Skip to Main content Skip to Navigation
Reports

Active SVM-based Relevance Feedback with Hybrid Visual and representation

Abstract : Most of the available image databases have keyword annotations associated with the images, related to the image context or to the semantic interpretation of image content. Keywords and visual features provide complementary information, so using these sources of information together is an advantage in many applications. We address here the challenge of semantic gap reduction, through an active SVM-based relevance feedback method, jointly with a hybrid visual and conceptual content representation and retrieval. We first introduce a new feature vector, based on the keyword annotations available for the images, which makes use of conceptual information extracted from an external ontology and represented by ``core concepts''. We then present two improvements of the SVM-based relevance feedback mechanism: a new active learning selection criterion and the use of specific kernel functions that reduce the sensitivity of the SVM to scale. We evaluate the use of the proposed hybrid feature vector composed of keyword representations and the low level visual features in our SVM-based relevance feedback setting. Experiments show that the use of the keyword-based feature vectors provides a significant improvement in the quality of the results.
Document type :
Reports
Complete list of metadata

https://hal.inria.fr/inria-00070448
Contributor : Rapport de Recherche Inria <>
Submitted on : Friday, May 19, 2006 - 8:31:27 PM
Last modification on : Thursday, February 6, 2020 - 2:16:06 PM
Long-term archiving on: : Sunday, April 4, 2010 - 9:14:46 PM

Identifiers

  • HAL Id : inria-00070448, version 1

Collections

Citation

Marin Ferecatu, Michel Crucianu, Nozha Boujemaa. Active SVM-based Relevance Feedback with Hybrid Visual and representation. [Research Report] RR-5558, INRIA. 2005, pp.20. ⟨inria-00070448⟩

Share

Metrics

Record views

269

Files downloads

427