Nuno Vasconcelos and Andrew Lippman,
Embedded Mixture Modeling for Efficient Probabilistic
Content-based Indexing and Retrieval,
In C.-C. Jay Kuo, Shih-Fu Chang and Sethuraman
Panchanathan eds., Multimedia Storage and Archiving Systems III (VV02),
Boston, Massachusetts, USA, Vol. 3527 of SPIE Proceedings, pp. 134-143, November 1998. (SPIE Symposium on Voice, Video and Data Communications)By formulating content�based retrieval as a problem of Bayesian inference we have previously developed a retrieval framework with various interesting properties: 1) allows the incorporation of prior beliefs about image relevance in the retrieval process, 2) leads to simple and intuitive mechanisms for combining information from several modalities, such as images, audio, and text during retrieval, 3) provides support for the development of interfaces that learn from user interaction, 4) allows retrieval directly from compressed bitstreams, and 5) lends itself to the construction of indexing structures which can also be computed as a side effect of the compression process. In this paper, we analyze the relationships between probabilistic retrieval using mixture models and two approaches commonly found in the retrieval literature, and introduce an alternative criterion, the random sample likelihood that is computationally cheaper and more robust. By performing all the modeling in the frequency domain, we obtain an embedded set of representations that operate on subspaces of the feature space. When only the coarsest of these subspaces is considered for retrieval, our model becomes a simple histogram leading to very efficient retrieval. However, as more subspaces are considered, the model is also able to capture higher order dependencies between image pixels, allowing for more accurate retrieval. The representation is therefore suited to the implementation of multistage filtering strategies where low�complexity and low�accuracy methods are used in the early stages to eliminate clearly inappropriate matches and higher accuracy methods are employed in subsequent stages to select the best matches within the set of surviving hypothesis. Simulation results show that the new retrieval criterion beats the standard ones even when the former is applied to a generic feature space and the latter ones to a feature space highly specialized to the type of imagery in the database.