Enhancing Ottoman word recognition via self-supervised pretraining using a siamese swin transformer

Official URL: https://dx.doi.org/10.1109/SIU66497.2025.11112258

The Ottoman archive contains millions of handwritten documents, and their efficient management and retrieval is of paramount significance to historians. In this paper, a semi-supervised method is presented for Ottoman word recognition from handwritten historical documents. The approach involves a two-stage process: self-supervised pretraining followed by super-vised training with triplet loss. A Siamese network architecture with a Shifted Window Transformer backbone is used to learn robust feature representations. It is shown that pretraining significantly enhances the model's ability to discriminate between similar Ottoman words. The method is evaluated on a historical dataset of labeled Ottoman word crops. To foster reproducibility and future research, we release the full code and dataset at https://github.com/Sadiq04/Ottoman-KWS-IEEE-SIU