图像二维鉴别特征矩阵融合及人脸识别  

Two-Dimensional Canonical Correlation Analysis and Its Application to Face Recognition

作　　者：刘永俊[1] 赵根林[1] 陈才扣[2] 杨静宇[3] 

机构地区：[1]常熟理工学院软件工程系,江苏常熟215500 [2]扬州大学信息工程学院,江苏扬州225009 [3]南京理工大学计算机科学与工程系,江苏南京210094

出　　处：《常熟理工学院学报》2007年第10期102-107,共6页Journal of Changshu Institute of Technology

基　　金：国家自然科学基金(60472060);江苏省高校自然科学基金(05KJB520152)资助项目

摘　　要：二维最大散度差鉴别准则和二维Fisher鉴别准则抽取的特征具有很强的相关性.本文在此基础上,通过对传统的基于向量的典型相关分析方法进行分析改进,提出了一种新的直接基于图像二维鉴别特征矩阵融合的二维典型相关分析方法,并将其应用于人脸识别的特征融合过程中.较基于向量的典型相关分析,该方法计算过程中构造的协方差矩阵维数大幅度减小.这在一定程度上避免了人脸识别中存在的"高维小样本问题",另一方面也使算法的速度明显提高.By analyzing the relativity between two-dimensional maximum scatter-difference discriminant analysis （2DMSLDA） and two-dimensional Fisher discriminant analysis （2DFLDA）, according to traditional canonical correlation analysis （CCA）, a novel method of combining different feature matrixes directly is proposed in this paper by using the main idea of image projection in face recognition. Compared with traditional CCA based on feature vectors, this method has the following two main advantages： first, the small sample size problem （SSS） occurred in traditional CCA is essentially inevitable as a result of the evidently reducing dimension of the covariance matrix. By the same reason, the second advantage is that much computational time would be saved if using the proposed method. Finally, extensive experiments performed on ORL face database verify the effectiveness of the proposed method.

关 键 词：典型相关分析 二维典型相关分析 特征融合 人脸识别 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]