric feature). Themain work can be summarized as follows:1. Aiming at the phase wrapping problem in the cross-range scaling ofinterferometric ISAR, a novel algorithm is proposed based on Randomized HoughTransform. Using the linear relationship between the azimuth positions derived from thewrapping interferometric phases and the Doppler frequencies of the dominant scatters inISAR image, the proposed algorithm estimates the scale factor between true azimuthposition and Doppler frequency and determines the ISAR image scale in the cross rangedirection, consequently avoiding the complex phase unwrapping procedure. Thesimulation results verify the validity and anti-noise capability of this algorithm, and thegeometrical dimension of target can be obtained according to the cross-range scaledISAR image.2. The micro-motion feature extraction is studied in this section, and the main workconcerns the following two aspects. Firstly, based on the analysis of narrowband echoesof target, we point out that its micro-motion feature extraction is equivalent to theinstantaneous frequency estimation of multi-component non-stationary signal. A novelmethod based on curve tracking algorithm is proposed to extract micro-motion feature.The proposed method separates the time-frequency curves successfully in time-frequency domain with the Nearest Neighbor Data Association (NNDA) algorithm,and the separated time-frequency curve of each component signal is smoothed by theextended Kalman filter, then the parameter of micro-motion can be estimated with thesmoothed time-frequency curves. Secondly the wideband echoes are analyzed withsome conclusions that the wideband echoes posess micro-range feature indownrange-slow time domain and micro-doppler feature in time-frequency domain.APES (Amplitude and phase estimation) is adopted to obtain the superresolution ofmicro-range, and then CT algorithm is used to separate and extract the micro-rangecurve; Signals of multi range cells are jointly analyzed to obtain the intacttime-frequency distribution of target, then CT algorithm is utilized to extractmicro-doppler feature. Simulation results on electromagnetic data verify the validity ofthe proposed algorithms.3. Attributed scattering center extraction is discussed in this section and it containstwo parts.(1) Considering the sparsity of the frequency-aspect backscattered data in theattributed scattering center model parameter domain, a novel method based ondictionary refinement is proposed to extract attributed scattering center and estimateparameters in frequency-aspect domain. Due to the high dimension of model parameter,one high dimensional joint dictionary needs to be constructed, which may cost a massstorage. Aiming at this problem, dictionary zooming and alternative optimization areexplored to reduce the dictionary dimension in the proposed algorithm, sparse signalrecovery problem is solved by utilizing OMP (Orthognal Matching Pursuit) algorithmcombined with RELAX to alleviate the influence of closely-spaced scattering centers oneach other, and attributed scattering centers are extracted from frequency-aspect domain.(2) A new method based on range characteristic and aspect characteristic decoupling isdeveloped to reduce the storage resources request further. Two low dimensionaldictionaries including localization and aspect attribute parameters respectively areconstructed to replace the high dimensional joint dictionary to decouple the rangecharacteristic and aspect characteristic and obtain the parameters estimation. With theextracted attributed scattering centers, geometrical dimensions of the target or its mainstructure can be estimated. Numerical results both on electromagnetic computation dataand measured data verify the validity of the proposed method.4. This section focuses on the extraction of fully polarimetric attributed scatteringcenter:(1) Considering the joint sparsity of the fully polarimetric measurements in theattributed scattering center model parameter domain, a novel method based on jointsparsity is proposed for attributed scattering center extraction; and the recovered sparse coefficient matrix is used for polarimetric signature extraction, the three dimensionalpose of target can be acquired by interferometric processing. Alternative optimizationand dictionary refinement are utilized for dimension reduction, which are also adoptedin the dictionary refinement based attributed scattering center extraction method, whilejoint sparsity constraint is imposed on the sparse coefficient matrix, and simultaneousorthogonal matching pursuit (SOMP) are adopted to find the solution of the joint sparseoptimization problem for attributed scattering center extraction.(2) Aiming to extractattributed features of overlapped attributed scattering centers, row sparse constraint andmatrix sparse constraint are imposed on the polarimetric decomposition coefficientmatrix of target based on the sparsity of scattering center in the parameter domain andscattering mechanism domain, coordinate decent technique is employed to optimizepolarimetric decomposition coefficient matrix and polarimetric scattering mechanismdictionary to extract attributed scattering center and polarimetric signaturesimultaneously. Numerical results on electromagnetic computation data verify thevalidity of the proposed algorithms.更多还原