Database Reference
In-Depth Information
Others were built on top of a DBMS, but with retrieval mechanisms detached from its query processor.
Nevertheless, this approach prevents using several optimization alternatives when executing complex
queries, especially when search conditions also contain operators over simple data, which are efficiently
handled by DBMS.
In order to add support for similarity queries in a RDBMS it is necessary: (i) to create a representa-
tion for multimedia data, (ii) to define how the similarity evaluation is carried out, (iii) to state how a
query involving similarity operations is written and (iv) to provide mechanisms to execute the queries
efficiently. The last requisite have been fulfilled by several works reporting basic algorithms to execute
similarity retrieval operations on data sets of multimedia objects (as described in the 'Indexing Methods
for Multimedia' section). The first two requisites have also been extensively addressed in the literature,
as it was discussed in the 'Similarity Evaluation' section. The third requirement is directly related to
the most widely employed language in DBMS, the SQL. Although a few works focused on languages
to represent similarity queries (Carey and Kossmann, 1997, Carey and Kossmann, 1998, Melton and
Eisenberg, 2001, Gao et al., 2004), none of them is able to provide a production-strength support seam-
lessly integrated with the other features of the language.
This section presents existing DBMS solutions aimed at similarity retrieval and their limitations, and
introduces an extension to SQL that allows managing similarity data inside a DBMS in a consistent way.
Current DBMS Solutions for Similarity Searching
The ISO/IEC SQL/MM (Melton and Eisenberg, 2001) is the piece of the SQL standard devoted to multime-
dia data. It defines data types and operators to manage multimedia information, such as multidimensional
data, images and large texts. Concerning similarity, the Still Image part of SQL/MM defines an abstract
data type that encapsulates a binary image and its metadata, a set of basic image processing functions
and data types aimed at executing similarity queries over the image contents. However, it addresses
too superficially the internals of the image representation and the query representation and processing.
A few DBMS vendors released modules to query multimedia data by content. The precursors were
the Audio, Image and Video (AIV) Extenders of the IBM DB2 database (IBM Corp., 2003), which were
based on QBIC (the first commercial strength system for multimedia retrieval) (Flickner et al., 1995).
Other examples are the Informix Excalibur Image Data Blade (Informix Corp., 1999) and the Oracle
interMedia (Oracle Corp., 2005) (now called Oracle Multimedia). They provide proprietary data types
with feature extractors and similarity functions that are used in SQL statements.
As an example, consider a relation to store photographs of landscapes defined as: (see Algorithm 1)
Algorithm 1.
CREATE TABLE Landscapes (
Id INTEGER,
Place CHAR(20),
Picture IMAGETYPE,
Photographer CHAR(30),...);
