Information Technology Reference
In-Depth Information
(a)
(b)
(c)
Fig. 7.14.
Segmentation and size normalization of meter value digits: (a) digit block with
occupancy index and segmentation of the left and the right digit of interest; (b) segmented
regions interpolated to higher resolution; (c) digits normalized to 8
×
15 pixels.
Because the segmented region usually contains fewer pixels than the normalized
digit, the resolution of the image is doubled in a first step by inserting interpolated
rows and columns. The value
v
of an interpolated pixel is computed from the values
v
1
and
v
2
of its neighbors as follows:
v
= (
v
1
+
v
2
+ max(
v
1
,v
2
))
/
3
. This makes
pixels next to the dark lines darker than simple averaging. The higher resolution
variants of the digits are shown in Figure 7.14(b). The values of the normalized
digits are set now to the response of a 3
×
3 binomial filter at the corresponding
position in the high resolution image.
Finally, the contrast of the normalized image is increased slightly by multiplying
the intensities with
1
.
25
, subtracting one fourth of the average pixel value, and clip-
ping the values to the interval
[0
,
1]
. This darkens the lines and sets the background
pixels to exactly zero (white). Figure 7.14(c) shows the digits after normalization
and contrast enhancement. Preprocessing worked well for these examples because
the digits were tightly framed by the segmentation and were clearly readable.
Some more problematic examples are shown in Figure 7.15. Part (a) of the figure
contains examples that were segmented successfully, although segmentation was not
easy. Note that even if a digit is broken into parts, these parts are grouped together
and placed at the correct position in the normalized digit image. Part (b) shows
some examples for which segmentation failed to select the two digits of interest.
This may have been caused by additional foreground structures, as in the example
in the second row. Another problem is the miscentering of meter values that leads
to the selection of the wrong digit, as in the first and the third row of the figure.
7.5.2 Digit Classification
A three-layered feed-forward neural network is used for digit classification. It is
sketched in Figure 7.16. The network's input layer contains the normalized digit to
recognize. In the second layer, 32 hidden units detect digit features. They are fully
Search WWH ::
Custom Search