Figures of Merit for Biometric Verification and a Means for Dimension Reduction

Optimum reduction of the number of dimensions of the feature vector in biometric verification requires a figure of merit for the verification performance. We show that, under certain conditions, various figures of merit can be approximated by the same expression. We select one, the discrimination distance, as an objective function for dimension reduction. This new method for dimension reduction is called maximum discrimination analysis (MDA). It is shown to have a better verification performance than linear discriminant analysis.