Recognizing Arabic (Indian) Hand-written Digits by Using Markov Hidden Models

Abstract

Hidden markov models (HMMs) are stochastic models which are widely used in speech recognition. Later HMMs began to be applied in handwriting recognition.

This thesis tested the performance of HMMs on Indian handwritten digits recognition using SUST-ARG-digits dataset. SUST-ARG digits dataset is a newly established Indian handwritten numeral dataset collected by Sudan University of Science and Technology Arabic Recognition Group. Features are extracted from digits images using chain coding method.

HMMs trained using Baum-Welch algorithm by a training dataset of size 1350 samples of the nine digits (150 samples for each digit). The best number of states has been searched experimentally. The experiments show that a model with 9 states gives the best results (97.78% recognition rate) on a testing data of size 180 of the nine digits (20 samples for each digit).