Discriminant face images taking both the feature correlation and Fisher criterion into account

Xu et al (A novel method for Fisher discriminant Analysis. Pattern Recognition, 2004, 37 (2): 381-384) proposed an interesting Fisher discriminant analysis method that not only requires that discriminant vectors have large Fisher criterion values but also requires that resultant features have small statistical correlation to each other. Indeed, this method is motivated by the following cognition: greater the Fisher criterion value is, more powerful the discriminant vector is and a set of uncorrelated features allow the set of features to contain more information of the original samples. However, when applied to appearance-based face recognition it encounters two problems. The first problem is that scatter matrices involved in the algorithm usually are very large and memory consuming. The second problem is that the method might suffer the small sample size (SSS) problem. In this paper, in order to overcome the two problems and take advantage of the idea of the method we propose a new discriminant method that is directly applicable for image data. This new method not only is free of the SSS problem but also achieves a good recognition performance.