Recognizing a painter's style is a continuously evolving research area in the field of machine learning and artificial intelligence. The approaches to this are manifold. They range from the extraction of certain features from digital images of works of art, which can serve as input for a wide variety of classification algorithms, to what are known as convolutional neural networks (CNNs), whereby the latter only use the images or image sections as input.

 

Originally, as part of a research project supported by the LKA Berlin (and before the company was founded), Reuter initially gained experience with features extracted from works of art. In particular, statistical parameters such as the average red, blue and green values, their standard deviation and other measured variables such as contrast gradients were used here. For several years, the start-up has been using CNNs exclusively for its analyzes - simply because the results of the models are much better than with extracted features, whatever type they are.

​​

Art Intelligence GmbH achieves a recognition accuracy of 90 percent in the pure recognition of painting styles with CNNs, which were trained with 50 painters from all style epochs. This is significantly better than the so-called Human Level Performance, i.e. the recognition accuracy of people interested in art, as was shown in a playful competition in 2018 in the Buchheim Museum ("Museum of Fantasy") in Bernried on Lake Starnberg.

​

Intuitively, the way a CNN works can be described as follows: The neural network consists of a good 23 million parameters, which are continuously optimized during so-called training to recognize the different painting styles.

​

To explain it simply, these networks “learn” characteristics that make up a painter's style. However, these characteristics are abstract – and are not presently understood by humans. However, with certain methods and techniques it can at least be visualized at which points in a picture the algorithms recognize the style of a painter.  

​​

The following two academic papers - representative of dozens of other papers on the subject - give a good overview of how CNNs work in recognizing artists.  

​​

http://cs231n.stanford.edu/reports/2017/pdfs/406.pdf

​

http://cs231n.stanford.edu/reports/2017/pdfs/410.pdf

​

The technique has already been used by various museums to assign works from inventories where pictures have not been handed down to specific artists.

 

For the detection of forgeries, Art Intelligence GmbH uses the CNNs trained in painter recognition, but also uses a slightly modified evaluation method that also considers the accuracy of detection of secured forgeries or imitations of the artist in question.

 

Other application examples, such as identifying a person portrayed in a painting, also use CNNs as a basis. Here, however, it is not the style of the painter in question that is recognized, but certain biometric points on the face of the person portrayed. A prerequisite for this type of analysis is that there are other illustrations of the suspected person, be they photos or other painted representations.