Database Reference
In-Depth Information
Although this approach is suitable to specific applications, the amount of operators needed to repre-
sent queries can be excessively high. The operators employed to identify objects should be specialized
to each application domain to acquire higher semantics, yielding a variety of functions to represent each
(class of) object(s). Moreover, although there are works aimed at retrieving images by spatial similarity
between segmented regions, such as (Huang et al., 2008, Yeh and Chang, 2008), queries are usually
represented using a sketch or an example of the desired spatial relationships, because representing the
association among various objects can be very difficult.
The representation of more complex classes of queries is even harder using content-based operators.
Queries Q3, Q4 and Q5 would require defining respectively: what determines a video to be copyrighted,
what is the visual pattern that represents the act of playing, and what a drum solo is. It is easy to notice
that this approach would lead to an explosion on the amount of operators to meet every situation and
the variety of relationships among complex elements and among their components.
The Similarity Operators Approach
We define the
Similarity Operators approach
as the one that provides a few similarity-based operators
to be employed in a wide range of situations. These operators have the goal of computing the similarity
regarding a predefined pattern between the stored data and the (set of) element(s) provided in the query,
verifying when this similarity satisfies a given criteria.
The support for the Similarity Operators approach is twofold: providing an example of the desired
pattern is generally more feasible than trying to describe it; and providing a restricted set of operators
is sufficient to represent queries over complex data, because the peculiarity of each domain is carried
by its definition of similarity.
The first supporting idea is derived from the query-by-example strategy that has been employed in
databases for long time. Most of the existing content-based multimedia retrieval tools use this method.
The second idea aims at simplifying the overall query answering process to allow developing a consistent
retrieval environment. This simplification can be done without losing the generality of the approach
because the definition of similarity is
problem-dependent
. For instance, a good similarity measure to
answer Q2 would focus on the color distribution pattern of the images. This measure would probably
produce poor results if applied to Q1, because what identifies a mass in a radiograph is usually the
difference between the mass and its surrounding texture tissue, which is not adequately captured by
a standard color-based evaluation, thus requiring a texture extractor. The domain knowledge of each
problem class is embedded into the similarity evaluation process, which must be specialized according
to the requirements of each case. This allows stating specialized query conditions to a broad range of
domains employing few query operators.
The reader may ask whether we are not just transferring the aforementioned problem of the explosion
on the amount of operators for multimedia data querying to the explosion of similarity measures. In part
this observation is correct. Nevertheless, maintaining a reduced set of similarity operators allows devel-
oping a common framework with well-defined query optimization alternatives and with basic primitives
and structures that can be applied to a variety of situations. Providing such common framework is not
feasible using the Content-based Operators approach.
Therefore, this chapter adopts the Similarity Operators approach for representing queries. The next
section discusses similarity evaluation regarding multimedia data.
