【作者】 李龙；

【导师】 刘峥；

【作者基本信息】 西安电子科技大学 ， 模式识别与智能系统， 2018， 博士

【摘要】 在如今日益复杂的地面战场环境下,利用传统雷达信号处理技术已无法实现对地面目标的准确探测,因此对雷达目标识别技术的需求愈加迫切。高分辨一维距离像可以提供目标在雷达视线上的结构信息,其成像条件简单、容易获取且运算与存储压力小的特点,使得基于高分辨一维距离像的雷达地面目标识别技术受到了广泛的关注和研究。目前,基于高分辨一维距离像的目标识别技术已经取得了一定的进展与突破,但是针对弹载这一特定的应用场合,目标非合作性导致的训练模板库非完备问题,地面目标相似导致的特征向量低可分性问题,复杂地面战场环境导致的低信噪比与假目标干扰问题,特征空间分布的随机性导致的分类边界不准确问题,严重地影响了目标识别的总体性能。本论文以实现复杂战场环境下的目标识别为出发点,围绕上述弹载雷达地面目标识别中存在的问题展开研究。具体包括以下几个方面:1.为解决雷达地面目标高分辨一维距离像识别中,非合作目标模板库非完备的问题,提出一种基于混合模型的雷达非合作目标高分辨一维距离像仿真方法。本方法基于模型匹配目标识别基本思想,利用有限的目标信息进行建模仿真,从而构建完备的目标训练模板库。本方法通过构建目标精细化散射点模型,并利用时域高频电磁散射计算方法获得散射点的散射强度,同时基于距离单元服从的统计分布特性,建立散射点间的统计相关性,以实现目标高分辨一维距离像电磁散射特性与统计分布特性的兼顾。通过与目标实测数据的对比,本方法所生成的目标高分辨一维距离像与实测目标数据具有较高的相似性。利用本方法生成识别模板库,并基于实测数据进行测试,验证了本方法在目标识别中的有效性。2.为提取高分辨一维距离像的低维度、高可分性特征,提出基于统计核函数相关判别分析的特征提取算法。本算法通过对目标高分辨一维距离像距离单元统计特性的分析,分别构建距离单元理想统计模型与非理想统计模型下的统计核函数,对不同统计模型下的目标特性进行描述,从而实现最小信息损失的特征分量提取。以此为基础,基于可分性判别分析与典型相关分析理论构建特征融合准则函数,实现特征空间中类内相关性与类间差异性最大化,同时减少目标特征中的冗余信息,保证特征向量的低维度特性。利用实测数据对本算法进行验证,结果表明,在保证目标特征向量低维度的条件下,本算法有效地提高了特征向量的可分性,从而改善了高分辨一维距离像目标识别系统的总体性能。3.为提高低信噪比下高分辨一维距离像目标识别性能,提出基于稀疏-低秩联合学习的噪声稳健目标识别方法。本方法通过对稀疏、低秩的联合表示,实现对目标高分辨一维距离像局部特征与全局特征的提取。以此为基础,在训练阶段利用支持向量理论与字典学习原理,对特征提取字典进行优化,从而提高特征向量的可分性;此外,为了更加精确的对目标特征空间进行描述,结合机器学习理论,采用基于联合可分性分析的多分类器加权融合字典学习方法,进一步提高本模型的识别性能。在测试阶段,利用因子分析模型匹配方法对去噪声字典进行优化,从而实现对噪声的有效抑制,保证了目标识别系统的噪声稳健性。实验结果表明,本方法可在低信噪比条件下有效地恢复目标高分辨一维距离像,并实现较高的识别正确率。4.针对雷达地面目标高分辨一维距离像识别中的目标鉴别问题,构建一种基于聚类-空间描述的目标鉴别器。在训练阶段利用基于相关系数预处理的K-Means聚类方法对库内目标特征空间进行区域划分;针对特征向量的多区域聚合性,采用改进的支持向量域描述方法确定特征空间边界;最终利用特征空间边界与加权K近邻原则实现目标鉴别。本方法解决了库内目标与库外目标的鉴别问题,完善了目标识别系统的功能。通过对训练特征空间的区域划分,在有效地减小了运算复杂度的基础上,实现了更为精确的特征空间描述。最后通过实测数据进行实验,验证了本方法具备优良的鉴别性能与实时处理能力。5.为实现雷达高分辨一维距离像目标识别中鉴别与分类的联合处理,构建一种基于多重支持向量模型的目标识别器。根据目标高分辨一维距离像特征空间的密度分布,对基于密度敏感的多重支持向量模型进行组合优化,实现对特征空间的区域分割,并构建多个子分类超平面,实现对不同密度、不同类别目标的特征空间区域划分,以得到更为精细化的目标特征空间描述。本方法有效地解决了特征空间密度分布非均匀给目标识别带来的分类偏差问题,同时实现了目标鉴别与分类的融合。此外,本方法基于支持向量模型,内存需求少、计算复杂度低,适合目标识别系统的实际工程应用。利用实测数据得到的实验结果表明,本方法具有良好的目标鉴别与目标分类性能,并具备一定的工程实用化潜力。更多还原

【Abstract】 Due to the growing complexity of ground battlefield environment,it becomes increasingly difficult to detect target accurately with conventional radar signal and information processing.Thus the radar target recognition method has received intensive attention from the radar scholars.Radar high resolution range profile(HRRP)has been widely used for practical target recognition system for its easy acquisition and low storage requirement.Besides,HRRPs contain the detailed target structure signatures,such as strength and distribution of scatters,and target size,etc.So far,numerous efforts have been devoted to verify the advantage of HRRP-based target recognition.However,there are some problems to be further studied,such as the incomplete training samples of non-cooperative target,extraction of discriminative and low-dimension feature vector,design of discriminator and classifier,etc.Based on the practical requirements of precision attack for ground target in the complexity environment,this dissertation aims at the design of feature extraction and recognition method.The main contributions are summarized as follow:1.To solve the incomplete training samples of non-cooperative targets,a novel HRRP simulation method is proposed.First the exquisite scattering point model is constructed based on the mixture model and statistical distribution of scatters.Taking the covering influence into consideration,the constructed scattering model is further simplified by removing overlapped scatters on the radar line of sight.Then the high frequency approximate electromagnetic scattering computing method is used to obtain the backscattering intensity of scatters.Therefore,the simulated HRRP can be generated.By comparison,it can be seen that the simulated HRRP shares similar scattering and statistical characteristics with the real-measured HRRP.It is noted that the proposed recognition method is based on the template matching theory.In the recognition process,the simulated HRRPs serve as training samples while the measured HRRPs are utilized for testing.The experiment result verifies the effectiveness of the proposed method.2.To extract low-dimension and highly discriminative features from HRRP,a novel feature extraction method is designed,which is named as statistics kernel function discrimination analysis method.Based on statistical analysis,the statistical kernel functions are exploited with ideal and nonideal statistical model of HRRP range cells.Therefore the integrated target information can be obtained for target recognition with minimum information loss.In addition,a novel criterion function is constructed based on canonical correlation analysis and discrimination analysis.In this criterion function,the within-class correlation and between-class discrimination are maximized to guarantee high discrimination for feature vectors.Besides,the redundancy and dimensionality of the feature vectors are reduced by the fusion operation within the criterion function,which reduces time consumption for practical radar target recognition system.Experimental results with measured datasets validate the efficiency of the proposed method.3.To improve the HRRP-based target recognition performance under low signal-to-noise ratio(SNR),a target recognition method based on sparse and low-rank joint learning is proposed.Specifically,sparse representation(SR)and low-rank representation(LRR)are applied to extract the local and global characteristics of target HRRPs,respectively.To obtain the noise-robust and high-discriminative features of HRRPs,dictionary learning is involved.In the training stage,the discriminative dictionary is constructed based on Fisher discriminant criterion and support vector theory.Besides,in order to achieve more accurate feature space description,jointly discriminative analysis multiclass classifier weighted embedding dictionary learning method is used.Denoising dictionary optimization is implemented for noise suppression in the testing stage.Experimental results based on the measured HRRP data demonstrate that the proposed method can recover the original HRRPs and significantly improve the recognition performance under low SNR conditions.4.To identify the out-of-database targets in HRRP-based target recognition,an improved target identifier is designed based on clustering strategy and feature space distribution.In the training stage,a K-Means clustering strategy based on the pre-processing of correlation coefficient is utilized to divide the training feature space.Then each sub-space boundary is determined by support vector domain description(SVDD)based on the distribution of the feature space.Finally,the target category can be decided with the sub-space boundary and the weighted K-neighbors principle.This method can work without the template of out-of-database samples,which improves the effectiveness of target identification.Due to the fact that the feature space of different targets has the characteristic of non-uniform aggregation under different attitudes,a procedure of region partition is applied to training dataset.Thus the computational load is relieved by decreasing the searching operation of template matching.The requirement of real-time processing can be satisfied.Finally,the experiment results based on both simulation and real-data verify the effectiveness and efficiency of the designed identifier.5.For radar HRRP recognition,three aspects are of great importance to improve the performance,i.e.discrimination for outlier,classification for inner and accurate description for feature space.To tackle these issues,a novel target recognition method is designed,denoted by multiple support vectors method.First,a treble correlate support vector model is constructed to segment the feature space into two regions according to the density of feature vectors.Then the description and classification hyperplane for each region are obtained.Based on the support vector framework,the computation complexity can be reduced significantly for practical radar HRRP recognition.Finally,the experiment based on the measured data verifies the excellent performance of the proposed method.更多还原