Application of artificail neural networks to optical character recognition
No Thumbnail Available
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Saudi Digital Library
Abstract
In this work we examine shape descriptors and neural network classifiers for the recognition of Arabic characters. There are a total of 29 characters in the Arabic alphabet. However, since the shape of character changes depending on its position in the word (beginning, middle, end, and stand-alone) we end up with more than 29 shapes. Shape descriptors are studied in this work. In particular the moment invariant approach due to Hu.[3], is studied and examined, with some modifications, on the Arabic alphabet. A new algorithm for training the feed forward neural network is developed in this thesis. This algorithm is shown to be faster and more stable than other schemes presented in the literature. The thesis presents the classification of shapes through shape descriptors and feed forward neural network classifiers. Testing on Arabic characters of different sizes and orientation is carried out. Four separate neural networks were trained for each character position with the seven moment invariants of Hu.[3] as an input. The output is trained to give five bits representing the ASCII code of Arabic letters plus one bit to serve as a parity check. Results of using the new training algorithm for the feed-forward neural network are presented. Very high recognition rate of Arabic fonts is achieved.