Content-based retrieval in distributed image databases

Abstract

Image databases on the Web have heterogeneous characteristics since they use different similarity measures and queries are processed depending on their own schemes. In the content-based image retrieval from distributed sites, it is crucial that the metaserver has the capability to find objects, similar to a given query object in terms of the global similarity measure, from different image databases with different local similarity measures. First, we investigate (1) the problem of finding databases, which contain more objects relevant to a given query than other databases, from many image databases dispersed on the Web, and (2) the problem of retrieving relevant images by content-based retrieval from selected image databases. The former is referred to as a database selection problem and the latter is referred to as a collection fusion problem. We introduce a new selection method to determine candidate databases. The selection of databases is based on the hybrid estimator using a few sample objects and compressed histogram information of image databases. The histogram information is used to estimate the selectivity of spherical range queries and a small number of sample objects are used to compensate the selectivity error due to the difference of the similarity measures between metaserver and image databases. The collection fusion of image databases is concerned with the merging of results retrieved by content-based retrieval from heterogeneous image databases on the Web. We also propose new collection fusion methods using two heuristic estimators: average ranking and average global similarity, the ordinary regression model, and the Bayesian regression model. Among them, the probabilistic technique using Bayesian regression model outperforms the other approaches for diverse sizes of result sets for a query. Its improvement in effectiveness becomes especially large with small sizes of result sets. And, we present a virtual optimal algorithm to which our algorithm ...