Learning Over Multiple Temporal Scales in Image Databases

Abstract. The ability to learn from user interaction is an important asset for content-based image retrieval (CBIR) systems. Over short times scales, it enables the integration of information from successive queries assuring faster convergence to the desired target images. Over long time scales (retrieval sessions) it allows the retrieval system to tailor itself to the preferences of particular users. We address the issue of learning by formulating retrieval as a problem of Bayesian inference. The new formulation is shown to have various advantages over previous approaches: it leads to the minimization of the probability of retrieval error, enables region-based queries without prior image segmentation, and suggests elegant procedures for combining multiple user speci cations. As a consequence of all this, it enables the design of short and long-term learning mechanisms that are simple, intuitive, and extremely e cient in terms of computational and storage requirements. We introduce two...

• Type: Conference
• Year: 2000
• Where: ECCV
• Authors: Nuno Vasconcelos, Andrew Lippman
• Keywords: Computer Vision, Content-based Image Retrieval, ECCV 2000, Long-term Learning Mechanisms, Prior Image Segmentation, Retrieval Error, Retrieval Sessions