Share Email Print
cover

Proceedings Paper

Multicriteria character classification using Pareto optimality
Author(s): Radovan V. Krtolica
Format Member Price Non-Member Price
PDF $14.40 $18.00

Paper Abstract

Robust recognition of machine-printed characters may involve several independent methods. For instance, available commercial PC-oriented software packages for optical character recognition are able to achieve high performance, each of them on a specific data set. Assuming that each of the underlying approaches to character recognition provides at least partial ordering of the available templates for a given character image, it seems possible to integrate these different approaches so that the integrated system exceeds the individuals in performance. Classification of patterns according to different criteria may lead to conflicting individual decisions. This paper discusses the use of the concept of Pareto-optimality to resolve these conflicts. It is proposed to eliminate the inferior candidates (i.e., the candidates that have better alternatives with regard to all of the criteria). The remaining candidates are noninferior or Pareto-optimal. The Pareto selection mechanism reacts to ranking contradictions by extending the set of solutions. In that way, it decreases the plausibility to eliminate the correct solution (in statistical terms, to commit an error of the first kind). The plausibility to adopt an incorrect solution (to commit an error of the second kind) increases with the cardinality of the Pareto set, and hence it increases with decrease in consistency of the ranking criteria. When consistency of the multicriteria rankings is satisfactory, additional information is sometimes needed to choose a unique solution among the elements of the Pareto set.

Paper Details

Date Published: 14 April 1993
PDF: 8 pages
Proc. SPIE 1906, Character Recognition Technologies, (14 April 1993); doi: 10.1117/12.143625
Show Author Affiliations
Radovan V. Krtolica, Canon Research Center America, Inc. (United States)


Published in SPIE Proceedings Vol. 1906:
Character Recognition Technologies
Donald P. D'Amato, Editor(s)

© SPIE. Terms of Use
Back to Top