Digitised art collections have outgrown classic metadata search, while the web is being flooded with machine-generated imagery. This work presents Warden Ship, an interactive application for exploring an art-historical collection with the help of artificial intelligence. The thesis has two parts: an interactive graphical interface, and a backend of machine-learning models. The application renders six equal features – semantic search, classification of style, genre and artist, creativity and influence charts, a map of art history, and visual-attribute filters – as scenes of a shared nautical world, connected by animated transitions. The frontend is built with React and React Three Fiber; the backend uses Flask with MongoDB and GridFS. At its core are a CLIP model for text-to-image search, a fine-tuned vision transformer that predicts style, genre and artist jointly, and four classical computer-vision channels for filtering by visual attributes. The models are evaluated on a held-out test set of 12,680 images: the fine-tuned ViT-L/336 variant reaches 77.0 % accuracy for style (27 classes), 86.3 % for genre and 97.3 % for artist, whereas zero-shot CLIP reaches only 25.8 %, 57.7 % and 56.0 % on the same classes. Three components are then adapted to the art domain or trained anew. Contrastively fine-tuned CLIP lifts style accuracy to 55.3 % and genre accuracy to 79.8 %, while an ArcFace metric embedding raises precision@1 for style from 41.6 % to 61.0 %. A new regression head over CLIP embeddings estimates the creation year of a work with a mean absolute error of 31 years, against a naive baseline of 73 years. Test accuracies closely track validation, confirming good generalisation, and the results exceed the targets that were set.
|
