Thematico-Visual Image Retrieval: How to Deal With Partially Indexed Corpora

Gérald Duffing 1
1 LANGUE ET DIALOGUE - Human-machine dialogue with a significant language component
INRIA Lorraine, LORIA - Laboratoire Lorrain de Recherche en Informatique et ses Applications
Abstract : It becomes very easy to access large amounts of images when surfing the Internet. All images, however, are not always thematically indexed. We think that partially thematically indexed corpora can be organised in a way that facilitates retrieval. We assume that, concerning visual properties, the corpus is totally indexed by means of ``generic'' features. Based on these indexes, a hierarchical clustering technique is used to bring together images that share some similarities: two distinct structures are built (``dendrograms''). We propose a new retrieval strategy based on a virtual image that captures the user's need along the retrieval session, taking into account both thematic and visual aspects. Clusters are successively selected in each dendrogram. A combined method, called tunnels, allows dendrograms cooperation. Images are then ranked according to the virtual image. After each retrieval step, the virtual image is enriched within a relevance feedback process. Theme, colour and general layout of each images can be rated and the query is updated accordingly. In our experiments, we used two different corpora (2470 and 1100 images) to assess the performance of our thematico-visual approach within different indexing conditions. Experimentation results confirm the relevance of our approach and suggests improvement possibilities.
Type de document :
Communication dans un congrès
Internet Imaging II, 2001, San Jose, CA, 12 p, 2001
Liste complète des métadonnées
Contributeur : Publications Loria <>
Soumis le : mardi 26 septembre 2006 - 08:56:47
Dernière modification le : jeudi 11 janvier 2018 - 06:19:48
Document(s) archivé(s) le : mercredi 29 mars 2017 - 12:38:36



  • HAL Id : inria-00099396, version 1



Gérald Duffing. Thematico-Visual Image Retrieval: How to Deal With Partially Indexed Corpora. Internet Imaging II, 2001, San Jose, CA, 12 p, 2001. 〈inria-00099396〉



Consultations de la notice


Téléchargements de fichiers