Reconnaissance de courriers manuscrits par HMMs contextuels et modèle de langage

Olivier Morillot (1) , Emmanuèle Grosicki (2) , Laurence Likforman-Sulem (1)

1 LTCI - Laboratoire Traitement et Communication de l'Information
2 DGA/CTA/DT/GIP - CEP Arcueil

This article describes a global approach for the recognition of handwritten lines which was submitted at the ICDAR 2011 French handwritten mail recognition competition. Its main contribution is to extend the context-dependent HMMs word recognition system to the line level without segmentation into words. Thus, the authors describes a complete solution for mail recognition: original preprocessing, context-dependent HMMs and an optimized bigram language model. Obtained results during the competition (73,2% of words correctly recognized) are very encouraging since they are similar to state-of-art recognition performances obtained with HMMs on single words.

Mots clés: Handwritten recognition; mail; text lines recognition; context-dependent HMMs; preprocessing; language model; bigram

Domaines: Traitement du texte et du document; Apprentissage [cs.LG]

Olivier Morillot, Emmanuèle Grosicki, Laurence Likforman-Sulem. Reconnaissance de courriers manuscrits par HMMs contextuels et modèle de langage. Semaine du document numérique et de la recherche d'information 2012 - Colloque international francophone sur l'écrit et le document, Mar 2012, Bordeaux, France. pp.23-36. ⟨hal-00737428⟩