Abstract

In this paper we present the BBN Flex multilingual OCR model configuration and experimentation platform which allows non-expert users to configure and test HMM models for the BBN Byblos OCR system. The Flex platform is designed to provide customers with model training capabilities for any machine printed language. Furthermore. training capabilities can be updated as new modeling techniques become available. This paper describes the HMM OCR modeling paradigm and the design ofthe Flex platform. Modeling techniques currently available on the platform as well as new techniques that we plan to add are discussed. Finally. we describe a test ofthe system on Korean machine printed data in which we achieved a character error rate of6. 84%.

Date

2005

Authors

Ehry MacRostie, Premkumar Natarajan

Journal

2005 Symposium on Document Image Understanding Technology

Pages

15