Optimal Estimation of Descriptor Scales for Multimedia Retrieval

Relevance feedback schemes use a given set of descriptors and at each feedback iteration they just modify the importance of each descriptor to the retrieval performance. This implies that the same type of descriptors is used for all feedback iterations. Therefore, such approaches cannot improve the retrieval performance beyond the efficiency of the descriptors used to represent the multimedia content. In this paper, an alternative concept is to adopt a progressive descriptor organization and to allow dynamic modification of the type of descriptors (expansion or shrink) at each retrieval iteration by exploiting the current user's information needs. In other words, descriptors that best fit the current user's preferences are expanded into more details, in constant to the descriptors that are far way from the user's needs which are shrinking. In this case, instead of using a large set of descriptors, most of them inappropriate with respect to the current user's information needs, a small number of dominant descriptors is selected for the retrieval process. Thus, the proposed approach significantly increases the precision performance by discarding descriptors that yield noise in the retrieval.