Biomedical Engineering Reference
In-Depth Information
ν
1
ν
2
0
Feature
Extractor
Encoder
ν
3
9
S
R
ν
N
A
Fig. 4.4 Structure of the Permutation Coding Neural Classifier
produces an output binary vector of large dimension, which is presented to the one-
layer neural classifier input. The output of the classifier gives the recognized class.
4.5.2 Feature extractor
In Fig.
4.5
a
, a grayscale image of the digit “0” is shown. The feature extractor
begins the work by selecting the points of interest in the image. In principle, various
methods of selecting these points could be proposed. For example, the contour
points could be selected as the points of interest. For handwritten digit recognition,
we selected the points of interest
P
ij
, which have the brightness
b
ij
higher than the
predetermined threshold
B.
These points correspond to the white area in Fig.
4.5b
.
The rectangle of area
h
*
w
is formed around each point of interest (Fig.
4.6
).
Multiple features are extracted from the image in the rectangle. The
p
positive
and the
n
negative points determine each feature. The positive point is the point of
the
S
-layer connected to the ON neuron (Fig.
4.2
). The negative point is the point of
the
S
-layer connected to the OFF neuron (Fig.
4.2
).
These positive and negative points are randomly distributed in the rectangle.
Each point
P
mk
has the threshold
T
mk
that is randomly selected from the range:
T
min
T
mk
T
max
(4.3)
The feature exists in the rectangle if all its ON and OFF neurons (Fig.
4.2
) are
active. In other cases, the feature is absent in the rectangle. We used a large number
of different features
F
i
(
i
¼
1,
,
S
). In the final experiments, we worked with
p
¼
...
3;
n
11;
T
min
=1;
T
max
= 254; and
B
= 127. The feature extractor
examines all
S
features for each point of interest. In the handwritten digit image, the
number of the
S
features was changed from 1,600 to 12,800.
¼
6;
h
¼
w
¼
