SwinArabert: An Encoder-Decoder Approach for Arabic Handwritten Text Recognition

Recent advancements in natural language processing have led to the development of highly efficient transformer-based models, yet Arabic handwritten text recognition remains challenging due to the script's cursive nature, character connectivity, and diacritical complexity. This paper introduces SwinArabert, a novel encoder-decoder architecture that integrates the Swin Transformer for localized feature extraction with AraBERT for robust linguistic modeling. By leveraging the hierarchical self-attention mechanisms of Swin and the contextual representations of AraBERT, SwinArabert effectively captures visual and linguistic patterns in Arabic handwriting. Our model is pretrained on a large-scale printed Arabic text dataset, King Fahd, and fine-tuned on KHATT, a diverse handwritten Arabic dataset, enabling it to generalize well across varying handwriting styles. Although our model does not achieve state-of-the-art because of the limitation of computation resources which leads to low training time, the model achieves a CER of 0.31, which is a high potential result for this low learning time.