【作者】 周代英；

【导师】 杨万麟；

【作者基本信息】 电子科技大学 ， 通信与信息系统， 2001， 博士

【摘要】 自动雷达目标识别作为一种自动目标识别技术，受到各国的广泛关注和重视，已成为各国未来武器的重要组成部分。采用不同的雷达作为观测手段，收集到目标的散射信息量会不同。使用高距离分辨率雷达可获得目标的一维距离像，一维距离像包含了更多的可用于目标识别的信息。本文针对一维距离像，对多种雷达目标识别方法进行了研究和讨论，同时对目标识别中的目标建库方法也进行了研究和探讨。其主要内容和创新之处如下： 1．讨论了雷达目标散射中心模型。 2．针对正则子空间维数受目标类别数限制的缺点，提出了修正正则子空间的一维距离像目标识别方法。修正正则子空间维数不受目标类别数的限制，同时其座标轴(基矢)在目标分离意义上优于正则子空间。以修正正则子空间变换进行目标特征提取和分类，可改善识别率。 3．提出了基于最优聚类中心的一维距离像识别方法。由目标训练样本集，建立最小平方距离准则下的最优变换矩阵。通过此最优变换可以减少同类目标之间的差异，而在变换空间选取最优的聚类中心可以增大异类目标之间的差异。 4．研究了基于最优投影平面的一维距离像识别方法。将多类目标识别问题转化为两类目标识别问题，对两类目标，使用最优投影平面方法进行分类，然后，采用投票机制确定目标最终所属的类别。将多类目标识别问题转化为两类目标识别问题这一思路可以应用到正则子空间法。 5．研究了目标识别中的建库方法，基于检测理论，提出非库属目标判别门限，在常规目标识别方法中的分类阶段引入该门限，可对非库属目标(没有参与训练的目标)进行判别，当输入目标被判别为非库属目标后，即可在训练目标库中加入该目标。 这些内容和创新点为多组仿真目标一维距离像数据和实测飞机一维距离像数据的识别实验所验证。更多还原

【Abstract】 The automatic radar target recognition system, which is one of the key components of present and future defense weapon systems, is widely and intensively focused by all of the countries. The scattering information collected by the different radars is different. The rangeprofile can be formed by high-range resolution radar (HRR). The rangeprofiles of the targets contain more information for recognition than the radar cross section (RCS). Some methods of radar target recognition are intensively and extensively studied in this paper. The method for building target library is also discussed. The main contents and new ideas include:The scatter-center model is discussed. The four kinds of simulated point targets are designed and the rangeprofiles at aspect angle are computed.The radar target recognition method based on the modified canonical subspace is investigated. First, the dimension of the modified canonical subspace are not limited by the number of target classes; Second, in the sense of separability, the modified canonical subspace transformation is optimal to the canonical subspace transformation. Therefore, the recognition performance is improved.The radar target recognition method based on the optimal cluster centers is studied. The optimal transformation matrix is formed by training sample set. The difference between the same classes is reduced by using this optimal transformation. The difference between the different classes is enhanced by selecting the optimal cluster centers in feature space.The optimal projection plane based radar target recognition method is discussed. The recognition problem for many classes is changed into the recognition problem for two classes. For two classes problem, the optimal projection plane method is used. The votes for each class are collected and a final decision made in accord with the class having the most votes. The ideathat the recognition problem for many classes is turned into the recognition problem for two classes may be applied to the canonical subspace method.The unknown target rejected method for building target library is investigated. The unknown target rejected threshold (UTRT) is used in classification phase of the conventional recognition approach to reject the unknown targets that are not included in training dataset. This may be used to build target library dynamically.All of these methods are proved by experiments on simulated data and real data of planes. These methods can be applied to automatic and real time recognition systems.更多还原