Szczegóły publikacji
Opis bibliograficzny
Machine learning-assisted vibrational spectroscopy for textile fiber identification in historical tapestries / M. KRÓL, P. STOCH, S. WÓJCIK, E. Nowak Przybyszewska, W. MOZGAWA // Journal of Cultural Heritage ; ISSN 1296-2074. — 2025 — vol. 73, s. 571-578. — Bibliogr. s. 577-578, Abstr. — Publikacja dostępna online od: 2025-05-15
Autorzy (5)
- AGHKról Magdalena
- AGHStoch Paweł
- AGHWójcik Szymon
- Nowak Przybyszewska E.
- AGHMozgawa Włodzimierz
Słowa kluczowe
Dane bibliometryczne
| ID BaDAP | 159907 |
|---|---|
| Data dodania do BaDAP | 2025-06-17 |
| Tekst źródłowy | URL |
| DOI | 10.1016/j.culher.2025.04.034 |
| Rok publikacji | 2025 |
| Typ publikacji | artykuł w czasopiśmie |
| Otwarty dostęp |  |
| Czasopismo/seria | Journal of Cultural Heritage |
Abstract
Wawel tapestries constitute an invaluable part of Polish and world heritage. To preserve this patrimony and support conservators, it is essential to correlate the type of material from which they were made with the presence of dyes. We structurally characterized a large set of wool and silk fibers collected from tapestries during conservation work using Fourier transform infrared spectroscopy coupled with multivariate statistical analysis. This allowed us to correlate the protein structure with the type of the samples and the presence of dyes. We found that Principal Component Analysis (PCA) effectively distinguished silk from wool, revealing greater homogeneity in wool fibers. However, PCA showed limitations in accurately classifying fiber colors, prompting the application of supervised machine learning models. Among the tested approaches – Logistic Regression (LogReg), Linear Discriminant Analysis (LDA), k-Nearest Neighbors (kNN), and Multi-Layer Perceptron (MLP) – MLP demonstrated the highest accuracy, particularly in handling complex spectral patterns. Notably, systematic misclassifications aligned with spectral similarities in dye compositions, suggesting inherent challenges in fiber color differentiation. These findings highlight the potential of machine learning in historical fiber characterization.
