When evaluating art, we have always relied on esteemed art critics and historians to tell us which are the most uniquely interesting pieces throughout history. We all know paintings that have withstood the test of time, bringing something unique to that time period, like Leonardo de Vinci's Mona Lisa, Claude Monet's Water Lilies, Pablo Picasso's Guernica, and Edvard Munch's The Scream.
But it's the job of the art historian to essentially be a walking library of all known facts about the artist, time periods and aesthetics. Art historians are also given the complicated task of classifying paintings by style, which requires the sharp eye to spot varying aesthetics.
But now this job is going to a machine.
Ahmed Elgammal and Babak Saleh from Rutgers University in New Jersey developed a machine-vision algorithm that can select the most creative paintings in history to provide a new way to explore art history and to see how creativity changes among the great artists.
Up until now, computer science and machines haven't been capable of identifying and analyzing paintings in terms of their style and creativity. Thanks to technological advances in machine learning, the two have been able to successfully train algorithms to accurately identify the artist and style of paintings, which has provided connections that art historians have struggled to make before.
"We investigate a comprehensive list of visual features and metric learning approaches to learn an optimized similarity measure between paintings," the researchers write.
Elgammal and Saleh created a database of over 62,000 photos of art paintings that includes more than 1,000 artists across 15 centuries, along with 27 different styles. The works of art are also categorized by genre.
The researchers then used machine-learning algorithms to pick out certain features like color or texture, using the database to study the art.
"We develop a machine that is able to make aesthetic-related semantic-level judgments, such as predicting a painting's style, genre, and artist, as well as providing similarity measures optimized based on the knowledge available in the domain of art historical interpretation," they write.
To test that the algorithm works, they used fine-art paintings the machine had not seen. They found that the technology was able to identify the artist correctly 60 percent of the time, and identify the style correctly 45 percent of the time.
The algorithm makes it possible to study art in a new way since it brings up links between different styles and influences (such as how expressionism can be linked to fauvism). Of course, it also provides more obvious links like which artists influenced others in different time periods, and the similarity and differences in fine art paintings. It will be useful to help archive and scan information from the number of fine-art collections that are digitized and publicly available. "With the availability of such large collections of digitized artworks comes the need to develop multimedia systems to archive and retrieve this pool of data," they note.
But the system isn't free from flaws. The researchers found that this algorithmic approach makes it hard to differentiate between similar artists who used similar styles in the same time period, such as Claude Monet and Camille Pissarro.
This experiment can help art historians evaluate and study art, but could also be applied to other areas such as literature or science.
Photo: Mike Licht | Flickr
Be sure to follow T-Lounge on Twitter and visit our Facebook page.
