Ensemble methods to improve the performance of an english handwritten text line recognizer

Abstract

This paper describes recent work on ensemble methods for offline handwritten text line recognition. We discuss techniques to build ensembles of recognizers by systematically altering the training data or the system architecture. To combine the results of the ensemble members, we propose to apply ROVER, a voting based framework commonly used in continuous speech recognition. Additionally, we extend this framework with a statistical combination method. The experimental evaluation shows that the proposed ensemble methods have the potential to improve the recognition accuracy compared to a single recognizer. © 2008 Springer-Verlag Berlin Heidelberg.