Popularly, art connoisseurs are portrayed as sophisticados who carry themselves with an aura of mystery, in command of an inner portal to truth that the rest of us inexplicably just don’t possess. Presented with an unassuming Renaissance painting purchased for $1,000 in New Orleans, for instance, one might be stricken with certainty that the painting was authored by no other than Leonardo da Vinci; another attributes hundreds of paintings to Rembrandt and claims that his genius is obvious to the “experienced eye.” The elusive certainty of connoisseurship has always come with raised eyebrows: can you tell a garage sale replica from the real deal, let alone a workshop painting from an Old Master one? Can we trust anyone who claims to know?
Recent developments in machine learning applied to photographs of artwork promise to lend more objectivity to processes of attribution when the provenance is uncertain. In September, AI analysis performed by Swiss company Art Recognition determined that “Samson and Delilah” (c. 1609-10) — a painting attributed to Peter Paul Rubens, touted by London’s National Gallery as a highlight of its collection and originally bought at auction for a then-record price of £2.5 million in 1980 (~$11.5 million today, accounting for inflation) — concluded with 91 percent certainty that the painting had not been painted by Rubens.
Now, a new study published by researchers at Case Western University shows that machine learning analysis of the “surface topography” of a painting can be consistently accurate in identifying who has done the brushwork. In their testing of 720 patches from paintings by four artists, the algorithm attributed 96.1 percent to the correct painter.
The researchers hypothesized that brushwork on a painting leaves behind a “fingerprint” that largely lies beyond human powers of identification but can be picked up by computation — and they were right.
“We’ve uncovered what could be considered the unintentional style of a painter,” Kenneth Singer, a lead researcher on the study, said.
Previous research has already established the potential for applying machine learning to high-resolution photographs of paintings to assess their style and era of origin as well as whether they are forgeries. Instead of analyzing digital images of paintings, the recent study works with topographical data, scanning the surface of a painting to collect information about brush patterns and how the paint was deposited and dried. It therefore presents an additional metric that can either count for or against attributing a painting to an artist.
Four art students from the Cleveland Institute of Art were recruited to produce three similar paintings each of a water lily, all using the same supplies and tools. The researchers then broke down the paintings into small square patches of various sizes ranging from 0.5 to 60 millimeters, using some of them to train the algorithm. They compared regions with different colors, and found in that scenario that their attributions were almost twice as accurate as an algorithm using photographs of paintings. Another unique advantage of analyzing the surface of a painting is that art historians may be able to attribute different regions of a painting to different hands — particularly useful for understanding how workshop paintings are made. It could be consequential in the valuation of such paintings: a painting largely forged by a master, compared with one with very little intervention by the master, might see their price tags quickly diverge.
Such developments don’t eliminate the role of connoisseurs, but will likely change the methods they employ. Pure intuition may no longer suffice as the justification for an attribution, and perhaps the mystery will no longer lie in a connoisseur’s convictions but in what a neural network sees that we don’t.
0 Commentaires