Biometric dispersion matcher versus LDA

This paper compares the recently developed biometric dispersion matcher (BDM) with the classical linear discriminant analysis (LDA) for biometric pattern recognition. BDM is extended to the BDM with simultaneously diagonalization (BDMSD) of the covariance matrices and, from a theoretical point of view, it is demonstrated that the feature selection of LDA and BDMSD are equivalent. However, LDA uses the between-class scatter matrix (S_B) only for feature selection and BDMSD also uses it for classification. This implies a set of advantages. Mainly the BDMSD offers better generalization capability for classifying samples of users that have not been used for training the classifier. Experimental results show that BDM and BDMSD outperform LDA in face recognition and hand-geometry recognition. These two cases correspond to very different situations: number of samples greater than their dimensionality (hand-geometry) and number of samples similar to their dimensionality (face recognition).