Using High Dimensional Indexes to Support Relevance Feedback Based Interactive Images Retrieval

1
Using High Dimensional Indexes to Support
Relevance Feedback Based Interactive Images
Retrieval Junqi Zhang Xiangdong Zhou
Wei Wang Baile Shi Jian Pei Fudan
University, China Simon Fraser
University, Canada
Motivation K-means cluster approach had been
widely used to improve the performance of high
dimensional index. But, there are still some
problems need to be discussed further, such as
how to preset the query radius and the number K
of the K-means cluster, etc. In this demo system,
we present a new cluster splitting based B-tree
index to deal with the above problems, and the
index has been applied to support relevance
feedback for content-based images retrieval.
Index Structure
Background The central idea of iDistance is to
cluster objects and find a reference point for
each cluster. Then, the distance between an
object and the reference point in the cluster to
which the object belong can be indexed in a
B-tree. It has been well observed that in the
high dimensional real data space, a majority of
clusters are intersected each other. Therefore,it
is often the case, that a query region covers
many clusters and causes lower query
efficiency. In order to improve the query
performance, the iDistance search algorithm
starts with a preset small search radius and
enlarges the search radius gradually if
necessary.
Experiment results

Challenges However, in the known work, the
initial query radius and the enlarging step need
to be preset by experiment or users experiences.
It is lack of theory to guide the estimation of
these parameters.
Demo system 1. Based on the query cost model of
metric space, we developed the formulas to
compute the optimal cluster splitting number
M. 2. In the interactive relevance feedback
processing, the query distance is updated using
users feedback and the index distance is
guaranteed to be a lower bound of the query
distance. Thus, the index structure does not need
to be changed.
,
Nc the number K of K-means cluster N size of
dataset H height of internal node u fanout of
node
Approach We present a new cluster splitting
based B-tree index to deal with the above
problems, 1. The optimal KNN search algorithm is
adopted to avoid the selection of initial query
radius 2. Through cluster splitting, the data
space is partitioned more finely to reduce the
intersection between query region and data
clusters