Towards Robust Arabic and Urdu OCR Systems: A Systematic Review of Deep Learning Techniques

Abstract

In recent years, deep learning has increasingly replaced traditional machine learning algorithms across various domains, including Machine Translation (MT), Pattern Recognition (PR), Natural Language Processing (NLP), Speech Recognition (SR), and Computer Vision (CV). Notably, deep learning-based systems for optical character recognition (OCR) have demonstrated substantial success. However, within the domains of pattern recognition and computer vision, handwritten character recognition remains one of the most complex challenges. This complexity arises from the variability in character height, orientation, and width, as individuals employ diverse writing instruments and exhibit distinct writing styles. As a result, handwritten recognition becomes a particularly difficult task. Additionally, research on regional languages such as Arabic and Urdu remains relatively underexplored. This article presents a review and comparative analysis of the most significant deep learning techniques employed in the recognition of Arabic-adapted scripts, specifically focusing on Arabic and Urdu languages.