Back Propagation Neural Network Arabic Characters Classification Module Utilizing Microsoft Word

Abstract

Problem statement: Arabic character recognition has been one of the last major languages to receive attention. This may be attributed to the inherent complexity of both printed and handwritten Arabic characters. The objectives of this study were to: (i) summarize the main characteristics of Arabic language writing style. (ii) suggest a neural network recognition circuit. Approach: A Neural network with back propagation training mechanism for classification was designed and trained to recognize any set of character combinations, sizes or fonts used in Microsoft word. Results: The proposed network recognition behaviours were compared with perceptron-like net that combines perceptron with ADALINE features. These circuits were tested for three character sets combinations; 28 basic Arabic characters plus 10 numerals set, 52 Latin characters and 10 numerals only. Conclusions: The method was robust and flexible and can be easily extended to any character set. The network exhibited recognition rates approaching 100% with reasonable noise tolerance.