Biomedical Engineering Reference
In-Depth Information
it is necessary to make its permutations
E
(
X
) times and to make an additional
permutation for
P
x
components of the vector. After that, it is necessary to shift the
code to the vertical direction, making the permutations
E
(
Y
) times and an additional
permutation for
P
y
components.
9.3.3 Neural Classifier
The structure of the proposed recognition system is presented in Fig.
9.7
. The
system contains sensor layer
S
, feature extractor, encoder, associative neural layer
A
, and reaction neural layer
R
. In the screw shape recognition task, each neuron of
the
R
-layer corresponds to one of the screw groups from the database. The sensor
layer
S
corresponds to the initial image.
The associative neural layer contains “binary” neurons having the outputs 0 or 1.
The values of its neuron outputs are produced as a result of the encoder's work. The
neurons of the associative layer
A
are connected to the reaction layer
R
with
trainable connections having the weights
w
ji
. The excitations of the
R
-layer neurons
are calculated in accordance with the equation:
X
n
E
i
¼
a
j
w
ji
;
(9.9)
j¼
1
where
E
i
is the excitation of the
i
th neuron of the
R
-layer;
a
j
is the excitation of
the
j
th neuron of the
A
-layer; and
w
ji
is the weight of the connection between the
j
th neuron of the
A
-layer and the
i
th neuron of the
R
-layer. The neuron-winner
having maximal excitation is selected after calculating the excitations.
We use the following training procedure. Denote the neuron-winner number
as
i
w
, and the number of the neuron, which really corresponds to the input image, as
i
c
.If
i
w
¼
i
c
, nothing must be done. If
i
w
6¼
i
c
ðÞ
8
j
w
ji
c
t
ð
þ
1
Þ¼
w
ji
c
t
ðÞþ
a
j
ðÞ
8
j
w
ji
w
t
ð
þ
1
Þ¼
w
ji
w
t
ðÞ
a
j
(9.10)
w
ji
w
t
if
w
ji
w
t
ð
þ
1
Þ <
0
ð
þ
1
Þ¼
0
;
v
1
v
2
0
Feature
Extractor
Encoder
v
3
4
S
R
v
4
Fig. 9.7 Permutative Coding
Neural Classifier
A
