التعرف على اﻷرقام الهنديه المخطوطه بإستخدام اّله المتجهات الداعمه

Abstract

Recent results in pattern recognition have shown that Support Vector Machine (SVM) classifiers often have superior recognition rate in comparison to other classification methods. Despite of this fact, the literature doesn't show intensive experiments for these classifiers on Arabic charcter recognition. This thesis, reports the experiments of one variant of SVM -namely One Against All- on a new Arabic (Indian) numeral data set (SUST ALT digits). SUST ALT digits data set is a new Arabic (Indian) numeral data set which contains 37,000 Almost digit images established by Arabic Recognition Group in Sudan University of Science and Technology. The Recognition error rate in these experiments is 7%. This error is not uniformaly distributed among the digits. For instance Quarter of this error occurred with two digits only (0&2). Therefore, more enhancements and post processing to remedy this localized error is expected to boost the classifier recognition rate.