Abstract

Previous algorithms of data partitioning methods (DPMs) to find the exact K-nearest neighbors (KNN) at high dimensions are outperformed by a linear scan method [J.M. Kleinberg, Two algorithms for nearest neighbor search in high dimensions, 29th ACM Symposium on Theory of computing, 1997; R. Weber, H.-J. Schek, S. Blott. A quantitative analysis and performance study for similarity-search methods in high-dimensional spaces. in: Proc. of the 24th VLDB, USA, 1998]. In this paper, we present a “plug&search” method to greatly speed up the exact KNN search of existing DPMs. The idea is to linearize the data partitions produced by a DPM, rather than the points themselves, into a one-dimensional array-index, that is simple, compact and fast. Unlike most DPMs that support KNN search, which require storage space linear, or exponential [J.M. Kleinberg, Two algorithms for nearest neighbor search in high dimensions, 29th ACM Symposium on Theory of computing, 1997; M. Hagedoom, Nearest neighbors can be found efficiently if the dimension is small relative to the input size, ICDT 2003], in dimensions, the array-index requires a storage space that is linear in the number of mapped partitions.

Keywords: Indexing methods; Image databases; KNN image search, array-index, plug&search method

Article Outline

1. Introduction

2. Array-index method

2.1. Partition linearization

2.2. Array-index construction

3. KNN search method

3.1. Traversal control of KNN search

3.1.1. Partition pruning condition

3.1.2. Partition ending condition

3.1.3. Point pruning condition

3.1.4. Point ending condition

3.2. KNN search algorithm

4. Experiments

4.1. Extraction of feature vectors

4.2. Reference node selection

4.3. KNN search efficiency

4.4. Insertion and deletion

4.5. Comparative study

5. Conclusion

Appendix A. Appendix

A.1. Proof of Lemma 1

A.2. Proof of Lemma 2

A.3. Proof of Lemma 3

A.4. Proof of Lemma 4

References

Vitae