Biomedical Engineering Reference
In-Depth Information
weights should be determined by the recognition system designer and are not
modifiable during training. In the simplest cases, the neuron numbered 1 can extract
the contour points. For this case, the weights of excitatory connections must be
inversely proportional to the area of the E-rectangle and the weights of the inhibitory
connections must be inversely proportional to the area of the I-rectangle. Another
simple case corresponds to extracting the bright areas of the image. For this case, the
weights of the excitatory connections should be inversely proportional to the area of
the E-rectangle and all the inhibitory connections should be “0.”We use the first case
for face recognition problems and micromechanical applications of the proposed
system; we use the second case for handwritten digit recognition. The output of the
neuron numbered 1 is “1” if the algebraic sum of the input signals is positive, and the
output is “0” if the sum is negative. We assume that the complex neurons will detect
the most informative points of the image. The neuron descriptor
D
has output “1” if
all the neurons 1-5 have output “1.” If at least one of these neurons has output “0,”
the output of neuron
D
is “0.” The neuron
D
can be termed AND neuron.
The neural layer
D
1
(Fig.
4.1
) consists of a large number of planes,
d
11
,
d
12
,
,
d
1M
. Each plane contains the number of AND neurons equal to the pixel number of
the input image. The plane
d
1j
preserves the topology of the input image, i.e., each
neuron of plane
d
1ji
corresponds to the pixel located in the center of the
corresponding rectangle of dimension
W
x H (Fig.
4.2
). The topology of connec-
tions between the sensor layer and neurons 2-5 (Fig.
4.2
) is the same in the range of
every plane
d
1j
(Fig.
4.1
). The topology of connections between the sensor layer and
the neuron numbered 1 is the same for all the neurons in all the planes of layer
D
1
.
The aim of each plane is to detect the presence of one concrete feature in any place
of the image. The number of planes corresponds to the number of extracted features
(in our system, each feature corresponds to one descriptor type). The large number
of features permits us to obtain a good description of the image under recognition.
In the MNIST database we used 12,800 features; in the ORL database we used 200
features. To estimate the required number of features, it is necessary to solve the
problem of structure risk. It is difficult to obtain an analytical solution to this
problem for such a complex recognition system as we propose, so we estimate
the required number of features experimentally for each recognition problem.
Layer
D
2
(Fig.
4.1
) also contains
M
planes of neurons. Each neuron of the
d
2j
plane is connected to all the neurons of the
d
1j
plane located within the rectangle.
The output of each neuron is “1” if at least one of the connected
d
1j
neurons has the
output “1.” Such a neuron is termed an OR neuron. The topology of the
d
2j
neurons
corresponds to the topology of the
d
1j
neurons and to the topology of the
S
layer. We
shall term all neurons having output “1” the active neurons.
The associative layer
A
contains
N
associative elements, which collect the
activities of
D
2
neurons selected randomly. The other function of the associative
elements is to reduce the number of active neurons up to the predetermined value.
We refer to this function of the associative layer as normalization. This layer also
performs the binding function [
40
,
41
]. The structure of the three associative
elements
a
1
,
a
2
, and
a
3
is presented in Fig.
4.3
in the form of electronic circuits,
but it could be implemented as a neural network.
...
