CAJ | 학술논문

传统的统计不相关鉴别分析方法使用样本的均值来估计期望，计算出总体散度矩阵。这些方法在数据不满足高斯分布的情况下会出现大的偏差，影响最优鉴别特征的提取。为了解决该问题，文中结合二维鉴别分析的思想，分别提出了基于局部的二维统计不相关鉴别变换( L2DUDT)方法和基于全局加权的二维统计不相关鉴别变换( WG2DUDT)方法。L2DUDT通过用样本的近邻中心来定义每个样本的期望，而WG2DUDT用样本间的欧几里得距离加权来定义期望。基于AR和FERET人脸数据库的实验表明，文中提出的方法与一些相关方法相比，有效地提高了识别性能。
전통적통계불상관감별분석방법사용양본적균치래고계기망，계산출총체산도구진。저사방법재수거불만족고사분포적정황하회출현대적편차，영향최우감별특정적제취。위료해결해문제，문중결합이유감별분석적사상，분별제출료기우국부적이유통계불상관감별변환( L2DUDT)방법화기우전국가권적이유통계불상관감별변환( WG2DUDT)방법。L2DUDT통과용양본적근린중심래정의매개양본적기망，이WG2DUDT용양본간적구궤리득거리가권래정의기망。기우AR화FERET인검수거고적실험표명，문중제출적방법여일사상관방법상비，유효지제고료식별성능。
The traditional uncorrelated discriminant analysis methods employ the mean of sample-set to estimate the expectation for all samples,thus computing the total scatter matrix. However,when the data are not Gaussian distributions,these methods may not extract op-timal discriminant features. In order to address this problem,propose two approaches named Local Two-Dimensional Uncorrelated Dis-criminant Transform (L2DUDT) and Weighted Global Two-Dimensional Uncorrelated Discriminant Transform (WG2DUDT) on the basis of two-dimensional discriminant analysis respectively. L2DUDT redefines the expectation for each sample using the sample's neighbor center,while WG2DUDT uses Euclidean distance between samples as weighted value to construct the expectation. The experi-mental results on AR and FERET databases demonstrate that the proposed approaches can effectively improve the recognition perform-ance,as compared with some related methods.