A Benchmark Study of Classical and U-Net ResNet34 Methods for Binarization of Balinese Palm Leaf Manuscripts
Year: 2025
Type: article
Abstract: Ancient documents that have undergone physical and visual degradation pose significant challenges in the digital recognition and preservation of information. This study adopts a digital forensic approach to segment the Terumbalan palm leaf manuscript preserved by the National Library of Indonesia, which contains symbolic illustrations and traditional Balinese script. The research aims to evaluate and compare the effectiveness of ten thresholding-based segmentation methods in extracting textual content from digitized palm leaf manuscripts. A total of 15 high-resolution images were used, with preprocessing steps including grayscale conversion and median filtering. To assess segmentation performance, ground truth masks were created and evaluated. The results indicate that the locally adaptive Sauvola method consistently outperforms the others, achieving the highest scores across both accuracy and perceptual quality metrics—specifically, an IoU of 0.934, a Recall of 0.971, and an MS-SSIM of 0.800. While Li’s cross-entropy method produced competitive recall performance, it demonstrated lower structural fidelity. In contrast, methods such as Niblack and K-Means yielded poor results due to fragmentation and high sensitivity to noise. These findings highlight the effectiveness of locally adaptive thresholding in recovering faint and degraded scripts in historical documents. Furthermore, this study contributes a replicable digital image forensic analysis and evaluation framework to support future digital preservation efforts involving visually complex manuscripts.
Source: Heritage
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,930
Funder: Universitas Airlangga
Award: G7482927884
OpenAlex ID: W4413279258