Transfer Learning for Artwork Attribution: Assessing the Importance of the Artist's Signature

In this paper, we propose a Transfer Learning approach for artwork attribution investigating the importance of the artist's signature. We consider AlexNet Convolutional Neural Network and different fully connected layers after fc7. Three architectures are compared for authenticating paintings from Monet and classifying other artworks into the 'Non- Monet' class. We considered paintings with a similar style and artist's signature to Monet's, as well as paintings with a very different style and signature. We assess the model's performance on images with the artist's signature extracted from the artwork, as well as on images of its near background not containing the artist's signature. Results demonstrate the importance of considering the artist's signature for art attribution, reaching a classification rate of 85.6% against 82% on images without signature. The analysis of the obtained feature maps allows highlighting the power of our approach based on Transfer Learning in extracting high-level features that simultaneously capture information from the artist's signature and the artist's style.