Computer analysis identifies many similarities between van Gogh and Pollock

The paintings of Van Gogh and Jackson Pollock may seem totally unrelated, but according to a computer scientist at Lawrence Technological University, the artistic techniques of the two famous artists have a lot in common.

Using a software program he developed for biological image analysis, Lawrence Tech Assistant Professor Lior Shamir has compared works of art by nine painters - Gustav Klimt, Marc Chagall, Paul Cezanne, Joan Miro, Claude Monet, Jackson Pollock, Paul Klee, Piet Mondrian and Van Gogh.

The surprising conclusion was that the two painters in that group with the most in common were van Gogh and Pollock.

Van Gogh was a pre-1900 Dutch post-impressionist painter known for the vivid colors and the emotional impact of his paintings. Pollock was a leading proponent of abstract expressionism in the 1950s who abandoned the brush and easel used by van Gogh and would typically fling paint from many angles and directions onto a canvas attached to the floor.

Shamir's initial goal was to test whether computers employing artificial intelligence can "understand" visual art. He wanted to test the hypothesis that computer programs used for image analysis could help art historians find links between different artistic movements and classify paintings by artist or artistic style.

In the process, he has demonstrated that it is possible for computers to quantify attributes of art that can go undetected by even the most knowledgeable art critic.

"Recent advances in computer vision and image processing have enabled basic automatic analysis of visual art. In this study we have used computer analysis to extract thousands of numerical low-level image content descriptors from digitized paintings, and use them to objectively compare the similarities between the artistic styles of different painters," Shamir said.

He was the lead developer of Wndchrm, a 2008 open source utility for biological image analysis. He used it to break down works of art into 4,027 numerical descriptors.These include high-contrast features such as edge and shape statistics, textures, statistical distribution of the pixel values, polynomial decomposition of the image, and similar geometric shapes known as fractals.

By applying statistical methods, the computer program performed an automatic search for those numerical image content descriptors that are best able to differentiate between the artistic styles of the painter, and ranked these features by their discriminating power.

Among the 20 most informative image content descriptors that distinguished Pollock from the other painters, 19 showed that his work was most similar to van Gogh's. The same was true for 57 out of the top 80 most effective content descriptors.

These results indicate that the two artists from very different schools of art used similar artistic techniques as measured by the mathematical lower-level descriptors of their works, according to Shamir.

"The research does not necessarily prove that Pollock was influenced by Van Gogh, but it does show they happened to use similar painting methods, at least as measured by mathematical low-level features," Shamir said. "This computer analysis goes beyond what the unaided eye can sense and quantify, or what the human brain can perceive."

Artificial intelligence may have an advantage because there isn't a single place in the human brain that analyzes art, according to Shamir. "The analysis of art is a highly complex cognitive task that is not yet fully understood, as the different elements of visual art such as colors, shapes, and boundaries are processed by different pathways and systems in the brain," Shamir said.

Shamir published his findings in Leonardo, an international journal devoted to the application of contemporary science and technology to the arts. This research was the subject of an article in July 30 issue of The Economist.

After earning his bachelor's and master's degrees in Israel, Shamir was awarded his PhD in computational science and engineering in 2006 from Michigan Technological University. His dissertation was on astronomical imaging processing using fuzzy logic. He did postdoctoral research on biomedical image analysis for the National Institutes of Health.