An arabic OCR using neural network classifiers

Abstract

A typewritten Arabic OCR (Optical Character Recognition system) is introduced. The system automatically predicts the size of the font, and uses it in separating lines, words and subwords. Then, it scans the text in a way similar to what the Arabic readers do to recognize and segment the different characters. A set of nine Neural Network modules is used for this process guided by a novel algorithm. The whole text is, then, rebuilt using the recognized characters and some error correction loops are applied. Using Neural Network classifiers allows an efficient hardware implementation and good “generalization” abilities.