Skip to Main content Skip to Navigation
Journal articles

Un modèle neuro markovien profond pour l’extraction de séquences dans des documents manuscrits

Abstract : In this paper, we propose a keyword extraction system able to extract keywords in handwritten documents. The base system rely on a HMM line model made of an Out-Of-KeyWord Vocabulary model and keywords model. In order to be more discriminant at the local level (the frame level), the standard gaussian mixture of the HMM are replaced by a deep neu-ral network (DNN) for computing the observations probabilities. Experimentations are carried out on an unconstrained handwritten document database used for the 2009 ICDAR handwriting recognition competitions. The results demonstrate the interest of the keyword extraction system as opposed to the sequential integration strategy of full text recognition prior to the detection of keywords. We also show the benefit from using the deep architecture instead of the gaussian mixtures.
Document type :
Journal articles
Complete list of metadata

Cited literature [27 references]  Display  Hide  Download

https://hal.inria.fr/hal-01105363
Contributor : Clément Chatelain <>
Submitted on : Tuesday, January 20, 2015 - 11:10:17 AM
Last modification on : Tuesday, November 24, 2020 - 9:38:05 AM
Long-term archiving on: : Tuesday, April 21, 2015 - 10:36:46 AM

File

docnum-thomas2013.pdf
Files produced by the author(s)

Identifiers

Citation

Simon Thomas, Clément Chatelain, Thierry Paquet, Laurent Heutte. Un modèle neuro markovien profond pour l’extraction de séquences dans des documents manuscrits. Document Numérique, Lavoisier, 2013, 16 (2), pp.20. ⟨10.3166/dn.16.2.49-68⟩. ⟨hal-01105363⟩

Share

Metrics

Record views

370

Files downloads

1097