Biomedical Engineering Reference
In-Depth Information
where
w
ji
(
t
) is the weight of the connection between the
j
neuron of the
A
-layer and
the
i
neuron of the
R
-layer before reinforcement, and
w
ji
(
t
+ 1) is the weight after
reinforcement.
9.4 Results
To examine the PCNC's recognition of shapes of micromechanical workpieces, we
have produced 40 screws of 3 mm diameter with a CNC lathe from the “Boxford”
company. Ten screws were produced with correct positioning of the thread-cutting
cutter (Fig.
9.4b
). 30 screws were produced with erroneous positioning of this
cutter. Ten of them (Fig.
9.4a
) had a distance of 0.1 mm less than necessary between
the cutter and screw axis. Ten screws (Fig.
9.4c
) were produced with a distance of
0.1 mm larger than necessary, and the remaining ten (Fig.
9.4d
) with a distance of
0.2 mm larger than necessary. We have made the image database from these screws
using a Samsung Web-camera mounted on an optical microscope.
Five images from each group of screws were selected randomly and used for
neural classifier training, and another five were used for neural classifier examina-
tion. The experiments were conducted with different values of parameter
B
for
specific point selection. In Table
9.1
, the first column corresponds to parameter
B
.
Four different runs were made for each value of
B
to obtain statistically reliable
results. Each run differs from the others by the set of samples selected randomly for
neural classifier training and by the permutation scheme structure. For this reason,
we obtained different error rates in different runs. The second column contains the
error number for each run. The third column gives the mean value of the error
number for each
B
value. The fourth column contains the mean percentage of
correct recognition.
A new neural classifier, PCNC, was developed to recognize different types of
images. It showed good results on the MNIST and ORL databases. In this chapter,
Table 9.1 Results of system investigation
Threshold for specific
point selection
Error number
(four runs)
Average number
of errors
% of correct
recognition
20
3
3
85
4
1
3
40
2
2.25
88.75
2
2
3
60
3
1.5
92.5
1
1
1
