Recognition of Meter Values - Hierarchical Neural Networks for Image Interpretation - page 138

same time, the lines of meter value should be enhanced to improve readability. The
filtering methods described in the following were developed from a low-level model
of the print that includes line width, line color, ratio of foreground to background
area, and typical noise.
Color Filtering. Because the meter stamps are printed with red ink, it is obvious to
use the color as a key to segment the print from the background. The RGB-images
were captured with a two-line camera. One sensor line consists of only green pixels,
while in the other line blue and red pixels alternate. Thus, the vertical resolution of
the green color is twice as high as the resolution for the other two color channels.
Since the lines of the print are only about one pixel wide, frequently they cover
only parts of a pixel's sensor area. This leads to color deviations. The red lines then
appear to be orange or magenta.
Figure 7.4(a) shows the original rectangular regions of three examples from the
Swedish Post database. Parts (b-d) display the three 8-bit RGB-color components of
these images. One can observe that the print is best visible in the green component,
as this is the complementary color to red. A lower contrast blurred version of the
print appears in the blue component. The red component contains almost no differ-
ences between the print and the background since the reflectivity for red light of the
red color is about as high as the one of the paper.
A pixel-based filter extracts the red colored lines as follows:

v = min(255 , max(0 , 24 + 2 r + 0 . 125 b − 2 . 25 g )) ,

(7.1)

(a)

(b)

(c)

(d)

(e)

(f)
Fig. 7.4. Color filtering: (a) original image (converted to grayscale); (b-d) red, green, and blue
components of original image; (e) red color filtered version of the input (shown inverted); (f)
output image produced using center-surround filtered green component combined with red
neighborhood mask.