摘要
多分类器系统是应对复杂模式识别问题的有效手段之一.当子分类器之间存在差异性或互补性时,多分类器系统往往能够获得比单分类器更高的分类正确率.因而差异性度量在多分类器系统设计中至关重要.目前已有的差异性度量方法虽能够在一定程度上刻画分类器之间的差异,但在应用中可能出现诸如"差异性淹没"等问题.本文提出了一种基于几何关系的多分类器差异性度量,并在此基础上提出了一种多分类器系统构造方法,同时通过实验对比了使用新差异性度量方法和传统方法对多分类器系统融合分类正确率的影响.结果表明,本文所提出的差异性度量能够很好地刻画分类器之间的差异,能从很大程度上抑制"差异性淹没"问题,并能有效应用于多分类器系统构造.
The multiple classifier system is one of the effective means to resolve pattern recognition under complicated environments. When the member classifiers are diverse or complementary, multiple classifier systems can usually obtain higher classification accuracy compared with a single classifier. Thus, diversity measures are crucial to multiple classifier systems design. Though the existing diversity measures can, to some degree, describe the difference among classifiers, they may lead to problems like ＂diversity submergence＂ in some cases. In this paper, a novel multiple classifier diversity measure based on geometric relationship and a multiple classifier system constructing method based on the new diversity measure are proposed. It is experimentally shown that the proposed diversity measure can well describe the diversity among classifiers and effectively suppress the problem of ＂diversity submergence＂. It can also be effectively used in designing multiple classifier systems.
出处
《自动化学报》
EI
CSCD
北大核心
2014年第3期449-458,共10页
Acta Automatica Sinica
基金
国家重点基础研究发展计划(2013CB329405)
国家自然科学基金(61104214
61203222)
国家自然科学基金创新群体(61221063)
中国博士后科学基金(201104670)
中央高校基本科研业务费专项资金(xjj2012104)资助~~
关键词
多分类器系统
差异性度量
差异性淹没
几何中心
Multiple classifier system, diversity measure, diversity submergence, geometric center
