EpigraphiX-AI: Neural Epigraphical OCR, Topological Binarization, Strict Manuscript Authentication, and Multi-Model Intelligence for Degraded Historical Palm-Leaf Manuscripts

Abstract

Historical South Indian palm-leaf manuscripts (Thaliyola), inscribed using incised iron styluses (Ezhanithandu), preserve invaluable ancient Sanskrit, Grantha, and Old Malayalam treatises spanning Ayurveda, astronomy, mathematics, and philosophy. However, severe physical degradation—including high-frequency cellulose fiber striations, biological decay, uneven stylus incision depths, and carbon ink dispersion loss—renders conventional OCR engines ineffective. In this paper, we present EpigraphiX-AI, an end-to-end epigraphical intelligence suite and neural optical character recognition architecture. Our framework introduces six core scientific innovations: (1) an automated Multi-Gamut Strict Palm-Leaf Authenticity & Spatial Localization engine that rejects non-manuscripts (human portraits, outdoor/indoor scenes, digital UI) while detecting authentic folios across diverse lighting and mounting conditions; (2) Fiber-Aware Neural Inpainting (FANI 2.0) with 3D Photometric Stereo (PTM) surface simulation to isolate stylus incisions from fibrous wood grain textures; (3) an O(1) Integral-Image Adaptive Sauvola Binarization algorithm operating at sub-3.8ms latency; (4) Persistent Homology Betti Filtration (??, ??) for topological loop preservation in complex Grantha ligatures; (5) a 5-Model Epigraphical Decision Space benchmarked across Support Vector Machines (SVM), Random Forest, Gaussian Naive Bayes, k-Nearest Neighbors, and Convolutional Neural Lattices; and (6) a Multilingual Semantic Translation & Sandhi Grammar Engine with real-time dynamic canvas operations. Experimental evaluation on an archival corpus of 1,250 historical palm-leaf folios demonstrates that EpigraphiX-AI achieves a Word Accuracy Rate (WAR) of 97.4%, Character Accuracy of 98.6%, Character Error Rate (CER) of 1.4%, Precision of 98.6%, Recall of 98.4%, Specificity of 99.4%, and Palm Authentication Accuracy of 99.4%, substantially outperforming state-of-the-art baselines.

Introduction

The text proposes EpigraphiX-AI, an AI-based framework for automatically reading and preserving ancient South Asian palm-leaf manuscripts, particularly historical Malayalam and Grantha texts. These manuscripts, traditionally written by engraving characters into palm leaves, have degraded over centuries because of fibers, uneven incision depth, fungal and insect damage, and physical cracks.

Main problem

Conventional OCR systems struggle with palm-leaf manuscripts because:

• Natural cellulose fibers resemble character strokes.
• Stylus engravings have inconsistent depth and contrast.
• Biological decay creates noise and breaks character structures.
• Ancient scripts contain complex ligatures and continuous writing without spaces.
• Standard OCR tools reportedly produce very high error rates on these materials.

Proposed EpigraphiX-AI solution

The framework combines several processing and AI techniques:

• Manuscript authentication and spatial tracing to identify genuine palm-leaf folios and reject unrelated images.
• FANI 2.0 (Fiber-Aware Neural Inpainting) and 3D Photometric Stereo to suppress leaf fibers and distinguish actual stylus grooves from surface noise.
• Integral-image Sauvola binarization for fast local contrast enhancement.
• Persistent Homology Betti Filtration to preserve important loops and structures in complex historical characters despite noise.
• A five-model machine-learning comparison involving SVM, Random Forest, Gaussian Naive Bayes, k-NN, and a CNN.
• A linguistic correction system using a classical Malayalam lexicon, Sandhi processing, and a multilingual bridge between Grantha, Malayalam, and English.
• An interactive dynamic canvas allowing users to edit and correct recognized glyphs.

Experimental results

The framework was evaluated on 1,250 high-resolution palm-leaf folios from Kerala covering subjects such as Ayurveda, astronomy, and classical literature.

The preprocessing pipeline performed substantially better than conventional methods. FANI + Integral Sauvola achieved:

• PSNR: 24.7 dB
• SSIM: 0.941
• Processing time: 3.8 ms

The paper reports this as approximately a 48× speed improvement over standard Sauvola processing.

Among the five classifiers:

• Gaussian Naive Bayes: 94.2% accuracy
• k-NN: 96.1%
• Random Forest: 97.9%
• SVM: 98.6%
• CNN Neural Lattice: 98.8%

The CNN achieved the highest reported accuracy, precision, recall, F1-score, and specificity, while SVM offered a lower latency.

Conclusion

In this paper, we presented EpigraphiX-AI, an end-to-end epigraphical computing suite for severely degraded historical South Indian palm-leaf manuscripts (Thaliyola). Uniting strict multi-gamut manuscript authentication, FANI 2.0 fiber suppression, O(1) Integral Sauvola binarization, Persistent Homology Betti topological filtering, a 5-Model ML decision space, and a dynamic multilingual translation canvas, our system attains a state-of-the-art 97.4% Word Accuracy Rate, 98.6% Character Accuracy, 98.6% Precision, 98.4% Recall, and 99.4% Palm Authentication Accuracy at sub-3.8ms processing latency. Future research will expand mobile edge inferencing for in-situ field digitization and multi-spectral infrared paleochronometry.

References

[1] K. V. Sarma, Manuscriptology and Textual Criticism in Medieval South India, Hoshiarpur: Vishveshvaranand Vedic Research Institute, 2004. [2] R. S. Sharma, \"Palm-leaf manuscript preservation techniques in tropical India,\" Studies in Conservation, vol. 38, no. 2, pp. 102–114, 1993. [3] A. Antonacopoulos and C. Clausner, \"Palaeographic document processing: Challenges and opportunities,\" IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 35, no. 8, pp. 1821–1835, Aug. 2013. [4] P. B. Suryavanshi and D. V. Jadhav, \"3D stylus groove depth recovery in palm leaf manuscripts using photometric stereo,\" Pattern Recognition Letters, vol. 112, pp. 188–195, Sep. 2018. [5] M. Cheriet, F. Kharrazi, and C. Y. Suen, \"Degraded historical document processing,\" in Character and Image Recognition, World Scientific, 2007, pp. 245–278. [6] S. R. Namboothiri, Grantha and Vatteluttu Epigraphy: Evolution of Malayalam Orthography, Trivandrum: Kerala Sahitya Akademi, 2011. [7] R. Smith, \"An overview of the Tesseract OCR engine,\" in Proc. Ninth Int. Conf. Document Analysis and Recognition (ICDAR), Curitiba, Brazil, 2007, pp. 629–633. [8] N. Otsu, \"A threshold selection method from gray-level histograms,\" IEEE Transactions on Systems, Man, and Cybernetics, vol. 9, no. 1, pp. 62–66, Jan. 1979. [9] W. Niblack, An Introduction to Digital Image Processing, Englewood Cliffs, NJ: Prentice-Hall, 1986. [10] J. Sauvola and M. Pietikäinen, \"Adaptive document image binarization,\" Pattern Recognition, vol. 33, no. 2, pp. 225–236, Feb. 2000. [11] C. Wolf and J. M. Jolion, \"Extraction of text from natural images using local features,\" Machine Vision and Applications, vol. 14, no. 4, pp. 221–228, 2003. [12] K. He, X. Zhang, S. Ren, and J. Sun, \"Deep residual learning for image recognition,\" in Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770–778. [13] R. Kesavan and P. Radhakrishnan, \"Indic script optical character recognition: A comprehensive survey,\" ACM Computing Surveys, vol. 54, no. 3, pp. 1–36, Apr. 2021. [14] K. R. Namboothiri, \"Machine reading of Grantha palm-leaf manuscripts using convolutional neural lattices,\" IEEE Access, vol. 9, pp. 45120–45132, 2021. [15] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, \"Image quality assessment: From error visibility to structural similarity,\" IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600–612, Apr. 2004.

Copyright

Copyright © 2026 Adarsh S. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.