bibtype	A - Abstract
ARLID	0648373
utime	20260417125917.9
mtime	20260402235959.9
title (primary) (eng)	Herculaneum Scrolls Ink Detection Using Textural Features
specification	page_count 2 s. media_type P
serial	ARLID cav_un_epca*0648796 title CAA Proceedings : CAA 2026: It's all about People publisher place Vienna name CAA International year 2026
keyword	Herculaneum Scrolls
keyword	Ink detection
keyword	Vesuvius Challenge
author (primary)	ARLID cav_un_auth*0506964 name1 Korotetskyi name2 O. country UA
author	ARLID cav_un_auth*0101093 name1 Haindl name2 Michal institution UTIA-B full_dept (cz) Rozpoznávání obrazu full_dept Department of Pattern Recognition department (cz) RO department RO full_dept Department of Pattern Recognition garant K fullinstit Ústav teorie informace a automatizace AV ČR, v. v. i.
source	url https://library.utia.cas.cz/separaty/2026/RO/haindl-0648373.pdf
cas_special	abstract (eng) This paper addresses pixel-level ink detection in unwrapped Herculaneum scroll imagery (the Vesuvius Challenge) using non-neural pattern-recognition and machine-learning methods combined with carefully selected textural features. Recovering ink from these volumetric X-ray/phase-contrast image stacks is crucial for making lost texts accessible to scholars-particularly in cases where neural methods may be impractical due to limited training data, interpretability requirements, or other constraints. Although deep-learning approaches currently dominate many recognition tasks, there is still large room for alternative strategies. This study explores whether a transparent, data-efficient, non-neural pipeline can still achieve meaningful ink-background separation and serve as a reproducible baseline for future comparisons. The data used in this study were obtained from the EduceLab-Scrolls dataset.\n\nIn this study, we explore how several classical methods can be combined to detect ink patterns effectively. We apply multiple texture descriptors together with unsupervised segmentation using K-means clustering to create homogeneous image regions. To improve class separability and stabilize covariance estimates, we use Principal Component Analysis (PCA) for dimensionality reduction. The resulting lower-dimensional features\n are then experimentally classified using pooled-covariance Bayesian classifier in couple with Linear Discriminant Analysis (LDA). When appropriate, simple spatial regularization is applied to smooth results and reduce speckle noise. Each method was chosen not only for its effectiveness but also for its ability to shed light on how individual features contribute to the overall classification.\n\nEvaluation was performed on a custom benchmark derived from segmented Vesuvius Challenge data, using pixel-wise comparisons between our binary predictions and the corresponding ground-truth masks as the primary performance metric. Although our results currently fall below reported neural-network baselines, the experiments demonstrate that non-neural methods can still be useful in ink recovery tasks with limited training data, while also revealing specific failure modes inherent to such approaches. These findings suggest that thoughtful integration of the discussed methods could potentially enhance existing ink-detection models, and several promising directions for further refinement are outlined. action ARLID cav_un_auth*0506965 name CAA : Computer Applications and Quantitative Methods in Archaeology 2026 (CAA 2026) /53./ dates 20260331 mrcbC20-s 20260404 place Vienna country AT RIV BD FORD0 20000 FORD1 20200 FORD2 20205 reportyear 2027 num_of_auth 2 presentation_type PR inst_support RVO:67985556 permalink https://hdl.handle.net/11104/0378169 mrcbC61 1 cooperation ARLID cav_un_auth*0401576 name ČVUT Fakulta informačních technologií institution ČVUT FIT country CZ confidential S article_num 125 arlyear 2026 mrcbU14 SCOPUS mrcbU24 PUBMED mrcbU34 WOS mrcbU63 cav_un_epca*0648796 CAA Proceedings : CAA 2026: It's all about People CAA International 2026 Vienna