Segmenting Dead Sea Scroll Fragments for a Scientific Image Set

TL;DR

The paper tackles the challenge of segmenting Dead Sea Scroll fragments from IAA image archives, where bars, labels, a black background, and backing substrates complicate standard segmentation. It introduces a multi-step, tailor-made pipeline that combines Faster R-CNN bar detection with image-alignment, infrared thresholding, and backing-substrate masking to produce accurate fragment masks, followed by contour extraction and filtering. A public evaluation dataset of 139 fragments with train/validation/test splits and ground-truth WKT annotations enables reproducible performance assessment, yielding high mean IoU ($0.97$) and precision/recall around $0.98$–$0.99$. The workflow is designed to be generalizable to other manuscript collections and provides a practical foundation for digital joins, reconstruction, and glyph-level analysis while promoting standards for benchmarking in fragment segmentation.

Abstract

This paper presents a customized pipeline for segmenting manuscript fragments from images curated by the Israel Antiquities Authority (IAA). The images present challenges for standard segmentation methods due to the presence of the ruler, color, and plate number bars, as well as a black background that resembles the ink and varying backing substrates. The proposed pipeline, consisting of four steps, addresses these challenges by isolating and solving each difficulty using custom tailored methods. Further, the usage of a multi-step pipeline will surely be helpful from a conceptual standpoint for other image segmentation projects that encounter problems that have proven intractable when applying any of the more commonly used segmentation techniques. In addition, we create a dataset with bar detection and fragment segmentation ground truth and evaluate the pipeline steps qualitatively and quantitatively on it. This dataset is publicly available to support the development of the field. It aims to address the lack of standard sets of fragment images and evaluation metrics and enable researchers to evaluate their methods in a reliable and reproducible manner.

Figures (13)

• Figure 1: Plot of average classification and regression losses of Faster R-CNN for bar detection on IAA images of fragments versus training and validation iterations.
• Figure 2: Color images of both recto and verso of a fragment on the left, with bar detection results overlaid. The images are also overlaid with ground truth in yellow. The bar detection was performed only on color images, as the network was trained on these. The results can be used for the infrared images on the right as well, as the bar positions are identical in both image types. The two strips of Japanese paper, or hinges, used to hold the fragment in place can also be seen at the top. Additional paper was used by conservators to reinforce the verso, and is visible through a hole in the parchment near the top of the fragment. Images courtesy of Leon Levy Dead Sea Scrolls Digital Library, Israel Antiquities Authority; photo: Shai Halevi.
• Figure 3: The $y$-axis represents the average number of inliers across the test images, while the $x$-axis represents the threshold value. Each plot line represents the results for a different feature extractor, with the highest number of inliers being returned using SIFT with a threshold value of $15$.
• Figure 4: Recto and flipped verso infrared images, with inliers depicted as red circles and their corresponding match connections as colored lines. Images courtesy of Leon Levy Dead Sea Scrolls Digital Library, Israel Antiquities Authority; photo: Shai Halevi.
• Figure 5: Image Thresholding. Images courtesy of Leon Levy Dead Sea Scrolls Digital Library, Israel Antiquities Authority; photo: Shai Halevi.