Face classification using a multiresolution principal component analysis

Multiresolution principal component analysis (M-PCA) uses principal component analysis (PCA) to obtain multiresolution features for a signal. Bischof (1995) and Bischof and Hornik (1996) used 3-layer networks to train principal component pyramids for image compression. M-PCA uses a single computational layer adaptive linear network trained with the generalized Hebbian algorithm (GHA). The multiresolution features were applied to automatic face recognition and tested against the Olivetti Research Lab database. Classification with multiresolution had an average (over 10 runs) error rate of 2.4%