The painting “The Wijk Mill”, recently discovered and attributed to Vincent van Gogh, has been checked for authenticity using Artificial Intelligence and Machine Learning. The Munich-based AI and Data Science specialists at Alexander Thamm GmbH did not detect any anomalies that would indicate a forgery. Their underlying neural network model is capable of recognizing a total of seven out of eight proven Van Gogh forgeries which were circulated by art dealer Otto Wacker in the late 1920s and early 1930s. The alleged Van Gogh painting “The Protestant Barn,” which surfaced in Italy in 2010, has also been classified as a forgery by the algorithm, an assessment which is backed up by numerous experts.
The model, which has been developed and periodically refined by the company’s Leading Data Scientist Wolfgang Reuter since the beginning of 2017, is generally able to detect forgeries with a relatively high degree of accuracy. The Berlin State Criminal Police Office has been supporting the project for several years – and has provided the AI expert with a total of 75 confirmed forgeries from six different artists during this time.
On average, the algorithm recognizes three-quarters of these forgeries, although the level of precision varies from painter to painter. For example, the model recognizes 67% of the 34 Penck forgeries provided by the authorities, but for supposed Max Beckmann paintings the rate of accuracy rises to 100 percent. For Heinrich Campendonk, the recognition rate is 90 percent, while only 80 percent of Max Liebermann forgeries were recognized
“It is worth noting, however,” says Reuter, “that the method used also always results in a certain proportion of false positives, i.e. originals being falsely classified as suspicious or conspicuous.” On average, this proportion is eleven percent. The developer is quite open about the fact that even Artificial Intelligence can never deliver a safe or court-proof verdict on whether a painting was actually painted by a particular artist or not. However, the same fallibility also applies to all other methods of verifying authenticity.
Indeed, human art experts and specialists have allowed themselves to be deceived in the past, as the case of the “forger of the century” Wolfgang Beltracchi, who was sentenced to six years in prison in 2011, clearly illustrates. Provenance research can also only provide clues, particularly if there are significant holes in the documentation of a painting in the years since its creation. Even chemical analyses are only of limited value, especially if the paint itself gives no immediate cause for suspicion. After all, many forgers are careful to only use pigments that were commercially available at the time of a painting’s supposed creation. In addition, there have been cases where newer colors were applied to a work of art during the course of restoration. Reuter therefore sees his algorithm as an additional indicator of authenticity, which does not make any of the previously used methods obsolete, but complements and supports them nonetheless. In the case of van Gogh, the precision in detecting counterfeits is 89 percent – however, just under eleven percent of originals also test positive (see graph below).
Each blue dot below the green threshold represents an original van Gogh painting that the model would classify as “suspect”. The red dots below the line represent detected forgeries. The somewhat obscured red dot at the probability value of 0.81 represents the “Mill of Wijk”, the even more obscured dot above it at 0.87 marks the only one of eight forgeries from art collector Otto Wacker that was not accurately recognized as such.
The Wacker forgeries were made available to Alexander Thamm by Henry Keazor, Professor of Modern and Contemporary Art History at the University of Heidelberg. They are shown below.
Paintings by the art dealer Otto Wacker recognized as forgeries:
Not recognized as a forgery painting of the art dealer Otto Wacker:
“Mill of Wijk” and Class Activation Map of the image:
The brightly marked areas show parts of the picture that the so-called convolutional neural networks have identified as distinctly ” typical for van Gogh”. However, this should be regarded as an ex-post observation, i.e. as a subsequent analysis of the model’s mode of operation: The algorithm receives as input just 100 randomly selected and unchanged image sections, each measuring 224 x 224 pixels, from each work of art. The algorithm learns by itself which areas or sections of an image are or are not “characteristic” for a painter during training.
Neither Alexander Thamm GmbH nor Wolfgang Reuter have any contact or connection with the current owner of the painting, the auction house or anyone else involved in the planned auction. The analysis was done based on their own initiative and out of purely scientific motives, without any commercial interest.
If you have any questions, please feel free to contact Wolfang Reuter by phone (+49 151/ 14659820) or e-mail (firstname.lastname@example.org).
Learn more about how the algorithm works in our whitepaper “Image Recognition with AI“.