基于分类器联合的反舰导弹HRRP目标识别与拒判研究

摘要

为了进一步提高存在库外目标情况下反舰导弹的目标识别率,利用高分辨距离像(HRRP)信息,构造了一个兼具拒判功能的反舰导弹目标自动识别系统.利用feko软件仿真建立了6类目标在不同方位角下HRRP数据库,并针对HRRP识别问题,提取7个基于目标结构的平移不变特征;选择二次判别分类器、7-最近邻分类器和决策树分类器作为基分类器,并利用IM-BP神经网络作为联合器构成多分类器系统;引入基于非参数估计的置信度评估方法,确保系统能够拒判置信度较低的待测样本.试验结果表明:LM-BP集成系统能够提高基分类器的分类性能,且在库外目标存在情况下,引入适当置信度判别阈值的多分类器系统能够兼顾识别正确率和宣判率.%In order to improve recognition rate of anti-ship missile with presence of outlier samples,the high-resolution range profile (HRRP) information is used to build an automatic target recognition system with rejection function.A HRRP database with six targets classes from various aspect angles is established,and seven physically based features are defined on the database.Three kinds of classifiers are used as base classifiers:quadratic discriminant classifier,7-nearest neighbor classifier and decision tree classifier.The LM-BP neural network is utilized as a classifier combiner.A confidence evaluation method based on non-parametric estimation is introduced,ensuring that the system rejects samples with a low decision confidence.Experiments show that the LM-BP ensemble system can improve the performance of the base classifiers.Besides,in the presence of outlier targets,the multiple classifier systems with advisable confidence threshold can make a suitable compromise between the probability of correct classification and rate of declaration.