A theorem on the uncorrelated optimal discriminant vectors

This paper proposes a theorem on the uncorrelated optimal discriminant vect ors (UODVs). It is proved that the classical optimal discriminant vectors a re equivalent to UODV, which can be used to extract (L - 1) uncorrelated di scriminant features for L-class problems without losing any discriminant in formation in the meaning of Fisher discriminant criterion function. Experim ents on Concordia University CENPARMI handwritten numeral database indicate that UODVs are much more powerful than the Foley-Sammon optimal discrimina nt vectors. It is believed that when the number of training samples is larg e, the conjugate orthogonal set of discriminant vectors can be much more po werful than the orthogonal set of discriminant vectors. (C) 2001 Pattern Re cognition Society. Published by Elsevier Science Ltd. All rights reserved.