Features Modelling in Discrete and Continuous Hidden
Markov Models for Handwritten Arabic Words Recognition
1LabSTIC,
University of 8 Mai 1945 of Guelma, Algeria
2LIASD,
University Paris 8, France
Abstract: The arab writing is originally cursive, difficult to segment and has a
great variability. To overcome these problems, we propose two holistic
approaches for the recognition of the handwritten arabic words in a limited
vocabulary based on the Hidden Markov Models (HMMs): discrete with wk-means and
continuous. In the suggested approach, each word of the lexicon is modelled by
a discrete or continuous HMM. After a series of pre-processing, the word image
is segmented from right to left in succession frames of fixed or variable size
in order to generate a sequence vector of statistical and structural parameters
which will be submitted to two classifiers to identify the word. To illustrate
the efficiency of the proposed systems, significant experiments are carried out
on IFN/ENIT benchmark database.
Keywords: Recognition of the handwritten arabic words,
holistic approach, DHMMs, CHMMs, k-means, wk-means, algorithm of Viterbi, modified
EM algorithm.