Biomedical Engineering Reference
In-Depth Information
OR
s
u
si
;
v
i
¼
(4.19)
where
v
i
is the
i
component of the vector
V
,
u
si
is the
i
component of the vector
U
*
s
,
which corresponds to the detected feature
F
s
, and
OR
s
is disjunction. This coding
process produces almost independent representations of all features because we use
independent random numbers for the feature mask generation. The weak influence
of one feature on the other appears only from the absorption of 1s in disjunction
(Equation 4.19).
To recognize the images, it is necessary to use feature combinations. For
example, the feature
F
a
is present in the image, but the feature
F
b
is absent. To
take into account such feature combinations, Context Dependent Thinning (CDT)
was proposed [
41
]. CDT was developed on the basis of vector normalization
procedures [
48
]. There are different procedures of CDT implementation, but here
we use the following procedure. The new permutation pattern (Fig.
4.10
), which is
independent of the
X
and
Y
permutations, is generated.
After that, we test each component
v
i
of the vector
V
.If
v
i
¼
0, nothing must be
done. If
v
i
¼
1, we consider the trajectory of this component during permutations
(according to arrows). If this trajectory contains at least one “1,” the value of
v
i
is
converted to 0. For example, in Fig.
4.10
, the trajectory of the
v
3
component is
v
4
!
0. The number of the permutations
Q
in CDT is the parameter of the recognition system. In our experiments, we used
Q
from 5 to 15. After the realization of CDT, the binary vector
V
is prepared for the
neural classifier.
v
8
!
v
7
.If
v
4
or
v
8
or
v
7
equals 1, we put
v
3
¼
v1
v1
v1
v1
v2
v2
v2
v2
v3
v3
v3
v3
v4
v4
v4
v4
v5
v5
v5
v5
v6
v6
v6
v6
v7
v7
v7
v7
v8
v8
v8
v8
Fig. 4.10 Permutation pattern for Context Dependent Thinning
