A 2021 controlled study showed that convolutional neural networks can attribute paintings to individual artists from optical-profilometry surface scans alone, with the discriminative signal concentrated at sub-millimeter scales comparable to one or two brush-bristle diameters — but the evidence came from nine student painters using identical materials on fresh canvases. The open question, posed by that study's authors as the route to real-world use, is whether surface-topography attribution transfers to historical paintings: whether the sub-millimeter signatures survive aging, varnishing, soiling, craquelure, and past conservation; whether tool- and material-induced texture can be disentangled from artist-specific motor signatures; and whether the contributions of multiple hands within a single workshop painting can be identified, and with what supervision. A 2026 follow-up by the same group (PATCH) gave a first label-free demonstration on two El Greco paintings, including a disputed workshop work, but explicitly left material-versus-artist disentanglement, alteration and damage confounds, and validation beyond two case studies unresolved.

Background

Attribution — deciding who actually painted a given work — is a foundational problem in art history, traditionally settled by connoisseurship: expert judgment of an artist's characteristic style from visible evidence. The problem is hardest for Renaissance and Early Modern paintings, because masters such as Rubens ran workshops in which apprentices and assistants contributed to canvases, so a single painting can contain several hands whose contributions are difficult to separate, and written records of workshop practice are sparse. Ji et al. (Heritage Science 9:152, 2021) introduced a quantitative approach based on surface topography. Optical profilometry is a non-contact scanning technique that measures the height of a surface point by point; their confocal profilometer recorded the paint surface at 50-micron lateral resolution. Nine art students each painted three copies of the same subject using identical paints, canvases, and brushes, so that any distinguishing surface signal would come from the painter's habitual motor and tool-deposition behavior rather than from materials or composition. The height maps were divided into small square patches (side 0.5 to 60 mm) — patches small enough to exclude the depicted subject — and an ensemble of convolutional neural networks (CNNs, image classifiers here initialized by transfer learning, i.e. reuse of a network pretrained on ImageNet photographs) assigned each patch to its painter with 60–96% accuracy, exceeding 96% at the best patch size. Filtering the height data into intrinsic mode functions (IMFs, the spatial-frequency-band components produced by empirical mode decomposition, a sifting technique borrowed from time-series analysis) showed that the discriminative information lives at the smallest length scales, comparable to one or two brush-bristle diameters (roughly 0.2–0.4 mm), and topography beat color photographs when training and test paintings differed in subject and palette. The controlled design deliberately isolated the artist's signature, and the authors acknowledged that this isolation is exactly what real paintings lack: a historical surface carries aging of paint and ground, varnish layers, soiling, craquelure (the network of fine cracks that develops in aged paint films), and conservation treatments, and its regions may represent several workshop hands or later restorers.

Problem Statement

Determine whether attribution of paintings by machine learning on optical-profilometry surface topography — established only under controlled classroom conditions — transfers to real historical paintings. Concretely, with answers validated on actual historical works rather than only on controlled replicas: (i) do the discriminative sub-millimeter topographic signatures, at length scales comparable to one or two brush-bristle diameters, survive the surface alterations a historical painting accumulates — aging of paint and ground, varnish layers, soiling, craquelure, and past conservation treatments — and can preprocessing separate original brushwork topography from later interventions; (ii) how do attribution performance and the identity of the informative length scales change when brushes, paint formulations, primers, and supports differ between the works used for training and the work under analysis, i.e. can tool- and material-induced texture be disentangled from artist-specific motor signatures; (iii) can the contributions of multiple hands within a single workshop painting be identified, and what supervision — labeled paintings per hand, or label-free pairwise approaches — does that identification require, with what accuracy as the number of candidate hands grows; and (iv) how do these constraints interact with the tradeoff between patch size and available training data on single, unique canvases. These four strands share one objective: a validated, quantitative, per-region attribution method for historical workshop paintings grounded in measurable surface topography. In strand (i), the listed alterations describe the class of surface changes that a historical painting accumulates rather than a checklist of conditions that must each be tested separately.

References

  1. F. Ji, M. S. McMaster, S. Schwab, G. Singh, L. N. Smith, S. Adhikari, M. O'Dwyer, F. Sayed, A. Ingrisano, D. Yoder, E. S. Bolman, I. T. Martin, M. Hinczewski, K. D. Singer, 'Discerning the painter's hand: machine learning on surface topography', Heritage Science 9:152 (2021), DOI: 10.1186/s40494-021-00618-w, https://doi.org/10.1186/s40494-021-00618-w (also arXiv:2106.07134, https://arxiv.org/abs/2106.07134)
  2. Andrew Van Horn, Lauryn Smith, Mahamad Salah Mahmoud, Michael McMaster, Clara G. Pinchbeck, Ina T. Martin, Andrew Lininger, Anthony Ingrisano, Adam Lowe, Carlos Bayod, Elizabeth S. Bolman, Kenneth Singer, Michael Hinczewski, 'PATCH: A deep learning method to assess heterogeneity of artistic practice in historical paintings', Science Advances 12:eaea0485 (2026), DOI: 10.1126/sciadv.aea0485, https://doi.org/10.1126/sciadv.aea0485 (preprint arXiv:2502.01912, https://arxiv.org/abs/2502.01912)
  3. Eric Chen, Patricia Alves-Oliveira, 'Patch-Based Spatial Authorship Attribution in Human-Robot Collaborative Paintings', arXiv:2602.17030 (2026), https://arxiv.org/abs/2602.17030
  4. Clarissa Loures, Caio Hosken, Luan Oliveira, Gianlucca Zuin, Adriano Veloso, 'Enhancing Authorship Attribution with Synthetic Paintings', Proceedings of the IEEE International Conference on Machine Learning and Applications (ICMLA), pp. 283–288 (2025), DOI: 10.1109/ICMLA66185.2025.00044, https://doi.org/10.1109/ICMLA66185.2025.00044
  5. Iris Bautista-Morenilla, Júlia María Olivera López, Cristina Ruiz-Recasens, Oriol Moret Viñals, Lluís Casas Duocastella, Gema Campo-Francés, 'Non-invasive methods for characterising letterpress prints: Surface microscopy and profilometry', Journal of Cultural Heritage 78:89–99 (2026), DOI: 10.1016/j.culher.2026.01.013, https://doi.org/10.1016/j.culher.2026.01.013
  6. Soojung Kim, Ji-o Kim, Gilles Nogues, Kyujung Kim, 'OCT-based quantitative analysis of craquelure morphology in an oil painting', Journal of Cultural Heritage 74:332–340 (2025), DOI: 10.1016/j.culher.2025.06.017, https://doi.org/10.1016/j.culher.2025.06.017