Nicolas Poussin: “Armide and Renaud”

Client image, Armide and Renaud, c. 1630

In 2022, Art Recognition was asked to authenticate the painting Armide and Renaud, which resembles a version displayed in the Gemäldegalerie in Berlin. The discovery of this painting considerably enriches our knowledge of Poussin’s early works, and allows a chronological readjustment to the early 1630s, eliminating the Berlin version dated in 1637.

According to the Berlin State Museums, Staatliche Museen zu Berlin, the painting’s attribution was disputed for more than a hundred years. After restoration and technical analyses, it was argued to be by Nicolas Poussin. Moreover, Armide and Renaud was produced in 1637 according to the historian Félibien. It has been reproduced subsequently into an engraving by Guillaume Chasteau.

The painting brought by our client represents an important addition to the knowledge we have so far of Poussin’s early work. As mentioned by Christopher Wright in his expertise, it is possible to see the differences between the client’s image and the work exhibited in Berlin, in which Poussin introduced many small changes and two important ones. The first is the introduction of a small column in the right background that replaces a small branch. The second is the removal of three tree trunks. This makes the whole composition lighter and much more open. The painting Armide and Renaud is part of the revision of the Catalogue Raisonné of Christopher Wright published by the Chaucer Press in London.

Art Recognition system at work …

Our proprietary deep convolutional neural network has first learned the artist’s discriminative features from the training data. This learning process is independent and there is no interference in any way.

We started by gathering 204 authentic images by Nicolas Poussin to train, optimize and validate our model. The system was also fed with 204 images that were not from Poussin to serve as a contrast set. Since the paintings exhibit high variability in terms of aspect ratios, we have employed a particular preprocessing strategy to capture both fine details and coarse structures. After having split images into non-overlapping patches, the final dataset contained 2604 data. Once the robustness of the trained neural network was tested on a validation set, we passed the image of the painting Armide and Renaud through the AI. It compared the features learned from the training images with those on the image in question. Based on this comparison, the AI returned a class probability of 78.90% for a positive response (‘authentic’).

The following are visualizations from our AI training. On the heatmap, the most important areas for the algorithm’s decision are those highlighted in red. These hotspots tend to appear in the regions which comprise more structure or are important for the overall composition.

Regarding the brushstrokes, we noticed that the visible strokes appear as some texture areas distinguishing themselves from other significant contours.

To conclude, the client’s painting provides a considerable source of information on Nicolas Poussin’s early works. It indicates that the composition originated around 1630 and not 1637 as suggested by the Berlin version. Thanks to this painting, it was possible to compare the differences and similarities with the version in the Gemäldegalerie in Berlin. Art Recognition’s AI system assessed the client’s painting as authentic with a 78.90% probability in favor of Nicolas Poussin.