Putu Ayu Febyanti, I Putu Agus Eka Darma Udayana, Aniek Suryanti Kusuma

An End-to-End Balinese Lontar OCR Framework Using Bayesian-Optimized Multiscale Retinex and MobileNetV3

Introduction

An end-to-end balinese lontar ocr framework using bayesian-optimized multiscale retinex and mobilenetv3. Digital preservation of Balinese lontar: An OCR framework using Bayesian-optimized Multiscale Retinex & MobileNetV3 for robust, efficient recognition of degraded manuscripts.

Abstract

The preservation of Balinese lontar manuscripts has become increasingly important due to their cultural, historical, and religious significance, while physical degradation such as uneven illumination, faded ink, texture interference, and manuscript aging continues to reduce readability and complicate digital preservation efforts. This study proposes an end-to-end Optical Character Recognition (OCR) framework for degraded Balinese lontar manuscripts by integrating Bayesian-optimized image enhancement, adaptive preprocessing, morphology-based segmentation, domain-specific augmentation, and lightweight deep learning recognition using MobileNetV3. The proposed enhancement pipeline combines Multiscale Retinex, adaptive gamma correction, edge-preserving filtering, and hybrid binarization to improve character visibility under degraded manuscript conditions. Bayesian Optimization with Optuna and Tree-structured Parzen Estimator (TPE) was employed to automatically optimize enhancement parameters according to manuscript quality characteristics. Experimental results demonstrated substantial improvements in manuscript image quality, where Laplacian Variance increased from 306.7596 to 6685.7641, RMS Contrast improved from 28.976 to 83.9085, Michelson Contrast increased from 0.8238 to 1.0, and Ink Ratio Score improved from 0.6096 to 0.9847. The MobileNetV3-based OCR recognition model achieved a test accuracy of 80.52% and a best validation accuracy of 83.78% across 102 Balinese script classes. The proposed framework demonstrates that adaptive enhancement optimization combined with lightweight OCR recognition can provide robust and computationally efficient recognition performance for degraded historical manuscripts while supporting scalable digital preservation and mobile-oriented cultural heritage applications.

Review

This study presents a timely and highly relevant end-to-end Optical Character Recognition (OCR) framework for the digital preservation of degraded Balinese lontar manuscripts, a critical task given their profound cultural, historical, and religious significance and the ongoing challenges posed by physical degradation. The paper appropriately identifies issues such as uneven illumination, faded ink, and texture interference as significant barriers to readability and effective digital archiving. By proposing an integrated solution that spans image enhancement to lightweight deep learning recognition, the authors address a niche yet vital problem within the domain of cultural heritage technology, offering a comprehensive strategy for overcoming these obstacles. The methodological approach is robust and well-articulated in the abstract, highlighting several innovative components. The core image enhancement pipeline, combining Multiscale Retinex, adaptive gamma correction, edge-preserving filtering, and hybrid binarization, directly targets the common degradation issues. A notable strength is the incorporation of Bayesian Optimization using Optuna and Tree-structured Parzen Estimator (TPE) to automatically tune enhancement parameters, an adaptive strategy that promises greater flexibility and effectiveness across varied manuscript conditions. The reported quantitative improvements in image quality metrics—such as a dramatic increase in Laplacian Variance and RMS Contrast, and near-perfect Michelson Contrast and Ink Ratio Score—substantiate the efficacy of this enhancement. For character recognition, the adoption of MobileNetV3 is a judicious choice, prioritizing computational efficiency and lightweight deployment, which is crucial for scalable digital preservation and mobile-oriented applications, achieving respectable accuracies of 80.52% (test) and 83.78% (validation) across 102 Balinese script classes. Overall, this framework represents a significant advancement in the digital preservation of Balinese lontar. Its end-to-end nature, coupled with the intelligent integration of adaptive enhancement techniques and an efficient deep learning model, makes a strong case for its practical applicability. The emphasis on robustness and computational efficiency is particularly commendable, paving the way for wider adoption in cultural heritage institutions. While the reported accuracies are encouraging, especially given the challenging nature of the data, future work could benefit from a more detailed analysis of the dataset size, specific error patterns, and a comparative evaluation against other state-of-the-art OCR systems designed for similar historical document types, even if not specifically Balinese. Nonetheless, this study provides a highly valuable contribution to both cultural heritage preservation and the field of document image analysis.

Full Text

Comments

You need to be logged in to post a comment.