(SIFT) [2] or Speeded Up Robust Features (SURF) [3] extracted
from images, can mainly be classified into three
categories of indices: tree-based index, hash-based index,
and visual-word-based inverted index [4–7]. In a treebased
indexing structure, such as KD-tree [8] or R-tree
[9], when the dimension of a feature descriptor grows
greater than 20, the construction efficiency deteriorates
rapidly. Both categories of hash-based index – the Euclidian
Locality Sensitive Hashing (E2LSH) [10] related methods
and the spectral hash methods [11–13] – are inef-
ficient on sparse descriptors. In a visual-word-based inverted
index, the traditional flat K-means [14–17] clustering
algorithm and the improved Approximate K-Means