Picasso, Matisse, or a Fake? A.I. for Attribution and Authentication of Art at the Stroke Level
Can artists be identified based on their individual strokes? Do individual strokes in a painting or a drawing carry artist’s unintentional signature, which would be hard to imitate of forge. Can these unintentional characteristics of strokes be quantified? These are essential questions in today’s Art market where one painting was just sold for an all time record-breaking price of $450 millions.
At Artrendex (Art Trend Analytics), in collaboration with the Atelier for Restoration & Research of Paintings, we developed an A.I. system that analyzes individual strokes in drawings. The goal is to test the hypothesis that artists can be identified based on individual strokes, which remains largely untested before scientifically. The goal also is to build a robust AI system that can help in attribution and authentication of art works, mainly based on the characteristics of strokes, and complementing other existing technologies.
The methodology we used is inspired by the “Pictology” methodology developed by Maurits Michel van Dantzig (1903–1960). Van Dantzig suggested several characteristics to distinguish the strokes of an artist, and suggested that such characteristics capture the spontaneity of how original art is being created, in contrast to the inhibitory nature of imitated art.
Among the characteristics suggested by van Dantzig to distinguish the strokes of an artist are the shape, tone, relative length of the beginning, middle and end of each stroke, direction, pressure, and several others. The list of characteristics suggested by van Danzig is comprehensive and includes, in some cases, over one hundred aspects that are designed for inspection by the human eye. The main motivation is to characterize spontaneous strokes characterizing a certain artist from inhibited strokes, which are copied from original strokes to imitate the artist style.
In our system we did not implement the exact list of characteristics suggested by van Dantzig; instead we developed methods for quantification of strokes that are inspired by his methodology but relevant to the digital domain and suitable for statistical analysis by machines rather than by human eye.
Our system isolates individual strokes using an algorithm for stroke segmentation (see Figure). The system then quantifies their shape and tone using various computer vision methods. The system also uses a deep neural network to track individual strokes and quantify their characteristics. In particular we used a recurrent neural network (RNN) architecture called Gated Recurrent Unit (GRU), which is a variant of LSTM (Long Short-Tem Memory). Such networks have been widely used recently in text translation and speech recognition. The characteristics of strokes, as quantified by the network and other methods, are compared to a large number of strokes by different artists using statistical inference and machine learning techniques.
We excluded using comparisons based on compositional and subject-matter-related patterns and elements. Most forged art works are based on copying certain compositional and subject-matter-related elements and patterns. Using such elements might obviously and mistakenly connect a test subject work to figures and composition in an artist known works. In contrast, if the hypothesis is true, the characteristics of individual strokes carry the artist’s unintentional signature, which is hard to imitate or forge, even if the forger intends to do so.
A collection of about 300 drawings were gathered from different sources to train, optimize, validate, and test the various classification methodologies used in this study. The collection included drawings and prints mainly by Pablo Picasso, Henry Matisse, Egon Schiele, Amedeo Modigliani, besides a small representative works of anther twelve artists, ranging from 1910–1950AD. These artists were chosen since they were prolific in producing line drawings during the first half of the Twentieth century. The collection included a variety of techniques including: pen and ink, pencil, crayon, and graphite drawings as well as etching and lithograph prints. Overall our collection contained around 80,000 strokes.
We performed extensive experiments with different settings. The settings included comparing the performance of models trained on specific drawing techniques (for example ink drawings or pencil drawings) and models trained across techniques to evaluate whether we can capture an invariant for the artist regardless of the technique used.
Our experiments show that the system can identify Picasso’s individual strokes across techniques with 79% accuracy, Matisse with 77% accuracy, and Schiele with 86%. This is classification based on one stroke at a time! By aggregating these results over a whole drawing, the system can identify Picasso’s drawings with 83% accuracy, Matisse with 80%, and Schiele with 83%. Similar results are obtained for the technique-specific models.
In order to validate the robustness of the system against being deceived by forged art, we commissioned five artists to make drawings similar to those of Picasso, Matisse, and Schiele, using the same techniques (see figure below for samples of these drawings mixed up with real ones). None of these fake drawings was used in training the models. We only used them for testing. The system successfully flagged these forged Picasso, Matisse and Schiele’s drawing 100% of the time using the across-technique trained models.
Attribution and Authentication of art works is a very essential task for art experts. Traditionally, stylistic analysis by expert human eye has been a main way to judge the authenticity of artworks. Various technical analysis methods are used today to analyze the surface of the painting, the pigments, the underpainting, and/or the canvas material. There is a wide spectrum of imaging (e.g. infrared spectroscopy and x-ray), chemical analysis (e.g. Chromatography), and radiometric (e.g. carbon dating) techniques that have been developed and used for this purpose. These techniques are complementary and each of them has limitations to the scope of their applicability.
The technology we are developing provides a quantifiable scientific way to approach the traditional stylistic analysis typically done by human experts, at the visual spectrum level, without the need for sophisticated imaging techniques. It would complement other existing technical analysis techniques and would provide a cost-effective solution compared to the cost of other lab-based methods that can be prohibitive at the massive scale to authenticate art not necessarily sold at the $450 Million level.
Key for real/fake images in the figure:
Fake, Fake, Matisse
Matisse, Fake, Fake, Matisse
Fake, Matisse, Picasso, Fake
Fake, Picasso, Picasso, Fake
Schiele, Fake, Fake, Schiele, Schiele, Fake
The work is published in a paper titled “Picasso, Matisse, or a Fake? Automated Analysis of Drawings at the Stroke Level for Attribution and Authentication ”, which will appear in the 32nd AAAI conference on Artificial Intelligence, to be held in New Orleans in February 2018. The paper can be accessed at https://arxiv.org/pdf/1711.03536.pdf
