支持向量描述鉴别分析及在人脸识别中应用.docVIP

支持向量描述鉴别分析及在人脸识别中应用 　　摘 要:数据降维是模式识别的重要组成部分。支持向量鉴别分析(support vector discriminant analysis, SVDA)依最优超平面法线方向投影对数据进行降维,克服了传统方法中假设数据满足高斯分布时,导致无法反映超平面单侧中多类数据间投影距离差异并影响了算法有效性的缺点。提出一种支持向量描述鉴别分析(support vector description discriminant analysis, SVDDA)算法,首先利用支持向量机最优超平面获取样本的类属信息,然后通过SVDD的超球面法线作为投影轴取得样本的投影距离,取两信息的组合作为样本的特征映射。算法利用SVDD的一类紧致超球特性,弥补支持向量鉴别分析的不足。通过人脸识别实验,验证了该算法的有效性。 　　关键词:特征降维; 支持向量鉴别分析; 支持向量数据描述; 支持向量描述鉴别分析; 人脸识别 　　中图分类号:TP391.41 　　文献标志码:A 　　 　　文章编号:1001-3695(2010)02-0488-03 　　doi:10.3969/j.issn.1001-3695.2010.02.022 　　 　　Support vector description discriminant analysis andits application to face recognition 　　 　　CHEN Chang-jun, ZHAN Yong-zhao, WEN Chuan-jun 　　 　　(School of Computer Science Communication Engineering, Jiangsu University, Zhenjiang Jiangsu 212013, China) 　　 　　Abstract:Dimension reduction of data is one of the most important components for pattern recognition. SVDA projected data according to the normal direction of SVM’s optimal separating hyperplane. SVDA could overcome the shortcomings of that traditional dimension reduction methods always assumes data meets Gaussian distribution. But the differences of some classes in the same side of SVM hyperplane could not be reflected by normal direction projection. This paper proposed a new dimension reduction method named as support vector description discriminant analysis. Obtained class information through SVM hyperplane and extracted projection distances through SVDD hypersphere projection normal, set the combination of class information and projection distance of sample data as corresponding feature component. And applied the algorithm to the feature extraction in face recognition. The results show the effectiveness of this algorithm. 　　Key words:feature dimension reduction; SVDA; SVDD; SVDDA; face recognition 　　 　　0 引言 　　在统计模式识别问题中,原始观察数据维数很高,存在大量的冗余信息,会导致算法复杂度过高、计算量太大、降低分类的性能,因此需要对原始数据降维[1,2]。传统的降维方法主要有主元分析(PCA)[3]、Fisher线性鉴别分解(LDA)[4]、独立分量分析(ICA)[5]、奇异值分解(SVD)[3]等。核化算法[6]将原始数据通过非线性