【摘要】 基于BP-Adaboost的目标分类算法用于雷达目标分类具有良好的效果.随着训练样本以及测试样本数增加,经典"一对多(One vs.Rest,OvR)"BP-Adaboost算法所需用时也随之增加.提出一种改进的多分类BP-Adaboost算法应用在雷达多目标分类上,在提高分类准确率的同时,有效地解决经典算法在多分类上时间开销过大的问题.该方法采用二进制方法重新表示样本数据类别,使用Adaboost算法将多个BP神经网络弱分类器集成起来学习,通过修改经典算法中的损失函数连续调整训练样本分布和弱分类器的权重,最终形成一个强分类器.对雷达高分辨率距离像(High Resolution Range Profile,HRRP)数据集进行分类仿真结果表明,相比于单个BP神经网络基学习器,所提算法的分类准确率提高了5%~10%,相比于经典的"一对多"BP-Adaboost算法,该算法所需用时仅为传统算法的1/2~1/3.更多还原

【Abstract】 BP-Adaboost based classification algorithm has a promising performance in the field of Radar target classification.However,due to the increase of the volume of training samples and testing samples,traditional BP-Adaboost based algorithm OvR(One vs.Rest)requires an amount of time.Therefore,this paper proposes an improved BP-Adaboost algorithm to solve the issue of Radar multi-target classification,which can improve accuracy and solve the time issue that traditional algorithm will encounter in multi-class classification effectively.Specially,we preprocess the initial categories in binary,then several BP neural networks are boosted up to form a new strong classifier,in which process we keep adjusting the distribution of training samples and weight of weak classifiers by modifying the loss function and the number of output nodes.Experiments using the range profile datasets have been demonstrated that the proposed scheme can raise accuracy up to 5%~10% compared to single BP neural network,and significantly im-proves the time issue efficiently compared to traditional algorithm OvR,which can save 1/2~2/3 running time under the same condition.更多还原