Rate-distortion approach to databases: storage and content-based retrieval

The problem of similarity search is central to a wide range of applications in multimedia databases. The degree of similarity is often quantified by a distance measure defined over the space of extracted feature vectors. In typical applications, since the volume of data is huge, and the feature vectors are of high dimensionality, it becomes necessary to access the feature vectors from a hard storage medium during the search. Since I/O operations on hard storage devices are slow, the time complexity of the search is dominated by the I/O time. One of the most effective approaches to reduce the time complexity is compromising the search accuracy by accessing compressed feature vectors. The processing time is then approximately proportional to the rate R at which the sequence of feature vectors is compressed. In this paper we characterize the set of all achievable time-accuracy pairs.