Biomedical Engineering Reference
In-Depth Information
Table 11.2. The results for
X
coordinate of interpolator investigation
e
for recognition
X
0.025 mm X
0.05 mm X
0.075 mm X
0.1mm X
0.15mm
X
X%
X
X%
X
X%
X
X% X
X%
Error
74
33.6
25
11.4
4
1.8
0
0
0
0
Correct recognition 146
66.4
195
88.6
216
98.2
220
100
220
100
Table 11.3. The results for
Y
coordinate of interpolator investigation
e
for recognition
Y
0.025 mm
Y
0.05 mm
Y
0.075 mm
Y
0.1 mm Y
0.15 mm
Y
Y%
Y
Y% Y
Y%
Y
Y%
Y
Y%
Error
101
45.9
48
20.9
11
5
0.3
0.14
0
0
Correct recognition 119
54.1
172
79.1
209
95
219.7
99.86
220
100
11.4 Discussion
The equipment for automatic assembly of microdevices must have high precision
because the tolerances of microdevice components are very small. To increase
precision, we use feedback based on computer vision principles. We propose an
image recognition method based on neural networks and have developed two types
of neural networks to solve this problem [
4
,
6
,
10
,
11
]. We call the first type a neural
classifier and the second a neural interpolator. They are used to obtain relative
positions of the micropin and microhole. The neural classifier gives a finite number
of relative positions, while the neural interpolator serves as an interpolation system
and gives an infinite number. An image database of relative positions of the
microhole and micropin of diameter 1.2 mm was used to test the neural classifier
and the neural interpolator. In this work, the recognition results of the relative
positions are presented for the neural classifier and the neural interpolator.
A special prototype was made to examine the proposed method. With this
prototype, 441 images were obtained. The distance between neighboring images
corresponds to 0.1 mm displacement of the pin relative to the hole. At the image,
this displacement corresponds to 1.8 pixels in the
X
direction and 1 pixel in the
Y
direction. The experiments show that the computer vision system can recognize
relative pin-hole position with a 0.1 mm tolerance. The neural classifier for this
tolerance gives the correct recognition in 100% of cases for the
X
axis and 86.7% of
cases for the
Y
axis (Table
11.1
). The neural interpolator gives 100% for the
X
axis
(Table
11.2
) and 99.86% for the
Y
axis (Table
11.3
). The neural interpolator also
permits us to obtain data for smaller tolerances. For example, for an
X
axis with a
tolerance of 0.05 mm, it gives an 88.6% recognition rate, and for a
Y
axis with a
tolerance of 0.05 mm, it gives 79.1%. The experiments show that the neural
interpolator gives better results in estimating the pin-hole relative positions.
It is interesting to observe that the 0.05 mm tolerance for the
Y
axis is less than
one pixel in the image. In this case, the recognition rate of 79.1% shows that the
recognition possibilities are not limited by the resolution of one pixel. This result
