A scalable and efficient content-based multimedia retrieval system

In this work the problem of content-based information retrieval is approached from a new perspective. We look at a probabilistic approach in CBIR from the angle of Bayesian networks. Our data structure serves to break two bottlenecks of retrieval performance: (1) high dimensionality of feature vectors and (2) poor mapping of raw features into highlevel content that a human understands (the semantic gap). We use the network structure instead of the feature space, and propose updating the higherlevel content description by utilising the relevance feedback obtained from the user. Strategies for display update for the next iteration are studied. A new approach for selecting the next display set is tied to our data structure.