A false acceptance error controlling method for hyperspherical classifiers

Controlling false acceptance errors is of critical importance in many pattern recognition applications, including signature and speaker veri$cation problems. Toward this goal, this paper presents two post-processing methods to improve the performance of hyperspherical classi$ers in rejecting patterns from unknown classes. The $rst method uses a self-organizational approach to design minimum radius hyperspheres, reducing the redundancy of the class region de$ned by the hyperspherical classi$ers. The second method removes additional redundant class regions from the hyperspheres by using a clustering technique to generate a number of smaller hyperspheres. Simulation and experimental results demonstrate that by removing redundant regions these two post-processing methods can reduce the false acceptance error without signi$cantly increasing the false rejection error. c © 2003 Elsevier B.V. All rights reserved.