New Transformer Approach to the Recognition of Mediaeval Arabic Historical Manuscripts

The transcription of historical Arabic manuscripts plays a crucial role in preserving cultural heritage and advancing academic research. By minimizing the physical handling of originals, the transcription of historical Arabic manuscripts contributes to the physical preservation of these manuscripts, thereby reducing the risks of physical deterioration. In this article, we present a new approach using a Decoder-Encoder deep learning model that combines a convolutional network (ResNet34) and a Transformer network for recognition text lines in Historical Arabic Manuscripts. Confronting specific challenges such as the cursive nature of Arabic writing and the degradation of ancient documents, the hybrid model leverages the capacity and robustness of ResNet34 to extract visual features and the power of the Transformer to sequentially model the data. Our experiments reveal that our model surpasses current state-of-the-art results on the VML-HD dataset, providing a Character Error Rate (CER) of 0.0131 and a Word Error Rate (WER) of 0.0133.