Thèse soutenue

FR
Auteur / Autrice : Sébastien Poullot
Direction : Michel Crucianu
Type : Thèse de doctorat
Discipline(s) : Informatique
Date : Soutenance en 2009
Etablissement(s) : Paris, CNAM
Partenaire(s) de recherche : autre partenaire : Institut national de l'audiovisuel (France1986-....)

Mots clés

FR

Mots clés contrôlés

Résumé

FR  |  
EN

This thesis essentially adresses the scability of the indexong methods of vectorial databases. The applications concern the similarity-based search of video descriptors in large volumes in order to perform content-based copy detection. On one hand we want to perform an online monitoring of a video stream on a reference database, containing here 280000 hours of video, which means 17 billions of descriptors. The proposed solution is based on a new indexing and probalistic searching method based on a Zgrid, but also on a distorsion of the video descriptors and on a local density model. The goal is to perform a more selective and so faster similarity search. Here we can handle the monitoring of one video stream on the 280000 hours database in a differed real time with a single standard PC. On the other hand we want to detect the occurences of the videos in a such a large database. The problem become quadratic, here a similarity self join of the descriptor database must be performed. Here we propose a new global description of the frames based on a local descriptions to reduce complexity while conserving a good tobustness. We also propose an indexing scheme apated to this task which presents moreover an easily parrallel scheme in order to mine the previously announced volumes. Our tests have been performed on dtabases containing up to 10000 hours of video in 80 hours with a single standard PC