一种基于奇异值分解的特征抽取方法  

A Method of Feature Extraction Based on SVD

作　　者：王文胜[1] 陈伏兵[1] 杨静宇[1] 

机构地区：[1]南京理工大学计算机系

出　　处：《电子与信息学报》2005年第2期294-297,共4页Journal of Electronics & Information Technology

基　　金：国家自然科学基会(60072034)资助课题

摘　　要：特征抽取是模式识别的基本问题之一,Fisher线性鉴别分析是特征抽取中最为经典和广泛使用的方法之一。该文分析了Fisher线性鉴别分析在求解过程中可能存在的问题:鉴别矢量的分量可能是复数;特征值对扰动的敏感性;鉴别矢量之间未必具有正交性。由此提出了均衡散布矩阵的概念,并利用均衡散布矩阵构造了一种新的线性鉴别准则。利用奇异值分解定理,将求取鉴别矢量转化为对矩阵求奇异向量。用该方法进行求解可以有效地避免前述的问题。试验结果表明,该鉴别准则具有良好的鉴别能力。Feature extraction is primary problem of pattern recognition. As one of the most classic methods in the field of feature extraction, Fisher linear discriminant analysis is applied widely. It may meet several possible problems in finding optimal set of discriminant vectors: the components of these vectors may not be real; the eigenvalue may be sensitive; these vectors may not be orthogonal each other. So the balanced scatter matrix is proposed in this paper. Based on the matrix, a discriminant criterion is formed. The optimal set of discriminant vectors can be acquired througn singular value decomposition theorem. The method can avoid the problems metioned above. The result of face recognition experiment shows that it has powerful ability of feature extraction.

关 键 词：特征抽取 线性鉴别分析 奇异值分解 模式识别 矢量 矩阵 正交性 奇异向量 特征值 分量 

分 类 号：TN911[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]