Information Technology Reference
In-Depth Information
13.5.3 Simulations
A large scale experimental study has taken place in order to extract useful conclu-
sions towards the improvement of the moment features' recognition performance
through the application of a selection mechanism. The Genetic Algorithm settings
are: population size 50, maximum generations 30, crossover with probability 0.8 and
2 points, mutation probability 0.01 and Stochastic Universal Approximation (SUS)
selection method. The k-NN classifier (k
1) is selected as the prediction model in
the case of the GA-based selection. Moreover, a 10-fold cross-validation technique
is applied in all datasets, while the moments are selected from a pool of the first 100
(up to 9 order for all moments, except Zernike computed up to 18 order) computed
moments for each moment family.
The corresponding mean recognition rates for each dataset are summarized in
Tables 13.2 , 13.3 , 13.4 and 13.5 . The best moment family alongwith the best selection
method is presented in these results.
By examining Tables 13.2 , 13.3 , 13.4 and 13.5 it is deduced that in almost all
cases the selected moment features show better or equal, in the worst case, perfor-
mance than the non selected (NoSel.) moments. This outcome enforces the initial
assertion regarding the needs for moment features selection. Moreover, among the
two examined selection methods, the GA-based one seems to be more efficient for
small sized moment subsets (up to 25-30), while the Relief algorithm is superior for
larger subsets (greater than 30-40). This observation can be justified by the fact that
for large moment subsets (greater than 30) the optimization problem, which needs
to be solved by the GA, is quite difficult. One solution to this limitation is to use
more advanced versions of the algorithm, where adaptive crossover and/or mutation
operators could guide the algorithm to more optimum solutions.
=
Table 13.2 Recognition performance of moment features subsets for the Yale dataset
Yale dataset
Number of moments
Moment type
Recognition rate (%)
Selection method
5
Krawtchouk
76.00
GA
10
Krawtchouk
86.66
GA
15
Zernike
88.00
GA
20
Zernike
87.33
GA
25
Dual Hahn
84.66
GA
30
Krawtchouk
72.66
Relief
40
Krawtchouk
75.33
Relief
50
Krawtchouk
74.66
Relief
60
Krawtchouk
76.66
Relief
70
Dual Hahn
76.00
Relief
 
 
Search WWH ::




Custom Search