Biomedical Engineering Reference
In-Depth Information
appeared. If the error percentage is less than 2-3%, it is difficult to compare the
classifiers using this mode.
The second mode permits us to conduct many different experiments with the
same database and to obtain statistically reliable results. In our experiments, we use
ten runs for each experiment. This method was used also in other works for
estimating the mean value of error rate [
11
]. They used six runs for each experi-
ment. Their results are given in Table
4.3
.
It is also interesting to investigate the cases where less than five images are
selected for training and the rest for testing. The data from such experiments are
presented in [
10
]. Unfortunately, the data are presented graphically and no table
is given. We tried to restore the data from the graphics. The results are given in
Table
4.4
:
The results of our experiments are presented in Table
4.5
:
The note tr./ex. reflects how many images were used for training (tr.) and how
many for examination (ex.). The comparison of results shows that our classifier, in
most of the cases, gives the best recognition rate. In this task, we used the classifier
that has
N
64,000;
S
= 200. The parameter
T
E
was 0.5. We mentioned above that
different methods of detecting points of interest could be used. For face recognition,
we used the points of the contours. For this purpose, the connections from the
E-rectangle (Fig.
4.2
) must have positive weights, and the connections from the
ΒΌ
Table 4.3 Best performances on 6 simulations
Simulation
1
2
3
4
5
6
NOM 2
6
7
2
1
6
(NOM - number of misclassifications or errors)
Table 4.4 Error rates for
different number of each
person's image presented for
training
Number of
training images
The best results from [
10
]
(restored from graphics)
Our
results
1
17.6
16.1
2
8.8
7.09
3
4.8
2.15
4
2.8
1.4
5
1.2
0.1
Table 4.5 Recognition results for ORL database
tx/ ex.
NOM
Total NOM E(%)
Run
1
2
3
4
5
6
7
8
9
10
1/9
64
66
55
45
52
63
72
44
58
60
580
16.1
2/8
14
29
18
18
25
26
19
13
21
44
227
7.09
3/7
6
13
6
0
9
8
10
3
4
1
60
2.15
4/6
3
3
7
2
5
1
0
7
4
2
34
1.4
5/5
0
0
0
0
0
0
0
0
0
2
2
0.1
