Annotation d'images sur de grands corpus réels de données

In this paper, we check the limitations of image annotation on a large real corpus. We built a corpus of documents containing text and pictures, using more than 25000 press articles. We use this data to compare the similarity between a text-based retrieval and an image-based retrieval. The retrieval systems we use are common Information Retrieval (IR) tools. Results show that the two IR systems we used do not work well together, so associating the descriptors we used is not suitable to annotate real data. Then it is necessary to find a model than can handle textual and visual descriptors together, in order to make image annotation possible. MOTS-CLÉS : Annotation d’images, propagation de mots-clefs, indexation texte-image, fossé sémantique, histogramme de couleurs