快速Foley-Sammon鉴别变换及脸象鉴别  

Fast Foley-Sammon Transform and Face Identification

作　　者：杨健[1] 涂庆华[1] 杨静宇[1] 

机构地区：[1]南京理工大学计算机科学系,南京210094

出　　处：《中国图象图形学报（A辑）》2002年第1期1-5,共5页Journal of Image and Graphics

基　　金：国家自然科学基金 (60 0 72 0 3 4)

摘　　要：为了解决小样本情况下 ,类内散布矩阵不可逆时 ,Foley- Sam mon最优鉴别矢量集的求解问题 ,给出了一种快速近似算法 .首先从理论上说明了当类内散布矩阵不可逆时 ,将在原始特征空间内求解最佳鉴别矢量集的问题映射到等于或小于 c- 1(c为样本类别数 )维的欧氏空间内进行是可行的 .由于样本类别数远远小于原始特征空间的维数 ,故该算法不仅大大减少了特征抽取的时间 ,也提高了分类识别的速度 .在 ORL 标准人脸库上的试验结果表明 ,该算法不仅在识别率和识别时间上优于传统的扰动法和补空间法 ,而且比经典的特征脸方法和It is well known that Foley Sammon transform is a very effective method for feature extraction and that face identification is a typical small sample size problem. In the problems, the dimension of original features is so high that the within class scatter matrix is always singular. In the singulas case, how to calculate Foley Sammon optimal discriminant vectors(FSDV) is a very difficult problem. This paper presents a fast algorithm for calculating FSDV for small sample size problems. The main idea of the proposed algorithm is to map the problem of calculating FSDV in the original feature space to another problem of calculating FSDV in a( c-1 ) dimensional (or less) Euclidean space(where c is the number of pattern classes). In the transformed space, the FSDV can be calculated directly. Generally speaking, the number of classes is much less than the dimension of original sample, so our approach needs less time for feature extraction. We do Experiments on Olivetti Research Laboratory(ORL) face database. The experimental results show that our approach is better than previous method such as Perturbation and Complementary Space in terms of discrimination power and computing time, and also superior to classical 'Eigenfaces' and 'Fisherfaces' method with respect to recognition rate.

关 键 词：最优鉴别矢量集 特征抽取 脸象鉴别 FOLEY-SAMMON 人脸识别 图象识别 求解 近似算法 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]