Biomedical Engineering Reference
In-Depth Information
2 and 4). If we permit a pin-hole position error of up to 0.1 mm, we obtain a rather
high quality of recognition (the right part of Table
11.1
, columns 5 - 8).
11.2 LIRA Neural Classifier for Pin-Hole Position Detection
In this chapter, we will describe the structure of LIRA neural classifier. It includes
three neuron layers (Fig.
11.9
). The
S
-layer corresponds to the input image, and the
A
-layer is a layer of associative neurons. We connect each
A
-layer neuron with
S
-
layer neurons randomly selected from the rectangle (
h
*
w
), which is located in the
S
-layer (Fig.
11.9
). The distances
dw
and
dh
are random numbers selected from the
following ranges:
dw
from [0,
W
S
-
w
] and
dh
from [0,
H
S
-
h
], where
W
S
and
H
S
stand for width and height of the
S
-layer. Very important parameters of this
classifier are the ratios
w/W
S
and
h/H
S
,
which were chosen experimentally. The
connections of the
A
-layer with the
S
-layer do not change during training.
The excitation of the
A
-layer neurons takes place only under the following
condition. Every neuron of the
A
-layer has
m
positive connections and
l
negative
connections with
S
-layer neurons. A positive connection is activated when the
image pixel that corresponds to this neuron has a value of 1. A negative connection
is activated when the image pixel that corresponds to this neuron has a value of 0.
The excitation of the
A
-layer neuron takes place when all
m
positive connections
and
l
negative connections are activated.
The
R
-layer presents the outputs of the classifier. All neurons of this layer are
connected to all neurons of the
A
-layer with trainable connections. The excitation of
R
-layer neurons is defined in accordance with the formula:
X
n
E
j
¼
a
i
w
ij
(11.1)
i
¼
1
Fig. 11.9 LIRA Neural
Classifier
