Making broad differentiations between modern and classic paintings can be fairly easy for the untrained eye, but telling the difference between an Impressionist and a Post-Impressionist painting may require a certain knowledge of art history. Well, it ain’t necessarily so when it comes to computers. An algorithm created and tested by computer scientists Lior Shamir and Jane Tarakhovsky, of Lawrence Technological University in Michigan has produced surprisingly accurate and expert results in art analysis.
The experiment was performed on approximately 1,000 paintings by celebrated artists. The technique is based on numerical image context descriptors, 4,027 of which were computed from each painting. These are numbers that identify the content of the image such as texture, color and shapes. Pattern recognition and statistical methods then analyze various aspects of the visual content of the paintings in order to detect patterns of similarities and dissimilarities between different painting styles. The feat is achieved without any human assistance in the process – or a degree in art history.
The algorithm succeeded in producing a network of similarities between painters that was largely consistent with the analysis that an art history expert would make. For example, the computer clearly identified the differences between classical realism and modernism. But it went further. Inside each of these two groups, it identified sub-groups that were part of the same movement. For example, it deduced that Gauguin and Cézanne should be clustered together, which is correct, since they are both Post-Impressionist painters. It also lumped together Raphael, Leonardo da Vinci and Michelangelo as representatives of the High Renaissance period.
A computer-generated graph of similarities between 34 different painters (Image: Lior Shamir)
In terms of application, the method can be used for discovery-driven research in art and history, Shamir told Gizmag. “Questions such as what artists or artistic movements inspired a certain painter can be addressed in a more quantitative fashion, and the methodology makes such assumptions testable,” he added.
The results of the experiment were presented in the recent issue of ACM Journal on Computing and Cultural Heritage.