Description:
Current art databases are organized based on ontologies produced by art experts, and the art images are then labelled with directly relevant keywords represented in such ontologies. Thus, this organization ignores two things: 1) relations between images in the database that are indirectly connected through the art ontology, and 2) quantitative representation of image features. The second innovation is then the proposal of a novel probabilistic model that addresses the two issues above. The model works with a graph of image similarities based on art and text ontologies, and global, local and pose image features. We also have another graph that links database images to art ontology keywords. These two graphs can be used to discover new relations between images and keywords that are not explicitly represented in the art ontology, which means that the model shall propose a new internal organization of the art database.