基于最小距离最大原则的模式分类  

Pattern Classifying Based on Principle of Maximizing Minimum Distance

作　　者：陈绵书[1] 付萍[1] 张春雨[1] 

机构地区：[1]吉林大学南岭校区通信学院,长春130025

出　　处：《计算机工程》2004年第9期28-30,共3页Computer Engineering

基　　金：国家自然科学基金资助项目(60172046)

摘　　要：广义最佳鉴别矢量集是用于高维空间模式分类的有效方法。在高维空间模式分类的情况下，通常训练样本数要小于特征空间维数，从而使类内散布矩阵奇异，不能直接求解。对于这种情况下求解问题，人们给出了一系列求解方法。但这些方法均是应用平均的类间距离，而没有考虑到类间分布的最坏情况。该文对此进行改进，提出了一种新的基于最小距离最大原则的广义最佳鉴别矢量集的求解方法。实验结果表明，提出的方法要优于已有的方法。The general optimal discriminant vectors method is an effective method for pattern classifying in high-dimension space. In the case of classifying in high-dimension space, in general, the number of training samples is less than that of number of dimension of the feature space. So the within-class matrix is singular. One cannot get the Fisher optimal discriminant vectors directly. People have proposed many methods for solving the problem of a small number of samples. However, all these methods use the mean distances between classes, without considering the worst case. This paper improves it and proposes a new algorithm for solving the general optimal discriminant vectors based on principle of maximizing minimum distance. The experimental results show that the method is superior to the existed methods.

关 键 词：式分类 最佳鉴别矢量集 最小距离最大 人脸识别 模式分类 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]