摘要
目前的许多分类器设计方法,如多层感知器网络(MLP)、支持向量机(SVM)、相关向量机(RVM)、径向基函数网络(RBF)等,实际是非线性映射加线性分类的方法,即将输入空间的非线性可分问题经非线性映射到另一空间,在那一空间实现线性分类.本文则开拓性的运用脉冲耦合神经网络神经元的点火捕获的思想,提出了一种基于耦合神经元点火捕获/抑制特性的分类器设计方法,使一类样本对应神经元总是较其它类样本对应神经元先点火以实现对样本的有效分类.所设计的分类器可实现对样本空间中任意复杂分布训练样本的非线性稳健分类,特别是有效实现复杂混叠模式的模式稳健分类,大量复杂混叠模式分类问题的仿真实验验证了本文方法的有效性和可行性,并应用于微波暗室实测一维距离像数据的自动目标识别中.
Nonlinear mapping plus linear classification,i.e. ,mapping inseparable input space into a hidden space via some nonlinear mapping, and designing a linear classifier to classify the data in the hidden space,is one of the most popular approaches for designing a classifier for an inseparable classification problem. In fact, multi-layer perceptron (MLP), support vector machine (SVM) ,relevance vector machine (RVM) and radial basis function (RBF) are the examples of such approach. In this paper, we propose a novel classifier which is not based on this approach, but on stimulating pulse-coupled neurons such that the neurons belonging to a same class activate each other while those belonging to different classes inhibit each other according to the stimulus and the geometry of neurons in the input space. The result of the competition between neurons is that all the neurons belonging to one class will fire earlier than all the other neurons in the net,which is used for dassification.A large number of experiments were conducted for classification of Iris data and simulated data where the overlap between patterns is very serious, as well as automatic target recognition for the model targets whose high-resolution range profiles were obtained from real microwave house. The experiments show that the proposed method is very effective in designing a classifier with good generalization ability and simple structure.
出处
《电子学报》
EI
CAS
CSCD
北大核心
2006年第12期2154-2160,共7页
Acta Electronica Sinica
基金
国家自然科学基金(No.60574039
60371044
60071026)
中意科技合作项目