Segmentation of Ink and Parchment in Dead Sea Scroll Fragments (github.com)

A new repository accompanying the preprint "Segmentation of Ink and Parchment in Dead Sea Scroll Fragments" releases both a dataset (Qumran Segmentation Dataset, QSD) and the Multispectral Thresholding and Energy Minimization (MTEM) method for separating ink and parchment in multispectral scans of Dead Sea Scroll fragments. The QSD contains multispectral data from 20 fragments—including full-color, first/last-band, and normalized last-band images—plus pixel-level annotations for rigorous evaluation. The work targets key historical-document challenges: low contrast between black ink and dark backgrounds or darkened parchment, and image degradation/noise, providing a ready benchmark and codebase for research in document restoration, paleography, and segmentation. Technically, MTEM combines multispectral thresholding with an energy-minimization step and achieves strong segmentation metrics: for ink segmentation MTEM records IoU 0.6713, precision 0.8935, recall 0.7029, F1 0.7676; for parchment MTEM achieves IoU 0.9764, precision 0.9945, recall 0.9818, F1 0.9877. A simple Otsu baseline fares much worse on parchment (IoU 0.5642, precision 0.5892). These results highlight the value of multispectral processing plus spatial regularization for separating subtle, degraded features and offer the AI/ML community a labeled multispectral benchmark and open-source implementation to develop and compare more advanced learning- and model-based segmentation approaches.