In this paper, we introduce a new large handwritten Arabic Digits dataBase (ADBase). The ADBase is composed of 60,000 digits for training and 10,000 digits for testing written by 700 persons of different ages and educational backgrounds. Also, a recognition system for handwritten Arabic digits with a recognition rate of 99.15% and low recognition time is introduced. The system is composed of two stages. The first stage is an Artificial Neural Network (ANN) fed with a short powerful feature vector for fast classification of non-ambiguous cases. The first stage has a reject option to pass the ambiguous cases to the more powerful, second stage. The Second stage is slow, yet powerful Support Vector Machine (SVM) fed with a large feature vector to classify the ambiguous cases rejected from the first stage.
展开▼
