Bernoulli HMMs for Off-line Handwriting Recognition

Adrià Giménez-Pastor, Alfons Juan-Císcar

2008

Abstract

Hidden Markov models (HMMs) are widely used in off-line handwriting recognition to model the probability (density) of an observation sequence, given its corresponding text transcription. Observation sequences typically consist of fixed-dimension feature vectors which are computed locally, using a sliding window along the handwritten text image. However, there is no standard set of local features being used by most of the systems proposed. In this paper we explore the possibility of raw, binary pixels instead of “complicated” features. To this purpose, we propose the use of Bernoulli HMMs, that is, HMMs in which the state-conditional probability (density) function is not a conventional Gaussian (mixture) density, but a multivariate Bernoulli (mixture) probability function. Promising empirical results are reported on two tasks of handwriting word recognition.

Download


Paper Citation


in Harvard Style

Giménez-Pastor A. and Juan-Císcar A. (2008). Bernoulli HMMs for Off-line Handwriting Recognition . In Proceedings of the 8th International Workshop on Pattern Recognition in Information Systems - Volume 1: PRIS, (ICEIS 2008) ISBN 978-989-8111-42-5, pages 86-92. DOI: 10.5220/0001740400860092


in Bibtex Style

@conference{pris08,
author={Adrià Giménez-Pastor and Alfons Juan-Císcar},
title={Bernoulli HMMs for Off-line Handwriting Recognition},
booktitle={Proceedings of the 8th International Workshop on Pattern Recognition in Information Systems - Volume 1: PRIS, (ICEIS 2008)},
year={2008},
pages={86-92},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0001740400860092},
isbn={978-989-8111-42-5},
}


in EndNote Style

TY - CONF
JO - Proceedings of the 8th International Workshop on Pattern Recognition in Information Systems - Volume 1: PRIS, (ICEIS 2008)
TI - Bernoulli HMMs for Off-line Handwriting Recognition
SN - 978-989-8111-42-5
AU - Giménez-Pastor A.
AU - Juan-Císcar A.
PY - 2008
SP - 86
EP - 92
DO - 10.5220/0001740400860092