Offline Arabic handwriting recognition using Hidden Markov Models and Post-Recognition Lexicon Matching

Arabic handwriting recognition is fairly complex operation due to the similarities between different letters under similar writing styles. This paper presents a new approach on offline recognition of handwritten Arabic words. The method does not require the segmentation of words into characters for recognition, but requires segmentation of training data into separate letters. The method trains a Hidden Markov Model (HMM) for each letter in the alphabet along with its various writing styles and taking into consideration the letter position variations. Having different types of features to extract from each image in the dataset helps to further improve the recognition rate of the whole system. The performance of the proposed method is demonstrated after various experiments carried out using the IFN/ENIT (Institut fur Nachrichtentechnik/ Ecole Nationale d'Ingénieurs de Tunis) reference database which contains an estimate of 32,492 handwritten Arabic words with various writing styles and formats of hundreds of writers. The overall recognition rate of the system is 87% after applying various methods of fine tuning and optimizations.