A maximum uncertainty LDA-based approach for limited sample size problems - with application to face recognition

Abstract

Abstract

A new ranking method for principal components analysis and its application to face image analysis

“…To address this problem for the Fisher's criterion, we have used a regularized version of the LDA approach called MLDA [14].…”

“…Eq. (14) shows the Zhu and Martinez criterion has a bias that favors the direction provided by the line joining the sample group means. In small sample and high dimensional problems where two-classes are well separated, such direction is stable and can be estimated with much lower variance than the ones provided by separating hyperplanes.…”

A multivariate statistical analysis of the developing human brain in preterm infants

“…In order to avoid both the singularity and instability critical issues of the within-class scatter matrix S w when LDA is used in such limited sample and high dimensional problem, we have proposed an LDA-based approach based on a straightforward covariance selection method for the S w matrix [42,45].…”

Regularisation of eigenfeatures by extrapolation of scatter-matrix in face-recognition problem

“…This way a more meaningful extrapolated curve could be obtained which will provide a reliable estimate ofŜ W . Using the arguments given in [7], we can select the predominant eigenvalues either as those eigenvalues which are greater than the average of all the eigenvalues in D W or by those which preserve 50% of the total sum of all the eigenvalues in D W . We found the later procedure to be better in the present classification task and, hence, it is used in this…”

“…The dimension size is 4980. The proposed method was compared with the following methods: the range space method, the null space based method (OLDA) [4], the PCA plus LDA method [10], the regularised LDA method [5] and the maximum uncertainty LDA (MLDA) method [7]. Table 1 shows the recognition accuracy on both the datasets for all the methods.…”