A multi-similarity algebra

Abstract

The need to automatically extract and classify the contents of multimedia data archives such as images, video, and text documents has led to significant work on similarity based retrieval of data. To date, most work in this area has focused on the creation of index structures for similarity based retrieval. There is very little work on developing formalisms for querying multimedia databases that support similarity based computations and optimizing such queries, even though it is well known that feature extraction and identification algorithms in media data are very expensive. We introduce a similarity algebra that brings together relational operators and results of multiple similarity implementations in a uniform language. The algebra can be used to specify complex queries that combine different interpretations of similarity values and multiple algorithms for computing these values. We prove equivalence and containment relationships between similarity algebra expressions and develop query rewriting methods based on these results. We then provide a generic cost model for evaluating cost of query plans in the similarity algebra and query optimization methods based on this model. We supplement the paper with experimental results that illustrate the use of the algebra and the effectiveness of query optimization methods using the Integrated Search Engine (I.SEE) as the testbed.