Databases Reference
In-Depth Information
calculated. The error per test object, P
err
, for a given number of test objects
o
t
is calculated for each feature using the following equation:
P
err
=
F
+ve
+ F
−ve
o
t
(7.1)
where F
+ve
and F
−ve
represent the number of false positives and negatives,
respectively.
The following scenario illustrates recognition accuracy calculations based
on error values. There are a series of patterns, P , containing n classes, P =
{p
1
, p
2
, . . . , p
n
}, and a set of features, F = {f
1
, f
2
, . . . , f
m
}. For each pattern
class, p
x
, x = 1, 2, . . . , n, select the feature, f
y
, y = 1, 2, . . . , m that minimizes
the recognition error, err
f
x
, for all test patterns. The recall accuracy, r
p
x
, for
each pattern class is derived using the following equation:
r
p
x
= arg min{err
f
1
: err
f
m
},
x = 1, 2, . . . , n
(7.2)
Note that the minimum error is not the only parameter that can be used to
determine the most effective recall accuracy for multi-feature pattern recogni-
tion. Other parameters include the normal mean, median, standard deviation,
and other statistical estimations, such as Bayes and maximum-likelihood es-
timators.
7.2.2 Complexity Estimation
In multi-feature recognition using the DHGN distributed pattern recog-
nition (DPR) scheme, the approach applied to recognizing features in each
pattern is similar to the original DHGN implementation described in Chapter
5. Therefore, the complexity of the basic recognition function (for recognition
at the subpattern level) is as low as the originally proposed scheme. However,
in the multi-feature scheme, the voting mechanism is applied at two levels,
i.e., at the SI module and the coordinator nodes.
At the SI module node, voting determines the matched pattern class for a
given pattern. At the coordinator node, voting selects the feature that gives
the optimal value for the specified accuracy parameter.
7.2.2.1
Voting Scheme at the SI Module
The voting scheme applied at the SI module assigns the test pattern into
a specific pattern class, based on a similar characteristic or feature value.
Inputs to this voting process are the indices retrieved from all of the DHGN
subnets. Each SI module handles a specific feature for a particular dataset. The
maximum voting scheme in this DHGN implementation finds the maximum
number of similar indices returned from the subnets. The voting scheme has
two stages, namely vote counting and maximum vote search.
