We do know that Artificial Intelligence (AI) can be an asset in the health field: for example, there have been studies to see if AI, trained on databases of mammograms collected from clinics and academic institutions, may help detecting breast cancer in early stages.
A research published today on the journal Heritage Science (Ugail, H., Stork, D.G., Edwards, H. et al. Deep transfer learning for visual analysis and attribution of paintings by Raphael. Herit Sci 11, 268 (2023)) analyses instead another application for AI in the field of artwork authentication.
We all know that art attribution and authentication pose intricate challenges for experts as the processes to analyse an artwork and attribute it to a specific artist are long and involve studies on provenance, materials, iconography and conditions, not to mention analysis of artists' careers, derivative works, and connoisseurship.
The paper focuses solely on connoisseurship (i.e. visual analysis of the composition, style, brush strokes, shading, etc.) and introduces a computational tool aimed at assisting scholars in analyzing and authenticating a restricted class of paintings, focusing, as suggested in the title of the study, on artworks by Italian painter and architect Raffaello Sanzio da Urbino, better known as Raphael.
Machine learning in art analysis is seen with suspicion among the scholarly art community, mainly because of the scarcity of high-quality data necessary for training robust multi-class classification models. Yet the research team, led by Hassan Ugail, professor of visual computing at the University of Bradford, found a way to avoid this by constructing smaller, artist-specific models for more accurate analysis and authentication of individual paintings.
For their study on Raphael, the team mainly employed a pre-trained ResNet50 (Residual Network50) deep neural network for detailed feature extraction, combining it with an SVM (Support Vector Machine) binary classifier for authentication. Additionally, edge detection and analysis algorithms, crucial for capturing Raphael's artistic style, including brushwork signatures, were integrated as authentication tools.
This combined and hybrid approach enhanced the model's accuracy: to classify a painting as likely by Raphael, the artwork had indeed to meet both the edge feature threshold and the SVM model's probability threshold. This composite methodology should significantly improve classification accuracy, something that was put in doubt in an analysis carried out a few months ago. In that case, two AI research groups, one from the University of Bradford/the University of Nottingham and one from a private company named Art Recognition, aimed to authenticate the de Brécy Tondo painting. Using facial recognition and brushstroke analysis, they yielded conflicting results: while the former group claimed with absolute certainty that the work is by Raphael, the latter reported an 85% certainty that it is not. The outcomes challenged the perceived infallibility of AI, while art experts criticised the results highlighting that the painting was a Victorian copy of Raphael's the Sistine Madonna.
The essay published today highlights how the system was trained with a collection of authenticated Raphael paintings and a mixed array of paintings from artists like Rembrandt, Peter Lely, van Dyck (labelled "Not Raphael"). The first results looked promising as the machine learning approach achieved a 98% accuracy in image-based classification tasks during validation with a test set of well-known and authentic Raphael paintings.
Artificial Intelligence correctly identified "The Sistine Madonna", the "Haddo Madonna" (initially attributed to Innocenzo Francucci da Imola but suspected, and later confirmed, to be a Raphael) and "Lo Sposalizio or The Marriage of the Virgin" as painted by Raphael, while the portrait of Diana Kirke, Late Countess of Oxford as painted by another artist (Peter Lely).
The most extensive research was carried out on the Madonna of the Rose (Madonna della rosa, 1518-1520, Museo del Prado, Madrid). The team's algorithm suggested a probability of just 0.57 that the entire painting was the work of Raphael. Individual sections of the artwork (mainly the heads of the main figures in the painting) were then analysed, leading to doubts and questions about whether Raphael painted the face of Joseph in this artwork.
It is worth remembering that some art historians previously questioned the full attribution of this painting to Raphael alone, suggesting that his associate, Giulio Romano, might have had a hand in it. The machine learning analysis done with the help of AI supports this academic theory.
The methodology developed in this research can be readily modified and adapted for authenticating paintings by different artists, benefiting art historians and collectors.
As machine learning and image processing technologies advance, this method could therefore become part of a comprehensive toolkit for artwork analysis, working alongside existing methods such as scholarly analysis and advanced imaging techniques like spectroscopic imaging, and dating techniques.
Yet the results confirm the obvious, reassuring art experts that their jobs will not be replaced any time soon: just like AI can be an asset for a radiologist or a doctor, providing faster and more accurate diagnostic analyses, the pre-trained ResNet50 and the SVM classifier may lead to interesting results when used judiciously by art experts, historians and collectors familiar with the context, as these tools can enhance traditional visual analysis in art authentication challenges and accelerate research processes, operating in conjunction with other methods of analysis.
So far authentication processes employing AI were only applied to the art field, but there is another discipline in which authentication has become crucial - fashion. Yet, authenticating garments and accessories may be trickier for Artificial Intelligence as authenticity in these cases also passes through a close-up tactile analysis. That said, maybe in future we will also have AI tools that may help fashion historians in their work.
In the meantime, those who want to discover more about the technical aspects of the research published today can access the image datasets used to train and test the machine learning model and check out the code behind it at this link.
Comments
You can follow this conversation by subscribing to the comment feed for this post.