Researchers Automated Historical Pigment Identification
A machine learning model achieved high classification accuracy by processing multispectral imagery of art.
Updated on Oct. 5, 2026 in Artificial Intelligence
Live Poll
Do you trust artificial intelligence to accurately identify and help preserve historical artifacts?
Published in 2025, researchers developed a random forest machine learning tool capable of identifying 40 unique pigments in historical paintings. The model relies on image data captured under varying spectral illumination conditions.
Why it matters
Traditional analytical methods for pigment identification are often time-consuming or spatially restricted, while hyperspectral imaging requires complex instrumentation. This approach offers a more efficient alternative for non-invasive art analysis.
The random forest model achieved 99.3 percent classification accuracy using 4,000 image patches sized 32 x 32 pixels. These were derived from 600 raw images captured under visible-light, ultraviolet, and infrared illumination.
The details
The classification framework generates feature vectors using statistical descriptors derived from RGB channels across three distinct imaging types. It processes visible-light reflectography, ultraviolet false-colour imaging — a technique that reveals hidden details by highlighting how pigments absorb or reflect UV light — and infrared false-colour imaging, which penetrates surface layers to show underdrawings. The model uses this multi-spectral data to categorize pigments without requiring the complex instrumentation typical of hyperspectral imaging.
Timeline
2025: The study was published in the journal Applied Sciences.
The Tech Race
This research marks a shift toward lower-complexity diagnostic tools in art conservation, moving away from the heavy data processing requirements of hyperspectral imaging. By using standard imaging modalities, the model aims to solve the bottlenecks seen in traditional analytical methods.
This research provides conservators with a more accessible framework for characterizing materials without using destructive sampling. It enables more efficient documentation of historical works by reducing the need for specialized laboratory equipment.
The takeaway
The study demonstrates that statistical descriptors from standard spectral imaging can achieve high classification accuracy. Researchers and institutions should monitor whether these models will be scaled to cover the remaining 160 pigment classes not included in this study.
Further reading
For broader trends in machine learning applications, visit Artificial Intelligence.
Source note: This article includes information reported by European Coatings.
Live Poll
Do you trust artificial intelligence to accurately identify and help preserve historical artifacts?
