Deep Learning Handwritten Text Recognition from Ancient Manuscripts using Hybrid CNN Transformer with PARSeq Model
A Jenefa, B Srinithi, T.M. Thiyagu, Catherine Joy R, P. Santhiya, Vidhya K · 2025
Historical documents contain a lot of knowledge; however, they are mostly inaccessible due to their fragile condition and complex ancient scripts. HTR transforms them into machine-readable data, making it much easier to use them for research. The old methods used in HTR are unable to properly understand the eccentricities of ancient manuscripts such as the diversity of scripts and the physical damage to manuscripts. Current methods mostly use regular convolutional neural networks (CNNs) or recurrent neural networks (RNNs), which frequently do not capture the extensive relations between sequential and spatial features of handwritten texts. In this paper, we propose a new hybrid CNN-Transformer model incorporating the PARSeq framework for improved feature extraction and sequence modeling for ancient scripts. We use a large dataset of more than 3000 images of Devanagari manuscripts that are classified as clean, degraded and heavily annotated from different archives and repositories. The proposed model show substantial improvements in terms of recognition accuracy by 98.5% on clean manuscripts and 92% on degraded documents. It substantially outperformed existing models. The HTR system with PARSeq helps to integrate traditional HTR systems with ancient manuscript recognition requirements to facilitate the work of historians and linguists to carry forward cultural preservation efforts.