改进D-S证据理论的多分类器决策层融合系统  

Multiple Classifiers Decision Fusion System Following Improved D-S Evidence Theory

作　　者：王成[1] 郭飞[2] 郑黎晓[1] 赖雄鸣[3] 

机构地区：[1]华侨大学计算机科学与技术学院,福建厦门361021 [2]西安交通大学电子与信息工程学院计算机系软件与理论研究所,西安710049 [3]华侨大学机电及自动化学院,福建厦门361021

出　　处：《小型微型计算机系统》2015年第5期1138-1141,共4页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(51305142;51305143)资助;福建省自然科学计划基金项目(2014JO1191)资助;中国博士后科学基金第55批面上项目(2014M5524)资助

摘　　要：针对传统D-S证据理论中基于识别率和误识率构造的基本概率赋值函数(Basic Probability Assignment,BPA)没有考虑训练样本分布的缺点,提出了一种将整体错误率分配给除了正确判别命题以外各个焦元的BPA构造新方法.针对传统D-S证据理论中所采用的基于正交和运算的合成规则不能融合矛盾证据的缺陷,提出一种能融合矛盾证据的大概率赋值法.在此改进D-S证据理论的基础上,给出了两分类器决策层融合流程和多分类器决策层融合系统.在ORL和Yale数据库上的实验结果表明,对几种典型分类器的决策层融合提高了系统人脸识别的正确率,且改进D-S证据理论比传统D-S和投票融合方法的正确率更高.In order to overcome the disadvantages of traditional D-S evidence theory method that it does not consider distribution of training samples in basic probability assignment （ BPA ） function based on construction of recognition rate and false accept rate, this paper presents a new BPA construction method considering distribution of training samples which can assign the overall error rate for each focal element except correct classification proposition. In order to overcome the disadvantages of traditional D-S evidence theory method that it is unable to fuse contradictory evidences in combination rules based on orthogonal calculation, this paper proposes a new maximum probability assignment method that can fuse contradictory evidences. After that, a flow diagram of two classifiers decision fusion and multiple classifiers decision fusion system are given in detail based on this improved D-S evidence theory. Numerical experiment results in ORL and Yale databases show that decision fusion of multiple typical classifiers could increases face recognition accuracy rate, and the improved D-S evidence theory has a better performance than the traditional D-S evidence theory and voting method.

关 键 词：D-S证据理论 BPA构造新方法 大概率赋值法 多分类器决策层融合 人脸识别 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]