Palm Leaf Digitization

Abstract

Palm leaf manuscripts have rich sources of knowledge and information reflecting cultural, historical, and linguistic knowledge. Extracting information from palm leaf manuscripts poses significant challenges for preservation and access, as they are fragile in nature. We propose an advanced multimodal deep learning framework for the digitization, character reconstruction, and decentralized federated learning of palm leaf manuscripts. The proposed approach integrates Transformer-based OCR models (TrOCR, LayoutLM), Vision Transformers (ViTs), and Contrastive Language-Image Pre-training (CLIP) to enhance character recognition for damaged and missing characters in the manuscripts. Natural Language Processing Algorithms are implemented to restore incomplete or faded characters while preserving the originality of the manuscripts. To ensure secure and decentralized access, we employ a blockchain-based federated learning system where metadata, translations, and reconstructed text are securely stored on a Zero-Knowledge Proof (ZKP) blockchain ledger. Federated learning across distributed nodes minimizes the centralized dependencies while enabling real-time collaborative OCR model updates. Scalability is ensured by Docker and Kubernetes where real-time processing is done across distributed nodes. Experimental results demonstrate superior OCR accuracy (96.3%), improved character restoration fidelity (92.7%), and enhanced blockchain security with minimal overhead (4.2%), outperforming traditional methods. The proposed method stores the manuscripts in a digitized form, providing easy access for researchers and scientists globally. The proposed work unlocks the hidden treasures, knowledge and information from the cultural treasures. Results obtained show that the efficiency and scalability of the proposed approach paves a path into the digital era by enhancing cultural preservation.