Biomedical Engineering Reference
In-Depth Information
4. If the condition (4.11) is true, then the procedure is complete; otherwise, go to
step 5.
5.
i
=
i
+ 1. Go to step 2.
This procedure satisfies all requirements of feature coding. In the initial codes of
the features, the unit element can appear with the probability determined during the
mask formation. Since the feature masks are formed independently of each other,
all the features in the assembled nonnormalized code have an approximately equal
number of representatives (unit elements). The vector
Y
produced in step 1 of the
procedure is the inversion of vector
X
; therefore, it has “0s” in the places where
X
has “1s”. Each “0” in the shifted vector
Y
could eliminate a “1” from vector
X
. The
probability that this unit belongs to any of the initial features is constant; therefore,
a quantity of unit elements that will be preserved by different features should be
approximately identical for all features. Consequently, the first requirement is
fulfilled.
Note that different
Y
vectors produce partial extinction of unit elements. Thus,
the feature that falls in different feature combinations loses different unit elements,
fulfilling the second requirement. It is also obvious that during repeated coding of
the same set of features, the same representatives remain in each feature because
vector
Y
contains the same initial components, and constants of the shift of this
vector are also the same (they represent the characteristic of the corresponding
buffer field). It is possible to see also that during coding of the similar feature sets,
many common representatives remain in each feature because the vectors
Y
are
strongly correlated for similar sets. This leads to the large correlation of elements
removed from the code of each feature, and therefore to the correlation of the
remaining representatives of each feature.
It is not difficult to build the normalization procedures taking into account the
priorities of different features on the basis of the described procedure. It is possible
to form the feature set while fulfilling the normalization procedure, joining some
features after this procedure is partially executed. Such features will preserve more
representatives than those that were located in the set from the very beginning. It is
possible to build the procedures of the formation and normalization of the feature
set in such a way that the representatives of features will enter into the normalized
set with practically any given weights.
Local connected coding was used to solve pattern recognition problems. The
results of applying this coding will be described. However, local connected coding
has specific deficiencies. One drawback of local connected coding during image
recognition is that the identical elements of patterns located in different image
places are represented by different stochastic codes, practically not connected with
each other. This prevents the formation of the neural ensembles that present the
separate elements of patterns and makes pattern recognition difficult when the
object in the image lies in different positions. To eliminate these deficiencies, a
new mechanism of neural coding is proposed. In this method, the name of the
parameter is represented by the binary stochastic code, and its numerical value is
represented by the shift of this code. We will identify this coding as shift coding.
