改进的多面最临近支持向量机  

Improved Multisurface Proximal SVM

作　　者：王晓辉[1] 荆晓远[2] 

机构地区：[1]韩山师范学院数学与信息技术学院,广东潮州521041 [2]哈尔滨工业大学深圳研究生院生物计算研究中心,深圳518055

出　　处：《计算机工程与应用》2006年第28期24-27,共4页Computer Engineering and Applications

基　　金：国家自然科学基金资助项目(编号:60402018)

摘　　要：基于泛化特征值问题的多面PSVM(GEPSVM)被O.L.Mangasarian证实是一种有效、简单、训练速度快的方法,但其仅对维数不高、样本数目也不大的数据集在实验中进行了比较和说明,而对上千维,甚至上万维人脸数据库,即小样本的、多类的问题并没有给出解决方法。文章把原算法加以改进,即把求解最小优化问题变成了求解最大优化问题,解决了因数据维数高、样本数较小而产生的奇异值问题,同时也实现了其多类算法,并用原GEPSVM算法和改进的算法来分别对这三个人脸数据库进行分类比较,从而使识别率和所用的处理时间两方面都得到了极大的改进。The effectiveness of Muhisurface Proximal SVM via generalized eigenproblem is demonstrated by tests on some simple examples or some public data sets in O.L.Mangasarian＇s paper.Although they proposed a very useful, simple and fast algorithm verified by their tests,they did not give a solution to deal with high-dimension sample sets, also the multi-class problems,just like face database.This paper gives the improved algorithm that can solve the small sample size problem which causes the singular problem,and also gives an implementation for multi-class problem.We also give the comparative results that show us the high recognition rate and low training tlme,on three public face datasets.

关 键 词：支持向量机 多类分类 人脸识别 泛化特征值 小样本集问题 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]