Enhanced Multi-Head Attention Mechanism and BGRU for Lexicon-Free Online Arabic Handwriting Recognition

Abstract

Lexicon-free Online Arabic Handwriting Recognition (OAHR) remains a formidable challenge due to the script's highly cursive nature, vast character set, inherent stylistic variability, and the critical absence of word-level constraints. Capturing the complex long-range temporal dependencies inherent in online stroke data is key to improving recognition accuracy. While conventional Recurrent Neural Networks (RNNs) often fall short, Multi-Head Attention (MHA) mechanisms have emerged as superior context aggregators, offering both enhanced dependency capture and efficient parallel computation. Motivated by this, we propose an enhanced BGRU-MHA hybrid network for lexiconfree OAHR. This novel architecture effectively integrates stacked Bidirectional Gated Recurrent Units (BGRUs) with a MHA module to construct robust, context-aware feature representations followed by a Connectionist Temporal Classification (CTC) output layer for character sequence generation. Spatial and temporal features are first extracted from the preprocessed handwriting input, then refined through the BGRU–MHA hybrid layers, enabling the model to effectively capture both local and global dependencies. The proposed approach was evaluated on two benchmark datasets: ADAB and Online-KHATT. The results demonstrate that our method establishes a new state of the art on both datasets. Specifically, the model achieved a Character Error Rate (CER) of 3.86% and 6.79%, and a Word Error Rate (WER) of 11.98% and 20.36% on ADAB and Online-KHATT, respectively. These results confirm that integrating MHA with BGRU significantly improves recognition accuracy, validating the effectiveness of the proposed lexicon-free OAHR framework.