Pour citer ce document : |
URI:
|
http://hdl.handle.net/2042/32611
|
Title:
|
Reconnaissance de l’écriture manuscrite arabe et latine par des modèles de Markov cachés multi-Flux |
Author:
|
KESSENTINI, Yousri; PAQUET, Thierry; BEN HAMADOU, AbdelMajid
|
Abstract:
|
Dans cet article nous proposons une approche de reconnaissance de l’écriture manuscrite. L’objectif étant de
proposer un système indépendant de la nature du script, nous procédons alors sans segmentation. Des
caractéristiques bas niveaux, basées sur les directions des contours et les densités de pixels, sont combinées
à travers une approche multi-flux. Nous évaluons l’apport de l’approche multi-flux ainsi proposée et nous la
comparons aux approches classiques de combinaison par fusion de représentations et par fusion de
décisions. Pour valider l’approche proposée nous avons effectué des expérimentations sur deux bases de
données de référence, la base de mots arabes IFN/ENIT et la base IRONOFF de mots latins. Les résultats
montrent que le système proposé donne de bons résultats comparables aux meilleurs approches rapportées
dans la litérature, aussi bien sur le Latin que sur l’Arabe. |
Description:
|
In this paper, we present a multi-stream approach for off-line handwritten word recognition. The multi-stream formalism
presents many advantages: it can combine several kinds of independent features. The combination can be adaptive:
some sources of information can be weighted, or even rejected if they are not reliable. The topology of the HMM can be
adapted to each source of information. It also allows asynchronous modelling of streams.
The proposed approach combines low level feature streams namely, density based features extracted from 2 different
sliding windows with different widths, and contour based features extracted from upper and lower contours.
Significant experiments have been carried out on two publicly available word databases: IFN/ENIT benchmark database
(Arabic script) and IRONOFF database (Latin script).
In order to model the Latin characters, we built 26 uppercase character models and 26 lowercase character models). In the
case of Arabic characters, we built up to 159 character models. An Arabic character may actually have different shapes
according to its position within the word (beginning, middle, end word position). Other models are specified with additional
marks such as “shadda”. In both Latin and Arabic script, each character model is composed of 4 emitting states. The
observation probabilities are modelled with Gaussian Mixtures (3 per state). Embedded training is used where all character
models are trained in parallel using Baum-Welch algorithm applied on word examples. The system builds a word HMM by
concatenation of the character HMM corresponding to the word transcription of the training sample. The recognition step is doing allowing the HMM-recombination algorithm that consists in building the product HMM and
using a classical Viterbi decoding algorithm.We investigate the extension of 2-stream approach to N streams (N=2,...,4)
and analyze the improvement in the recognition performance. The computational cost of this extension is discussed.
The developed system has been tested on two publicly available databases. For both scripts the results show significant
improvement while using a multi-stream approach. The comparison of the multi-stream performances to the classical
combination strategies namely, fusion of features and fusion of decisions shows the superiority of the multi-stream
approach. Moreover, the proposed recognition system provides significant results comparable to the best results
reported in the literature on both databases.The recognition step is doing allowing the HMM-recombination algorithm that consists in building the product HMM and
using a classical Viterbi decoding algorithm.We investigate the extension of 2-stream approach to N streams (N=2,...,4)
and analyze the improvement in the recognition performance. The computational cost of this extension is discussed.
The developed system has been tested on two publicly available databases. For both scripts the results show significant
improvement while using a multi-stream approach. The comparison of the multi-stream performances to the classical
combination strategies namely, fusion of features and fusion of decisions shows the superiority of the multi-stream
approach. Moreover, the proposed recognition system provides significant results comparable to the best results
reported in the literature on both databases. The recognition step is doing allowing the HMM-recombination algorithm that consists in building the product HMM and
using a classical Viterbi decoding algorithm.We investigate the extension of 2-stream approach to N streams (N=2,...,4)
and analyze the improvement in the recognition performance. The computational cost of this extension is discussed.
The developed system has been tested on two publicly available databases. For both scripts the results show significant
improvement while using a multi-stream approach. The comparison of the multi-stream performances to the classical
combination strategies namely, fusion of features and fusion of decisions shows the superiority of the multi-stream
approach. Moreover, the proposed recognition system provides significant results comparable to the best results
reported in the literature on both databases. |
Subject:
|
Reconnaissance de l’écriture manuscrite hors-ligne, écriture arabe, écriture latine, combinaison d’information, modèles de Markov cachés multi-flux; Off-Line handwriting recognition, Hidden Markov Models, Latin script, Arabic script, multi-stream, information combination |
Publisher:
|
GRETSI, Saint Martin d'Hères, France |
Date:
|
2009 |