An analysis of a multi-layered competitive net for invariant pattern recognition

This paper describes an analysis of a multi-layered competitive net for pattern recognition invariant to linear and/or nonlinear coordinate transformations such as projective transformations involving projections, translations, rotations, magnifications, and so on. The 2nd layer of the net performs a competitive learning of a number of modified patterns which are transformed by coordinate transformations and multiplied by the Jacobian of the transformations, which achieves the invariance to the transformations as well as retains the discrimination ability of different patterns. The invariant recognition of transformed test patterns is possible after the learning of original test patterns even if they contain different font patterns. Via computer simulation, we verify the ability of the invariance recognition by the net