Journal of Intelligent Information Systems

, Volume 19, Issue 2, pp 207–230

Combining Approximation Techniques and Vector Quantization for Adaptable Similarity Search

  • Christian Böhm
  • Hans-Peter Kriegel
  • Thomas Seidl
Article

DOI: 10.1023/A:1016515829761

Cite this article as:
Böhm, C., Kriegel, HP. & Seidl, T. Journal of Intelligent Information Systems (2002) 19: 207. doi:10.1023/A:1016515829761

Abstract

Adaptable similarity queries based on quadratic form distance functions are widely popular in data mining application domains including multimedia, CAD, molecular biology or medical image databases. Recently it has been recognized that quantization of feature vectors can substantially improve query processing for Euclidean distance functions, as demonstrated by the scan-based VA-file and the index structure IQ-tree. In this paper, we address the problem that determining quadratic form distances between quantized vectors is difficult and computationally expensive. Our solution provides a variety of new approximation techniques for quantized vectors which are combined by an extended multistep query processing architecture. In our analysis section, we show that the filter steps complement each other. Consequently, it is useful to apply our filters in combination. We show the superiority of our approach over other architectures and over competitive query processing methods. In our experimental evaluation, the sequential scan is outperformed by a factor of 2.3. Compared to the X-tree on 64 dimensional color histogram data, we measured an improvement factor of 5.7.

multimedia retrieval content-based search ellipsoid queries 

Copyright information

© Kluwer Academic Publishers 2002

Authors and Affiliations

  • Christian Böhm
    • 1
  • Hans-Peter Kriegel
    • 2
  • Thomas Seidl
    • 3
  1. 1.University for Health Informatics and Technology TyrolInnsbruckAustria
  2. 2.Institute for Computer ScienceUniversity of MunichMünchenGermany
  3. 3.Department of Computer and Information ScienceUniversity of ConstanceKonstanzGermany

Personalised recommendations