Online Farsi handwritten words recognition using a combination of 3 cascaded RBF neural networks

In this paper, we propose a method for online Farsi handwritten words recognition. At first, words are broken to their sub-words. Each sub-word is made of some strokes. The sign of the sub-word is found from the positions and shapes of its sub-strokes. After that, we classify sub-words according to their signs. Some online features are extracted from the main-stroke after the preprocessing stage. Preprocessing contains operations such as dehooking, smoothing, normalization and boundary size equalization. A combination of 3 cascaded RBF neural networks are learned and used in hierarchical recognition system. The first RBF net divides sub-words into classes, while the second one subdivides each class into sub-classes. The third RBF network recognizes sub-words in each sub-class. In this paper, we use a 1000-sub-word database of the most frequently used Farsi words. The performance of the system in the first and the second RBF classifiers is 99.7% and 98.9% respectively. The rate of correct performance of the third RBF net is 82.46% making the total recognition rate of the system on the database 81.3%.