Arabic Handwriting Recognition at Line-Level: A Comprehensive Review and Hybrid Modeling Strategy*

Abstract

Arabic Handwriting Recognition (AHR) constitutes a challenging research problem primarily due to the cursive form of the script, the context-dependent character shape variations, and high inter-writer variability. We introduce a comprehensive review of recent progress on AHR, with a special emphasis on line-level recognition, where word- and character-level intricacies are folded into a smooth sequence. We discuss the development of AHR methods, from early segmentation-based pipelines to deep learning strategies with CNNs, RNNs, and, more recently, Transformer-based architectures. To illustrate the state of the art, we review the current systems, compare their architectures and performance, as well as their advantages and limitations. Motivated by these observations, in this work, we propose our ongoing work: a hybrid CNN-Transformer model which can leverage CNNs' strength in spatial feature extraction, and the long-range context modeling power of Transformers. In this, we try to increase the recognition accuracy and efficiency especially for real-life line-level handwriting situations.