Les commentaires sont clos.

HdR Frédéric Precioso - "Apprentissage pour la classification et la recherche par le contenu dans les bases multimédia"

23 Novembre 2010

Catégorie : Soutenance HdR

Frédéric Precioso présentera son habilitation à diriger des recherches, intitulée "Apprentissage pour la classification et la recherche par le contenu dans les bases multimédia", le 3 décembre 2010 à l'Université de Cergy-Pontoise.


English title:

Learning for classification and content-based retrieval in multimedia databases.


The very challenging context of video content retrieval in large databases requires classification methods able to deal with highly semantic concepts and multi-class objectives while being compliant with user interaction (as on-line learning). When considering current state-of-the-art or even reference methods for video content retrieval, two key issues appear: the scalability becomes critical when dealing with video data; the temporal information is underused in data representation avoiding to handle the semantic complexity of video concepts.

In the first part, we focus on building new similarity measures scalable with respect to the size of the database, considering the context of retrieval when images are represented by a single feature histogram then proposing a scalable active learning strategy and finally designing a kernel-based similarity approach handling powerful data representation while preserving efficiency.

In the second part, we first integrate learning processes in active contour segmentation. On one hand we integrate a statistical model learning shape prior and, on the other hand, we design a new semi-supervised interactive segmentation scheme which iteratively learns the object to segment.

Finally, we present a new algorithm based on boosting for interactive object retrieval in images. In this last part, we address more specifically the problem of video classification and retrieval, providing a new representation of video objects which fully holds spatiotemporal information then designing a system able to either learn semantic classes of objects or provide accurate object recognition.