Skip to Main content Skip to Navigation
Theses

Indexation symbolique d'images : une approche basée sur l'apprentissage non supervisé de régularités

Abstract : This work deals with automatic indexing of personal photographs by highly abstract visual concepts. We make a case for unsupervised learning, by putting forward arguments against the use of supervised learning alone. We introduce a new unsupervised learning paradigm based on two kinds of regularities, implementing respectively to the notion of structure and the notion of contextual similarity. These regularities are inducted from a stream of visual data and are stored as new nodes in a growing network. New data is systematically recoded in terms of previously acquired knowledge, thus continuously changing the lens through which the data is seen. Experiments on real visual data, as well as synthesized data, show that our approach creates ‘relevant' recoding features, yielding better indexing results. From these very promising results, we draw a number of ambitious directions for future works.
Document type :
Theses
Complete list of metadata

Cited literature [80 references]  Display  Hide  Download

https://tel.archives-ouvertes.fr/tel-00011315
Contributor : Stéphane Bissol <>
Submitted on : Friday, January 6, 2006 - 6:00:11 PM
Last modification on : Friday, November 6, 2020 - 4:02:17 AM
Long-term archiving on: : Saturday, April 3, 2010 - 8:57:29 PM

Identifiers

  • HAL Id : tel-00011315, version 1

Collections

UJF | IMAG | CNRS | UGA

Citation

Stéphane Bissol. Indexation symbolique d'images : une approche basée sur l'apprentissage non supervisé de régularités. Interface homme-machine [cs.HC]. Université Joseph-Fourier - Grenoble I, 2005. Français. ⟨tel-00011315⟩

Share

Metrics

Record views

413

Files downloads

1264