基于保持投影的最大散度差的特征抽取方法  

Feature Extraction Method of Maximum Scatter Difference Based on Preserving Projection

作　　者：王建国[1,2] 杨万扣[1] 杨静宇[1] 

机构地区：[1]南京理工大学计算机科学与技术学院,南京210094 [2]唐山学院网络教育中心,唐山063000

出　　处：《模式识别与人工智能》2009年第4期610-613,共4页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金重点资助项目(No.60632050)

摘　　要：对非监督鉴别投影(UDP)准则进行修正,并在修正的准则基础上提出基于保持投影的最大散度差的特征抽取方法.该方法利用非局部散度与局部散度之差作为鉴别准则,从而避免UDP线性鉴别分析中所遇到的小样本问题引起的局部散度矩阵奇异的问题.在标准人脸数据库Yale和FERET上进行实验,实验结果表明本文方法的有效性.Firstly, the unsupervised discriminant projection （UDP） criterion is modified. Then, the feature extraction method of the maximum scatter difference based on preserving projection is proposed on the basis of the modified discriminant criterion. The proposed method adopts the difference of both nonlocal scatter and local scatter as discriminant criterion. Thus, the singular problem of local scatter caused by small sample size problem in UDP linear discriminant analysis is avoided. Finally, experimental results on Yale and FERET face databases demonstrate the effectiveness of the proposed method.

关 键 词：特征提取 人脸识别 散度差鉴别准则 非监督鉴别投影(UDP) 子空间分析 流形学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]