A multiple feature/resolution scheme to Arabic (Indian) numerals recognition using hidden Markov models

This paper describes a technique for the recognition of optical off-line handwritten Arabic (Indian) numerals using hidden Markov models (HMM). Features that measure the image characteristics at local, intermediate, and large scales were applied. Gradient, structural, and concavity features at the sub-regions level are extracted and used as the features for the Arabic (Indian) numeral. Several experiments were conducted for estimating the suitable number of image divisions, and the best combination of features using the HMM classifier. A number of experiments were conducted to estimate the best number of states and codebook sizes in terms of the highest recognition rate possible. In this work, we did not follow the general trend of using the sliding window technique with HMM. Instead, a multi-resolution feature extraction approach was implemented on the whole digit. A database of 44 writers, with 48 samples per digit resulting in a database of 21 120 samples was used. The achieved average recognition rate is 99%. The classification errors were analysed and attributed to bad data, different writing styles of some digits, errors between digit pairs, and genuine errors. The presented technique, which is writer independent, proved to be effective in the automatic recognition of Arabic (Indian) numerals.