Biomedical Engineering Reference
In-Depth Information
texture in this window. The histogram of brightnesses gives as many texture
parameters as the number of intervals into which the entire range of brightness
was divided. Sometimes it is useful to add to these parameters an estimate of the
mathematical expectation and dispersion of brightnesses in the window.
The histogram of contrasts is similar to the histogram of brightnesses, only
instead of brightness, the difference in the brightnesses between adjacent pixels
must be taken. In the histogram of the orientations of contour elements, the role of
brightness plays the orientation angle of every contour element. The latter is defined
as a vector, perpendicular to the gradient of brightness, determined by four adjacent
pixels. The algorithm that helps us determine the orientation angle is developed
using the Schwarz method [
8
]. The orientation angle of the contour element is
considered in the histogram only when the value of the gradient exceeds a certain
threshold.
In general, few textures have only a small number of sufficiently clear distin-
guishing features. Most frequently, it is necessary to take into account the
complex combinations of a large quantity of features during texture recognition.
In our case, each component of each histogram is considered a separate textural
feature. Depending on the quantity of intervals in each histogram, the total
quantity of textural features can vary from several to several hundred. This feature
space can seem too large, but if we want to obtain a sufficiently universal system
of technical vision, we must take into account the fact that, a priori, we do not
know what features will be more important during the recognition of concrete
textures. It is necessary to create a recognition system in which the excess of
parameters does not interfere with recognition and does not greatly decrease the
operating speed.
To enumerate all the features described above, it is possible to propose very
simple and inexpensive hardware, which will perform this procedure in real time.
We will consider that such hardware exists in the neurocomputer system of technical
vision.
6.1.2 The Coding of Texture Features
The code of the texture sample is assembled from the obtained codes of feature
names and the codes of the numerical values of these features as follows:
k
M
np
i
M
i
Þ
Z
¼
U
1
ð
&
(6.1)
i
¼
where
k
is the number of features;
M
i
np
is the mask of the feature name; and
M
i
j
is
the mask of the number corresponding to the numerical value (
j
) of the feature
i
.
The vector
Z
is supplied for the training of the associative-projective neural
network.
