Evaluation of prototype learning algorithms for nearest-neighbor classifier in application to handwritten character recognition

摘要

Prototype learning is effective in improving the classification performance of nearest-neighbor (NN) classifier and in reducing the storage and computation requirements. This paper reviews some prototype learning algorithms for NN classifier design and evaluates their performance in application to handwritten character recognition. The algorithms include the well-known LVQ and some parameter optimization approaches that aim to minimize an objective function by gradient search. We also propose some new algorithms based on parameter optimization and evaluate their performance together with the existing ones. Eleven prototype learning algorithms are tested in handwritten numeral recognition on the CENPARMI database and in handwritten Chinese character recognition on the ETL8B2 database. The experimental results show that the algorithms based on parameter optimization generally outperform the LVQ. Particularly, the minimum classification error (MCE) approach of Juang and Katagiri (IEEE Trans. Signal Process. 40 (12) (1992) 3043), the generalized LVQ (GLVQ) of Sato and Yamada (Proceedings of the 14th ICPR, Vol. I, Brisbane, 1998, p. 322) and a new algorithm MAXP1 yield best results.