Slashdot Mirror
Software To Authenticate Paintings
eldavojohn writes, "There's a new software tool out and about called Authentic which analyzes paintings to determine if they are indeed authentic works of the artist. If you don't think this is a serious problem to tackle, some experts estimate up to 15 percent of 'original prints' sold at auction houses are actually fake. From the article: 'By dividing 145 digitized paintings into pixels and analyzing the colors of each and how they compared with nearby pixels, the system was able to spot patterns unique to the painter. The software also showed Van Gogh's use of complementary colors (PDF) increased during his most active period from 1885 to 1890, according to the study published in Pattern Recognition Letters... In tests, Authentic performed as well as 15 human volunteers who were each given a small segment of a painting to study.' I've heard of many tools that analyze texts to verify the author but this is an extra dimension and a new frontier for pattern recognition. Tacking on another dimension, how much longer until we are able to analyze video in the same way?"
What amazes me is how many tools are out and available online regarding this sort of pattern recognition development. Since a lot of people know Java, I'm would encourage you to use the Java Media Framework (free from Sun). Once you have those libraries installed, it's quite easy to start editing sound, images & video. You might need to grab and install codecs if you're doing video analysis but I think almost all image codecs are supported.
I'm not going to lie, the video computation can be quite heavily but thankfully that framework is implemented such that the entire video doesn't have to be loaded into memory, just a one frame buffer analysis can be used if you want.
The last thing you would need is simply the know-how on programming these analysis algorithms. There are sites out there with a large wealth of up-to-date algorithms. An example would be the text book style site of pattern recognition or image processing. While this doesn't teach you how to do things, it does contain the raw resources and algorithms. General resources like the computer vision homepage exist that serve as links to all kinds of resources. Unfortunately, I know of no real solid books that contain everything out there because this field is so rapidly developing. My professors taught me from hand printed slides in a large compendium they had accumulated over the last couple years.
The last piece missing is the data to analyze. While you might not have the ultra high resolution Van Gogh images to do this yourself, it may be possible to visit museums with 6 MP cameras to obtain your own data. Failing that, there are repositories online that sometimes contain image information you can start with. While this may not satisfy your specific needs, it sure is great for the lazy developer like myself.
Lastly, I will mention citeseer and Google Scholar for cutting edge papers that you might want to try implementing. Distributing these algorithms and building a good GUI can be tricky but really anyone can build the backend. I heavily recommend experimenting with this if it interests you.
My work here is dung.
Sorry, this software does not detect fakes, as claimed. All it can possibly do is detect whether or not a painting resembles other paintings by an individual artist. Speaking as a painter myself, I know that most artists undergo radical changes through their career, and painting styles may change radically due to such simple factors as buying a different brand of oil paint. Some artists never repeat the same style twice. Some artists create works in a unique style and then abandon that style after only a few works. Some artists emulate the style of their teachers so closely that even experts can't tell their works apart. Software is not likely to help these situations.
And to further complicate the problem, the biggest problem in the art market is not forged oil paintings, it is forged prints. I know one famous atelier that keeps the plates from famous artists works they've printed (they are supposed to be destroyed at the end of a printmaking edition) and once in a while they'll reprint a few, forge the artist's signature, and sell them under the table as unnumbered Artist's Proofs. These forgeries sell for tens of thousands of dollars, and are undetectable from image analysis, they are printed from the same plates as the originals and are 100% identical. But they're fakes by any standard, since they were not authorized by the artist and are not numbered.
Conventional analysis is more than sufficient to deal with fakes. Chemical analysis of pigments or grounds, and IR, UV, or XRay imaging, etc. are well developed techniques for identification of forgeries. I know of some Matisse fakes that were identified because an art historian looked at the thread count in the canvas and determined this type of machine-woven canvas was not manufactured until after Matisse's death. You can't teach this to a computer, it requires experience and long study.