Art forgery is big business, and detecting the forgeries is time-consuming and difficult. New research from scientists published in the newest issue of PNAS shows a computerized algorithm based on vision research and sparse coding models that has great promise.
In their experiments, the researchers used the works of Flemish artist Pieter Bruegel the Elder as guinea pigs. Bruegel's art often depicts pastoral scenes, an image type that sparse coding is very good at processing, and his work spawned many imitators in his time, making him an ideal subject. First, the researchers took several known authentic Bruegel works, rendered them in grayscale, and processed them with sparse coding in tiny patches of 8×8, 12×12, and 16×16 pixels in different trials, making them into maps of math functions.
Once this was done, they processed patches from other drawings onto the authentic work, and looked at the probability of whether the new image's functions mapped onto the original. If the average probability peak for each pixel comparison was very sharp, the second image was more likely to be authentic; if the peak was more spread out or less prominent, the second image was probably a fake.
It has limitations, however. The paintings used as reference need to be of similar subjects as the suspected forgery, and the algorithm is best at landscapes. However, it’s a interesting piece of research that shows promise if further developed.
