Handwritten Arabic numerals recognition using multi-span features & Support Vector Machines

In this work, a technique for handwritten Arabic (Indian) numerals recognition using multi-span features is presented. Angle, ring, horizontal, and vertical span features are used. All combinations of these features are tested and the combinations that result in the best recognition rates using Support Vector Machine (SVM) are identified. The SVM classifier is trained with 15840 digits and tested with the remaining 5280 digits. It is shown that the recognition rates using angle & horizontal span features achieved better recognition rates than all other combinations including using all features. The recognition rates of SVM are compared with published results using Hidden Markov Model (HMM) and the Nearest Mean (NM) classifiers. The achieved average recognition rates are 99.4%, 97.99% and 94.35% using SVM, HMM and NM classifiers, respectively. The use of SVM and angle & horizontal span features give the highest recognition rates and are superior to HMM and NM classifiers for all digits.