In the past few years, the quick success of Convolutional Neural Networks in detecting and classifying images led to their adoption in a wide range of domains. This includes the automated analysis of paintings, where new technologies are expected to assist art historians and help maintain digital collections. Our work in this thesis proposes a novel approach for detecting similarities and influences across paintings based on information used for their classification by CNN. First, we adapt three previously trained models to predict artistic genres, styles and provoked emotions, showing improvements in their results when compared to some state of the art solutions. Then we combine the models’ internal representations of paintings to build a graph of similarities between artists across a dataset of 100.000 artworks. The relations found using this approach are largely corroborated by available documentation on well-known influences in art history and in some cases present valuable new artistic insights.
|la date de réponse||19 janv. 2021|
|Superviseur||Elio Tuci (Promoteur)|