On Linear Discriminant Analysis and its Variants in Face Recognition

Research areas:

• Volume 4,Issue 1, July 2015

Year:

2015

Type of Publication:

Article

Keywords:

Between Class Scatter, Linear Discriminant Analysis LDA, Within Class Scatter, Singularity

Authors:

Kai Li; Zhen Liu; Peng Tang

Journal:

IJAIM

Volume:

4

Number:

1

Pages:

29-34

Month:

July

Abstract:

Feature extraction is a key technology of face recognition, and it can directly affect the performance of the face recognition system. Among them, linear discriminant analysis (LDA) based on Fisher criterion is a classic and widely used method for the feature extraction, which takes the separability of the pattern data as its goal, and finds the optimal discriminant vectors that minimize the within-class scatter matrix and maximize the between-class scatter matrix. Nevertheless, as a kind of algebra feature extraction based on statistical techniques, the traditional LDA in the case of small sample size encounters two practical problems: one is the problem of singularity in the scatter matrix; the other is the problem of estimation error of the scatter matrix. In this paper, aiming at the problem of singularity, we make a detailed research and compare LDA with its variants such as PCA+LDA, D-LDA, R-LDA, N-LDA, O-LDA, U-LDA and 2D-LDA. Some experiments for the above algorithms are performed on the ORL and YALE face databases. And we analyze the advantages and disadvantages of each method from experimental perspectives.

Full text: IJAIM_466_Final.pdf