Discerning the painter's hand: machine learning on surface topography
It addresses attribution challenges in art history, particularly for workshop practices, but is incremental as it applies existing CNNs to new surface data.
This study tackled the problem of attributing paintings to artists by using machine learning on surface topography data, achieving 60-96% accuracy across patch sizes and nearly doubling accuracy compared to color image-based methods.
Attribution of paintings is a critical problem in art history. This study extends machine learning analysis to surface topography of painted works. A controlled study of positive attribution was designed with paintings produced by a class of art students. The paintings were scanned using a confocal optical profilometer to produce surface data. The surface data were divided into virtual patches and used to train an ensemble of convolutional neural networks (CNNs) for attribution. Over a range of patch sizes from 0.5 to 60 mm, the resulting attribution was found to be 60 to 96% accurate, and, when comparing regions of different color, was nearly twice as accurate as CNNs using color images of the paintings. Remarkably, short length scales, as small as twice a bristle diameter, were the key to reliably distinguishing among artists. These results show promise for real-world attribution, particularly in the case of workshop practice.