This thesis is part of the ANR project AAA – Augmented Artwork Analysis, which aims to develop tools for the assisted interpretation of artistic images. The project is also rooted in a broader reflection on the genealogy of forms in art history. 
The work fits within the development of a system for representing, querying, navigating, and visualizing images, with the goal of designing an interactive content-based image retrieval tool. This tool integrates deep neural networks to enable the exploration of large artistic corpora.

This thesis focuses on the development of an interactive system for searching artistic images based on visual content, using deep neural networks to explore large corpora. This thesis proposes a new approach to the study of artistic forms using computer vision tools. Recent advances in object detection and recognition can be used to automatically identify specific patterns, shapes or elements within large corpora of images. These automated analyses could then yield visual relationships that have hitherto been difficult to access by manual observation, paving the way for a renewal of methods in art history.
It is in this context that we became interested in the classification of works by genre, style and artist and the detection of objects in paintings. The main objective of this thesis is to develop a content-based image search tool for a corpus ranging from the 15th to the 20th centuries.

To tackle this problem, we have implemented two complementary approaches. The first is to use classification models to recognise the artistic genres of the paintings. This step guides the search by matching paintings belonging to the same genre while remaining visually close. 
The second approach is based on the detection of objects in the paintings and the geometric and topological relationships that link them. The aim is to analyse the work and bring together similar works. Particular attention has been paid to the faces in the paintings, and in particular to the direction of gaze of the figures depicted. 
These two methodological approaches make it possible to cross-reference the results obtained by the algorithms with hypotheses derived from art history, and pave the way for a large-scale visual exploration of artistic forms.