Neural networks are "trained" to recognize fake paintings
ARTinvestment.RU   18 июля 2018

It is an old dream of scientists and neo-алхимиков: just push the button — art-анализатор blinked the lights and drew the conclusion: "fake" or "original". Progress is being made, but the critics on-прежнему laugh at the "robot-экспертом"

Now experts in complex and critical cases, work the "old way". First rasmotrenie the people doing the initial examination to decide deserves thing attention or everything is clear, even nothing to watch. Then do "technology": an x-ray, look in the infrared, carried out a chemical study of the pigments and binders, etc. This allows you to immediately "take" things, not much falling at the time of creation in the prospective period. But the correct Dating little. Then begins, perhaps, the most important stage — art research. The expert or experts who specialize in the work of the alleged author, looking for authentication on the basis of the artist's style, his manners, and techniques, check out the biographical matching (where he lived, worked, traveled, etc.). And it is these experts have-искусствоведов are trying to take away the bread, the developers of artificial intelligence systems — scientists and programmers.

"Well, what is there impossible-то?" — argue, in-видимому, they. Obviously every artist has his own style, characterized by the techniques of application stroke and of the stroke. If the "feed" algorithm-нейросети a certain number of reference images, it'll soon learn to recognize the hand of the artist scans. And come all happiness.

Such a neural network has already appeared. And pass field tests. One of them last year in testing at Rutgers University (an hour by train from new-Йорка) and the Dutch Atelier for Restoration & Research of Paintings. On campus at Rutgers, by the way, is the Museum of Zimmerly from the collection of Norton Dodge — the best collection of Soviet nonconformist art in the world. But this is irrelevant.

the Researchers "fed" recurrent neural network (RNN) about 300 drawings by Picasso, Matisse, Modigliani and other artists, and she learned to distinguish between 80, 000 individual types of stroke, as well as to recognize the shape of the lines, to distinguish the effect of depression and other technical components of the style of the artist. And the system started to cope with the recognition. Test invited artists and asked them to draw pictures, imitating the style of Picasso and other famous authors. Useless: the neural network quickly saw through the simulation.

However it is not omnipotent. Artificial intelligence can perform only pictures with sharp lines or painting, where there are pronounced individual brushstrokes. That is, with the French Impressionists, she was soon right, and with Aivazovsky — is unlikely.

What's next? Whether and how soon artificial intelligence to replace art historians-экспертов? Yet, with whomever I spoke, all react upon the ability of the robots-экспертов condescending. All-таки everything to do with creativity and sense of beauty, too hard is formalized mathematically. Understanding and feeling the power of art — is all too human thing. For example, Krasnopevtsev superpoddelki experts is determined not by the strokes (although the form in the signature also may issue), but rather bythe balance of the composition — its equilibrium. The machine is not under power. So in our age of experts of flesh and blood works accurately enough. What neural network? Believe that they can become an auxiliary tool in determining the authenticity. And still believe that manufacturers of fakes of the highest class also will take these technologies on Board. And will strive to create "seleuchi" is able to cheat and pass the machine test.

Vladimir Bogdanov, AI

Sources: technologyreview.com, hightech.fm