Biomedical Engineering Reference
In-Depth Information
Fig. 6.9
Different ways of
writing the letter “
u
”
elements. This coding takes into account the relative position of separate features.
Using traditional methods of visual pattern recognition, there is a problem of
repeated entry of the same feature in the description of an object. If the same
feature is encountered several times, this requires the sorting out of algorithms,
which slows down the recognition rate. Let us examine, for example, the hand-
written letter
u
(Fig.
6.9
).
Figure
6.9
depicts an example of the description of symbol “
u
” in the form of the
features represented with line segments of different orientations (in Fig.
6.9
, only
vertical orientation is shown). Two versions of handwritten characters are given.
Each line segment is characterized by the angle of the slope and position on the
image. We will assume that the slope angle corresponds to the name of the feature
(in the real tasks, we used ten different slope angles, which corresponded to ten
different names). The position of each feature is given by two coordinates of the
center of the corresponding segment.
Let us consider the task of recognition of the handwritten symbols described by
this feature set. Let us assume that the method of potential functions is used for
recognition. For this method, it is necessary to describe each concrete image by the
fixed set of the parameters. In this case, difficulties are caused by the following.
Firstly, some features can be absent in the image of a handwritten symbol; there-
fore, it is necessary to reach an agreement concerning what coordinates are assigned
to the absent feature. Secondly, some features can be present in different quantities,
so the problem here is how to order the set of features. It is very difficult to
recognize and to fix different occurences of the same feature in the handwritten
character description. Figure
6.9
shows an example of this due to the distortion of
the handwritten symbol. In one case, the features 1, 2, 3, and 4 were extracted and in
the other case, 1, 2, and ? were extracted. However, the system does not know the
numeration of similar features, and in general, it is necessary to sort out all com-
binations of the correspondence of the feature numbers and to attempt to estimate
the results of recognition for each combination. But since a quantity of features can
reach several tens, it is practically impossible to sort out all combinations. In order
to avoid this complex problem, it is expedient to use a coding method that would
make it possible to recognize symbols independently of their position on the image.
One appropriate method is shift coding.
