Databases Reference
In-Depth Information
An advantage of the recall-percentage implementation for recognition at the
pattern level is the high recall value precision, in terms of the percentages of
pattern indices being recalled. For a given input pattern, the DHGN is able to
provide a precise recall value. The DHGN also has the ability to use previous
input patterns that have been stored in the network to analyze the pattern
composition.
The recall-percentage method also comes with a number of limitations.
These include its effect on the DHGN recognition accuracy, as described in
Section 4.3.2. The nature of the DHGN recognition process implies that a
slight change in the structure of the subpattern will affect the index calculation
of the entire subnet.
The recall-percentage method also raises issues regarding the level of con-
fidence in the outputs of the system. For instance, assuming a recognition
output of a pattern obtained from a DHGN network consists of three pat-
terns previously stored — P
1
: 0.4; P
2
: 0.3; P
3
: 0.3; — the result of this
recognition will favor P
1
as the recalled pattern. However, criteria of P
2
and
P
3
have also been detected in the pattern. Therefore, there is a need to es-
tablish a level of confidence for this type of result from the perspective of
recognition.
5.1.3.2.2 Voting method
Most of the existing pattern recognition
schemes apply rejection techniques to remove highly distorted patterns from
its classification procedure. This technique adopts the rejection/accuracy rate
as a parameter to indicate levels of similarity between patterns. The technique
offers a precise means to obtain a good classification measurement. However, it
is most suitable for deployment in a single-decision system, i.e., the classifica-
tion is conducted using a single classifier/recognizer. A decision-making mech-
anism is needed to combine all of the decisions (in terms of accuracy/rejection)
made by each of the classifiers.
One possible method for combining decisions on classification is the voting
method. There are several forms of voting available in the literature. These
include majority, common-consent, unison, and unanimity voting [69, 70]. In
a DHGN implementation, majority voting is used as a means to obtain a com-
bined decision on the recalls made by each of the subnets within a recognition
network.
For each recognition process, whether the input pattern has been recognized
(i.e., recall) or is new to the network (i.e., store) is decided by obtaining
the majority consent from all of the DHGN subnets. For a pattern to be
recalled, the network should confirm that most of the subpatterns belong to
the respective input pattern. The adopted majority voting concept follows the
work by Cruz, Sossa and Barron [71] and is described in [4].
In this pattern reconstruction and recognition process, the SI module will
initially receive all of the results from the recognition at the subpattern level
in the form of signal messages from the DHGN subnets. After all of these
