Skip to Main content Skip to Navigation

Indexation sémantique des images et des vidéos par apprentissage actif

Abstract : The general framework of this thesis is semantic indexing and information retrieval, applied to multimedia documents. More specifically, we are interested in the semantic indexing of concepts in images and videos by the active learning approaches that we use to build annotated corpus. Throughout this thesis, we have shown that the main difficulties of this task are often related, in general, to the semantic-gap. Furthermore, they are related to the class-imbalance problem in large scale datasets, where concepts are mostly sparse. For corpus annotation, the main objective of using active learning is to increase the system performance by using as few labeled samples as possible, thereby minimizing the cost of labeling data (e.g. money and time). In this thesis, we have contributed in several levels of multimedia indexing and proposed three approaches that outperform state-of-the-art systems: i) the multi-learner approach (ML) that overcomes the class-imbalance problem in large-scale datasets, ii) a re-ranking method that improves the video indexing, iii) we have evaluated the power-law normalization and the PCA and showed its effectiveness in multimedia indexing. Furthermore, we have proposed the ALML approach that combines the multi-learner with active learning, and also proposed an incremental method that speeds up ALML approach. Moreover, we have proposed the active cleaning approach, which tackles the quality of annotations. The proposed methods were validated through several experiments, which were conducted and evaluated on large-scale collections of the well-known international benchmark, called TrecVid. Finally, we have presented our real-world annotation system based on active learning, which was used to lead the annotations of the development set of TrecVid 2011 campaign, and we have presented our participation at the semantic indexing task of the mentioned campaign, in which we were ranked at the 3rd place out of 19 participants.
Document type :
Complete list of metadatas

Cited literature [175 references]  Display  Hide  Download
Contributor : Abes Star :  Contact
Submitted on : Wednesday, December 19, 2012 - 11:12:13 AM
Last modification on : Thursday, November 19, 2020 - 12:59:57 PM
Long-term archiving on: : Wednesday, March 20, 2013 - 11:32:50 AM


Version validated by the jury (STAR)


  • HAL Id : tel-00766904, version 1



Bahjat Safadi. Indexation sémantique des images et des vidéos par apprentissage actif. Autre [cs.OH]. Université de Grenoble, 2012. Français. ⟨NNT : 2012GRENM073⟩. ⟨tel-00766904⟩



Record views


Files downloads