Applications of Emergent Phenomena - Systematic Design for Emergence in Cellular Nonlinear Networks - page 136

The values of the feature vector components (ranging each between 0 and 1) can
be calculated as follows:
The scaled feature

ms

(aspect ratio of the box) was added, where f

7

sx / sy /

sx ( ix )

. ms max ( )

sy ( ix )

ix

The other six features were slightly scaled to face the situation of hand-
written characters, which are drawn with the same thickness independently of

their size. For instance instead of . Here f Nb / O sqrt ( N ) f Nb / N

N 1 1 1 1 1 1

represents the total number of pixels in a given sub-block, Nb is the number of

“black” pixels within that sub-block and O is a coefficient to be optimized such
that the maximum value of the feature will reach 1. By doing so, the rate of mis-
classification was reduced almost four times in the case of handwritten characters.

1

After the feature extraction, a classifier may be called to categorize the con-

tent of the corresponding character box. For the purpose of evaluation here we
consider RBF-type (Gaussian) SVMs [68]). However, similar results are expected
to be obtained with VLSI-tailored neural systems [69, 70] which are more conven-
ient from the implementation point of view.

8.2.4 Experimental Results

Here a real world problem is considered. A number of 143 digits were drawn on a
sheet of paper as shown in Fig. 8.6. The image was acquired with a digital camera
and it was binarized using a threshold adapted to reduce the noise.

Fig. 8.6. A real world problem for the recognition of handwritten digits. Left: the input image,
right - the segmented image after 50 CA iterations