Hybrid CNN transformer and BiLSTM CRF pipeline for text restoration in Tamil palm leaf manuscripts

Ancient Tamil palm leaf manuscripts contain valuable medical and cultural knowledge but are challenging to digitize due to their fragility, scriptio-continua writing style, and systematic omission of diacritics. This work presents an end-to-end pipeline for automated transcription and diacritic restoration. The pipeline introduces a hybrid Convolutional Neural Network and Transformer model that combines local feature extraction with global context modeling, achieving 97.10% accuracy in classifying 125 Tamil characters and outperforming CNN-only, RNN-only, and Transformer-only baselines. Following character recognition, a Bidirectional Long Short-Term Memory with Conditional Random Fields model performs sequence labeling to restore missing diacritics and word boundaries, leveraging bidirectional context and enforcing coherent label transitions to achieve a Character Error Rate of 2.33. Together, these modules form a scalable system for digitizing fragile, low-resource historical Tamil manuscripts into accurate, human-readable text.