On the maximization of divergence in pattern recognition (Corresp.)

This correspondence considers the problem of maximization of the divergence between a pair of unequal mean and unequal covariance matrix Gaussian distributed pattern classes. The original pattern space is transformed into a new space such that the sum of the covariance matrices is a unit matrix. From this relationship, a set of orthonormal directions are obtained sequentially such that, when the patterns are projected onto each of these directions, the divergence between the pattern classes is maximized.