Attribution by Algorithm

Eric Postma is a Dutch artificial intelligence expert who has been working on ways to authenticate art works using algorithms derived from visual information. The International Herald Tribune asked him how prevalent fakes are and what his technology can do about it:

Within the museum world there are two classes of paintings: established works and doubted works. In the first class the majority will be authentic, with perhaps 1 percent forged. In the second class the number of fakes, or misattributions, is much greater — typically, between 50 percent and 80 percent. These figures have not changed much over the years. […]

An important clue in identifying the artist’s style is the configuration of the brush strokes. Our software is able to break down these brush strokes to find a complex pattern unique to every artist. That is just one of many algorithms we use. We also look at pigment and canvas weave.Working in cooperation with the Van Gogh Museum and the Kröller-Müller Museum, both located in the Netherlands, we have been able to demonstrate the accuracy of digital analysis. […]

We are currently analyzing the works of Rubens, Monet and Gauguin. Provided we have a large enough database of paintings to work from, I see no reason why we could not apply our methods to Old Masters and modern works of art alike.

I was recently asked if we could tell whether a “drip” painting by Jackson Pollock was authentic. Clearly there are no brush strokes to work from here, but Richard Taylor at the University of Oregon performed a fractal analysis of Pollock’s paintings using computer algorithms and succeeded in demonstrating how these algorithms could distinguish a true Pollock from a forgery.

A few artists present challenges. When we started our analysis in 2000, we had trouble authenticating the works of Rembrandt because his paintings are dark and the brush strokes difficult to identify. Technology has improved since then and the digital images that we work from are of a much higher quality.

Ultimately, we would like to come up with an algorithm that goes beyond brush strokes and captures the visual structure of a painting.