An algorithm has been developed that determines the author of a painting by the characteristics of the strokes in it, as well as is able to distinguish between real paintings and fakes painted by other artists. The developers have trained the program on a set of nearly three hundred paintings by famous artists, such as Picasso and Matisse, according to MIT Technology Review. The development of American and Dutch specialists will be presented at the AAAI conference on artificial intelligence in February 2018, a preprint of the article is published on arXiv.org.
Since paintings by famous artists, as a rule, exist in a single copy, prices for them can amount to tens and hundreds of millions of dollars. Because of this, some paintings are forged by malefactors, and this is not always noticeable even to people who are versed in painting. To protect against such counterfeiting, various methods are proposed, for example, equipping paintings with unique identifiers, which are almost impossible to counterfeit due to their complex microstructure.
Researchers from the United States and the Netherlands, led by Ahmed Elgammal of Artrendex and Rutgers University, have created an algorithm that can recognize the authors of a painting by the features of their strokes. In 2015, this group of researchers has already created an algorithm that can classify paintings by authors and even styles based on their individual characteristics, such as colors. In the new work, the researchers decided to focus on one component of the paintings - strokes.
Each stroke can be described by many characteristics, for example, shape, length, uniformity of thickness along the stroke, and other parameters. The researchers decided to extract these characteristics using computer algorithms. Initially, the paintings were divided into separate strokes using a special algorithm. As a dataset for the algorithms, the researchers used 297 paintings by famous artists such as Picasso and Matisse, executed in the style of lithography, ink drawing, and others. The algorithm has broken down these pictures into more than 80,000 individual strokes.
Data set for training and testing algorithms / Elgammal et al. / arXiv.org, 2017
To assess the strokes, the researchers decided to use two approaches. They described basic characteristics such as stroke thickness and longitudinal profile using different descriptors and taught a support vector algorithm to classify strokes. The second approach was to use a recurrent neural network with controlled recurrent blocks, which independently searched for features characteristic of certain artists.
An example of fake paintings. Top row: fake; fake; original by Matisse. Middle row: original Matisse; fake; fake; original by Matisse. Bottom row: fake; original by Matisse; original by Picasso; fake / Elgammal et al. / arXiv.org, 2017
After preparing the algorithms, the researchers tested them on the same dataset, and by combining both approaches, they achieved 80 percent artist recognition accuracy. They also asked five artists to paint copies of paintings by Picasso, Matisse and Schiele. Having received 83 paintings, they checked them using their algorithms, and found that their combination is capable of recognizing a fake in all of these paintings.
Promotional video:
In the past few years, strong progress has been made in image processing and analysis using neural network algorithms. For example, such algorithms are able to mix several artistic styles in one image, turn sketches into full-fledged paintings, and even create original works of art. Also, similar algorithms work well with video recordings. For example, a system was recently presented that allows you to insert third-party speech into the video sequence, almost accurately recreating the articulatory facial expressions of the speaker.
Grigory Kopiev