Class-oriented recognizer design by weighting local decisions

Abstract

The paper presents a new technique for the design of class-oriented recognizer. For each recognizer, a generic technique is used to determine, in an optimal way, the weights to balance the local decisions obtained from the analysis by parts of the patterns of the specific class. The experimental results, that have been obtained in the field of handwritten numeral and character recognition, demonstrate the superiority of the new technique with respect to other traditional approaches.