Database Reference
In-Depth Information
from its tables, SIREN joins the system tables and the user tables, removing the feature attributes, thus
the user never sees the table split nor the features. This is the same approach for the treatment of BLOB
data in DBMS, which are stored apart from the original table into system-controlled areas and only
references to them are stored in the table.
When the user poses queries involving similarity predicates, SIREN uses the extracted features to
execute the similarity operators. The current version of SIREN has three types of feature extractors
regarding the STILLIMAGE data type: a texture extractor (TEXTUREEXT), a shape extractor based
on Zernike Moments (ZERNIKEEXT) and a color extractor based on the normalized color histogram
(HISTOGRAMEXT). For sound objects storing music, there are the a sound-texture extractor (SOUND-
TEXTUREEXT), which extracts the Mel-Frequency Cepstral Coefficients (MFCC) and features based
on the Short Time Fourier Transform (STFT). The 'A Look on the Feature Vector Component' section
presented references of these extractors.
The similarity operators implemented consist of the similarity selections for single query centers, the
similarity selections for multiple query centers and the similarity joins, as presented in the 'Similarity
Queries' section. The traditional single center similarity selections, that is, the similarity range query
and the
k
-nearest-neighbor query operators, are available in several MAM presented in the literature.
However, regarding the multiple center similarity selections, there are operators available only for
the Slim-tree MAM. Therefore, the Slim-tree MAM (Traina-Jr et al., 2002) is employed to index the
multimedia attributes. The Slim-tree is implemented in a C++ access method library, which is called
from the SIREN Indexer subsystem to execute the similarity-related operations. Unfortunately, there is
no procedure already published to execute similarity joins in this MAM. Therefore, in SIREN they are
always executed using sequential scan implementations.
Following, it is used the previous example to illustrate how SIREN executes a query asking for the
five closest landscapes from a given picture. (see Algorithm 14)
This command is analyzed and rewritten by SIREN following the steps shown in Figure 4. Initially,
the application program submits the SQL command. The command interpreter analyzes the original
command and detects that it contains a similarity predicate refering to a query image that is not stored
in the DBMS (Step 1). The interpreter also identifies the type of similarity operation that needs to be
executed (a
kNN
query in the case), the multimedia attribute involved (Picture) and the parameters of
the predicate (the query center sq = 'example.jpg', the number of neighbors
k
=5 and the metric that is
the attribute's default). Thereafter, it queries the SIREN data dictionary (Step 2). The data dictionary is
searched to obtain information regarding the complex attribute Picture: the attribute's default metric
(Texture in the example), the feature extractors that are employed by the metric (the TextureEXT extrac-
tor), the distance function to be used (), and the index structure
Ix
to be employed (Step 3). The query
image
s
q
is submitted to the required feature extractors (Step 4) and the extracted feature vector
V
is
Algorithm 14.
SELECT * FROM Landscapes
WHERE Picture NEAR 'example.jpg'
STOP AFTER 5;
