Arabic Handwriting Recognition Using Deep Learning: A Comprehensive Review
Abstract
Arabic handwriting recognition (AHR) remains challenging due to the context-dependent, cursive script that’s affected by diacritic marks, overlapping strokes, and high writer variability. This review presents an overview of deep learning approaches for AHR in both online and offline settings. The literature search was carried out using IEEE Xplore, SpringerLink, and Google Scholar, after screening titles, abstracts and reviewing the text of relevant studies, and the focused set were included. The most selected articles which meet the criteria were published between January 2015 and January 2026, in addition, one of the early articles which published on 2010 was included because it is representing the foundation for subsequent researches in this filed. This review covers recognition pipelines, and used datasets such as Arabic Handwritten Characters Dataset (AHCD), Institut für Nachrichtentechnik/ Ecole Nationale d'Ingénieurs de Tunis (IFN/ENIT), KHATT Handwritten Arabic Text Database (KHATT), Hijja, Modified Arabic Digit Database (MAD-Base), and Arabic Handwritten Writer Paragraphs Dataset (AHAWP), and major model families including convolutional neural networks (CNNs), convolutional recurrent neural networks (CRNNs), hybrid convolutional neural network-bidirectional long short-term memory models with connectionist temporal classification Convolutional Neural Network-Bidirectional Long-Short Memory/Connectionist Temporal Classification (CNN-BLSTM/CTC), attention-based methods, Explainable Artificial Intelligence (XAI) and Transformer-based recognizers. The reviewed literature shows that CNN models perform well on isolated characters, while hybrid sequence models are more suitable for cursive word and line recognition. Although cross-dataset generalization, segmentation difficulty, limited annotated Arabic datasets, resilience to degraded image, and computational cost added to the challenges. Evaluation under noise, self-supervised learning, lightweight deployment systems, and larger datasets ought to be pointed out in future researches.