A Blog by Jonathan Low


Feb 13, 2015

Computer Algorithms Can Accurately Identify Jackson Pollock Painting...And Maybe Lots More

Pattern recognition enhancement is one of the salient advantages big data offers to enterprises.

Whether it's identifying painting styles or recognizing sales trends, making sense of hard to decipher arrangements is an increasingly important benefit for a world drowning in information with limited capacity to interpret it.

In the world of art, where prices are becoming ever more stratospheric, recognizing forgeries is a serious financial obligation.

In the worlds of business, commerce and government, the process of optimizing everything from budgetary options, to national security strategies to consumers proclivities can have significant economic and social consequences. We are moving from not knowing what something is but knowing what we like to better understanding why we like it.JL

John Timmer reports in ars technica:

A series of image analysis algorithms can discriminate between real Pollocks and pieces painted in an attempt to mimic his style.
Like his work or hate it, it's clear that the painter Jackson Pollock pioneered a distinctive visual style with his drip paintings. If you admire his work (which I do), then you probably admire his distinctive visual flair. If you don't (heathen!), then you probably think that any old schmuck could produce an equivalent work simply by grabbing a paintbrush and having a seizure.
I am now fortunate enough to be able to argue for my perspective with math on my side. Lior Shamir, a computer scientist at Lawrence Technological University, has taken a series of image analysis algorithms and shown that they can discriminate between real Pollocks and pieces painted in an attempt to mimic his style.
This isn't the first time that computer science has intruded into the world of fine art. In several cases in the past, it has done so via the art community's invitation, as scientific analysis can help determine whether a suspicious work is likely to be genuine or not. Since Pollock painted very recently, however, it's relatively easy for a forger to match his materials. Whether they match his style or not is debatable given the disagreements noted above about whether there's a distinctive style involved at all.
There are, however, a number of possible Pollocks that are, as Shamir puts it, of "controversial authenticity," and computer scientists have attempted to use image analysis to weigh in on the issue. The idea that an artist paints in ways that create a distinctive style that's possible to identify based on multiple paintings isn't at all controversial. The question is whether that style can be reduced to a value through the mathematical processing of the information in an image of those paintings. In Pollock's case, past attempts indicated that his paintings have a distinctive fractal nature that can be recognized by algorithms.
Shamir, however, has developed a software package (originally intended for analysis of medical images) that can best be described as "throw everything at the wall and see what sticks." First, the software runs through a variety of processes that identify basic features of the image such as "textures, colors, edges, shapes, fractals, polynomial decomposition of the image, and statistical distribution of the pixel intensities." Next, it performs various transformations on the image (Fourier transforms, wavelet analysis, and so on), and computes values from those.
By comparing multiple images from a single source, it identifies the individual computations that provide the most consistent information about the source. Then, given a new image from an unknown source, it performs the same calculations and sees if the results are similar to its training set or not. If they're sufficiently similar, the work is likely to be in the same style as the artist's and probably by the artist him or herself.
For the Pollock paintings, Shamir obtained 26 paintings known to be by the artist and a second set that was painted by artists who were inspired by Pollock and attempted to create works in his style. The pieces were normalized to contain 640,000 pixels and then divided up into 16 equal-sized areas. Twenty works were used to train the software, then the remaining six were used for testing purposes. The analysis was then repeated multiple times, each with a different set of 20 training images, in order to provide a greater statistical power. The piece was determined to be Pollock or not based on a majority voting system, with each of the 16 sections of the painting getting one "vote."
Given these paintings as input, the computer was accurate over 90 percent of the time when asked to determine whether a painting was by Pollock or another artist. But previous computer analyses had suggested Pollock's style evolved over time, so Shamir went back and found 26 paintings that were done in the first half of the 1950s. With these as the training set, the accuracy reached 100 percent.
Looking back at the features that the algorithm ended up using, Shamir is able to identify a number of things beyond fractals that help define Pollock's style. (For the curious, these include Chebyshev statistics, a statistical analysis of the Fourier transform, Haralick textures, and Zernike polynomials taken from other image transforms.) It's not clear whether these are simply different ways of identifying a limited number of stylistic features or whether Pollock's style is multidimensional.
In any case, it's clear that there's something distinctive there. Whether you like it or not is a separate issue.
Shamir has placed the source code for this analysis package, termed "Wnd-charm," online.


Post a Comment