摘要
经典的基于距离测度的SVDD(Support Vector Domain Description)方法在解决两类(多类)识别问题时具有误判率较高、识别率低于普通二类SVC分类器等缺点.本文在分析其原因的基础上,提出了一种更能反映样本与类别本质关系的推广能力测度,并由此提出了具有多层结构的多类SVDD模式识别方法.对实测雷达一维距离像数据的测试表明,该方法在保留了经典SVDD识别器算法复杂程度低、扩充性强、对训练样本数据规模上要求低等优点的同时,有效地降低了误判率,识别率已接近甚至达到二类SVC的水平.
Classic SVDD classifiers, which use distance measures, have lower recognition rate than normal two-class SVM classifiers. After analyzing the causes, a new measure is proposed, which can represent the more essential relationship between sampies and categories.And then,a multi-level SVDD is proposed. The experiment on real one-dimensional range profiles data shows that,the multi-level SVDD reserves the lower complexity,higher expansibility and fewer requirements to sample size,ERs are reduced effectively, CR is increased even to the level of two-class SVM classifiers.
出处
《电子学报》
EI
CAS
CSCD
北大核心
2009年第3期464-469,共6页
Acta Electronica Sinica
关键词
模式识别
SVDD
多层结构
多分类算法
pattern recognition
SVDD
mulfi-level architecture
multi-class classification
