RelaxImage: A Cross-Media Meta-Search Engine for Searching Images
from Web Based on Query Relaxation
Akihiro Kuwabara and Katsumi Tanaka
Graduate School of Informatics, Kyoto University, Japan
{kuwabara, tanaka} @dl.kuis.kyoto-u.ac.jp
Abstract
We introduce a cross-media meta-search engine
RelaxImage for searching images from Web. Notable
features of the RelaxImage are as follows: (1) Each
user’s keyword query is “relaxed”, that is, by
gradually relaxing the search terms used for image
search, we can solve the problem of conventional
image search engine such as Google. (2) For
searching images, our RelaxImage sends a different
keyword-query to each search engine of different
media-type. We show several examples of how the
relaxation approach works as well as ways that it can
be applied. That is, our RelaxImage shows a great
improvement for increasing recall ratio without
decreasing of precision ratio.
1. Introduction
In this paper, we describe a query relaxation approach
designed to realize a cross-media meta-search engine
for images[1]. The cross-media meta-search engine for
images intuitively means a meta-search engine, in
which (1) it delivers a user’s query with some
modification to multiple search engines dedicated for
searching different media type content, (2) each search
engine returns collections of Web pages containing the
specified keywords, (3) the intersection of the searched
collections is taken, and (4) images contained in the
intersection are returned.
Our
query
relaxation
approach
intuitively
decomposes a user-input conjunctive keyword query
into sets of conjunctive subqueries, in which the
subqueries are automatically routed to the various
types of search engines. This idea was motivated by
the fact that conventional image search engines, such
as Google does not succeed in finding images for
conjunctive keyword queries consisting of too many
keywords. Evaluation of our query relaxation approach
shows that it can dramatically improve the recall ratio
for image retrievals. By the result of an experiment,
query relaxation can improve recall considerably,
maintaining precision to about 70%.
2. Query Relaxation for Image Search
We assume that users wish to have multimedia
content (texts, images, sounds, etc.) that is related to
their keywords k1, k2, … , kn. Our cross-media metasearch engine uses various search engines (namely, E1,
E2, …, Em). For example, E1 is Google, E2 is AltaVista,
E3 is Google image search, and so on. Thus, the key
feature of our system is that it can use a mixture of
search engines with different media types by
translating or relaxing the user-input conjunctive query
Q, and then sending the relaxed sub-query to each
search engine.
Let Ans (Q) be the set of answers for a user-input
conjunctive keyword query Q. In our query relaxation
approach, the conjunctive query Q is divided into a set
of tuples bound by sub-query:
In this case, the query is divided into two separate
sub-queries, because two kinds of search engines are
required for the query. Generally, the number of subqueries is dependent on the number of search engines.
The sub-query for each set of tuples is translated for
the image search engine. For example, the first tuple
(i.e. [ I , {k1 and k2 and ... and kn}]) shows that there
are no keywords for the text search engine, and k1, k2,...
and kn are the input keywords for the image search
engine.
Answers for query Q are retrieved as unions of all the
sub-query tuples:
Ans(Q) ( Ans(k1 ,... k n , E 2 ))
( Ans(k1 , E1 ) Ans(k 2 ... k n , E2 ))
( Ans (k 2 , E1 ) Ans (k1 k 3 ... k n , E 2 ))
...
( Ans(k1 k 2 , E1 ) Ans(k 3 ... k n , E2 ))
( Ans(k1 k 3 , E1 ) Ans(k 2 k 4 ... k n , E2 ))
...
( Ans(k1 ... k n , E1 )) where Ans(Q, Ei) means the answers of the query Q by
the search engine i. In this case, E1 is a text search
engine and E2 is an image search engine. The engines
return URLs which match the queries. Then the
intersection of the answers is calculated. Finally,
Ans(Q) can be retrieved as unions of all the answers.
Proceedings of the 21st International Conference on Data Engineering (ICDE 2005)
1084-4627/05 $20.00 © 2005 IEEE
3. Query Relaxation Experiment
We performed a simple experiment to relax the query.
The number of keywords is set at three. We randomly
chose keywords from the headlines in the news Web
pages. We define the degree of relaxation as the
number of keywords in a user-input query Q that are
actually used by a Web text search engine. That is,
when three of the keywords are used for an image
search engine, the degree of relaxation is considered to
be zero. When two keywords are used for an image
search engine and the third keyword is used for a text
search engine, the degree of relaxation is considered to
be one. We compare the results of 0 of relaxation, to 1
of relaxation, and 2 of relaxation respectively.
(i) Type A
Fig.1 Lattice of query Q
(Q = Mt. Fuji ш snow ш sunset)
The results of the experiment are shown in Fig.2. The
recall and precision graphs for each keyword are also
shown in Fig.2. The precision and recall are displayed
for every degree of relaxation. Precision is defined as
the average precision for every degree of relaxation.
Recall is defined as the total number of pertinent pages
found for each of experiment. It is assumed that the
recall for 2 of relaxation is 100%. Therefore, the
results are relative.
The recall ratio for the results increases from 0 to 1
for every degree of relaxation. That is, increasing the
number of keywords for the text search engine is an
effective means of obtaining pertinent Web pages.
In the graph, there are two patterns. Type (A) is
normal: the recall ratio is increased and the precision
ratio is decreased according to the degree of relaxation.
On the other hand, type (B) shows from 0 to 1 of
relaxation with and increased precision and recall ratio
that is particularly apparent. No answers were obtained
at 0 of relaxation, but by relaxing the degree, many
answers were obtained. By the result of an experiment,
query relaxation can improve recall considerably,
maintaining precision to about 70% at one of degree.
(ii) Type B
Fig.2 Experimental results
4. Conclusion and Future Work
We have described a cross-media meta-search
method that uses a query relaxation approach to obtain
pertinent Web pages. We have demonstrated our
approach and shown examples of how to calculate it.
We believe that the query relaxation approach is an
efficient method for image searches.
5. References
[1] Akihiro Kuwabara, Kazutoshi Sumiya, Katsumi Tanaka,
“Query
Relaxation
and
Answer
Integration
for Cross-Media Meta-Searches,” Proc. of IEEE ICME2004,
June 2004.
Proceedings of the 21st International Conference on Data Engineering (ICDE 2005)
1084-4627/05 $20.00 © 2005 IEEE
© Copyright 2026 Paperzz