Improving CBIR Systems by Integrating Semantic Features

• Venue: CRV '04 Proceedings of the 1st Canadian Conference on Computer and Robot Vision
• Year: 2004

Nowadays, access to information requires to manageeffectively multimedia databases, and among challengesoffered to scientific community since last decades,multimedia retrieval techniques (particularly imagesretrieval) are became an active research direction.Introduced to overcome the main drawbacksencountered by text-based images retrieval, which arethe subjective and manual annotation of images, contentbased images retrieval (CBIR) systems index imagesaccording to low-level visual features such as color,texture, shape to retrieve similar images. However,despite the progress achieved in the content based imageretrieval, in particular with the relevance feedbackapproach where the user refine the search via thespecification of relevant or not relevant items, thecurrent CBIR systems still have a major difficulty that ithas yet to overcome: how to negotiate the "semanticgap"? This problem comes from the mismatch betweentheir capabilities and the needs of users.In this paper, we address the problem of how relate low-levelfeatures to high level to bring out semanticconcepts from images. Our aim is to combine content-basedand metadata-based approaches for imageretrieval from a user perspective to yield better resultsand overcome to the lacks of these techniques whenthey are taken separately. To represent the semanticcontent of images, we propose a model which takesaccount of the interaction between the user and themetadata. In particular, we model the semantic user'preference by analyzing its answers through theRelevance Feedback process. Furthermore, weintroduce a new machine learning technique that modifythe weights (i.e. relative importance) of metadatarepresenting the semantic content of images.