An Arabic OCR Approach Using Levenshtein Distance and CNNs

Abstract

Arabic Handwritten Character Recognition (AHCR) systems encounter various challenges arising from the unique characteristics of the Arabic language and the limited availability of public databases. Consequently, numerous research endeavors have aimed to enhance the recognition accuracy of AHCR. In this study, we propose a solution inspired by the intricate functions of the human visual cortex and hippocampus. Our proposed system employs a segmentation method to break down Arabic characters, and a Convolutional Neural Network (CNN) is then utilized for character recognition. To recognize entire words, we employ the Levenshtein distance with a personalized database containing an extensive collection of Arabic words. Experimental results demonstrate that our system yields a word error rate ranging from 1% to 25%, contingent upon the number of accurately recognized characters.