Description
Date depot: 1 janvier 1900
Titre: Recherche mentale dans les espaces d'images / Subspace methods for mental image search
Directeur de thèse:
Hichem SAHBI (LIP6)
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Non defini
Resumé:
At least, two interrogation modes are known in content based image retrieval (CBIR); keyword-based search and query by image examples. In the first mode a user submits keywords and the search engine returns a set of images annotated with those keywords. When keywords are not available or difficult to suggest, querying by example is a good alternative in which a user submits an image as an example of her/his “class of interest” and the search system displays the closest image(s) using a feature space and a suitable distance [1]. A variant of this mode, known as relevance feedback (eg. [2,3,4]), consists in interactively choosing multiple relevant examples (and possibly irrelevant ones), learning classifiers and iteratively updating results until finding the target. Nevertheless, it is not always possible to find examples that accurately express the semantic(s) wanted by users.
A novel image search paradigm has emerged, during the last years, referred to as “mental search” [5-8]. In this mode, no query examples are available; a user is instead asked questions, and according to her/his answers, decision criteria are learned in order to iteratively refine the results, close the semantic gap and find the mental target. A possible scenario is a user looking for some location in a large database of photos without remembering the exact name of that location. It is clear that neither keyword-based search is helpful (as photos may not be annotated, and if so the user may not remember the name of the location) nor image examples as the latter are not always available.
Doctorant.e: Oliveau Quentin