基于特征标权的中文签字核聚类研究

Kernel Clustering of Chinese Signature Based on Feature-Weighting

针对不同书写者书写同一字的分类问题,介绍了签字的五个全局特征的提取方法.在特征总数不多的情况下,使用特征标权而不是特征选择的方法来反映各特征对于签字分类的区分度不一样的事实,并着重讨论了如何利用待分类的模式,无监督的进行特征标权以得到权重向量的方法.将权重向量加入到作为核函数的高斯函数中,以核聚类方法对签字进行分类,实验显示,采用同样的核聚类步骤,加入权重向量后分类正确率较没有权重向量时的分类正确率有明显提高,权重向量自学习较同类方法指导性更强,说明该方法适用于文中提出的中文签字的分类问题,是可行且有效的.

This paper describes features extraction method for five globe features of Chinese signature for classifying signatures of different writers. Feature weighting not selection is selected to reflect the fact that the effect of each feature is not same for classifying the signature when the number of features is small. How to use all samples to get the weight vector in unsupervised method is discussed too. The kernel clustering method using weighted gauss function classifies signatures according to the degree of effect of every feature. Experiments show this weighting-based method improves the right rate of classifying and indicate it is fit to use this method to classify the problem of Chinese signature raised in this paper. The results of experiments show this method is feasible and effective too.