Recognition of Meter Values - Hierarchical Neural Networks for Image Interpretation - page 142

Fig. 7.8. Preprocessing of meter values. For ten randomly selected examples of the Swedish
Post database, the original region of interest (converted to grayscale) and the result of the
preprocessing are shown.
The second minimum is used to center the 32 × 16 window horizontally. Its ver-
tical center is chosen to be the vertical component of the center of foreground mass,
indicated by the horizontal lines in Figure 7.7(b) for three examples.
Figure 7.7(c) shows the windows cut from the three examples. The two digits
of interest have been centered successfully. Because the other digits, as well as the
pixels near the upper and the lower window edge, are less important for recognition,
image contrast is faded towards the borders of the window. This also reduces border
effects in the block recognizer.
Figure 7.7(d) displays the final result of the preprocessing for the three exam-
ples. In Figure 7.8 the original regions of interest and the result of the preprocessing
are shown for ten randomly selected examples. The two digits of interest are quite
salient, centered, and most of them are readable.
7.4 Block Classification
The task of the block classifier is to recognize a meter value from a preprocessed
image. Although the preprocessing discarded some of the variances present in the
examples and increased their readability, the problem is still challenging.
As can be seen in Figure 7.8, the print varies considerably. The digits come
in different sizes and different fonts with varying spaces between them. Some ex-
amples contain a delimiting point, while others do not. Some of the loops enclose
background pixels, while others only have some brighter foreground pixels in the
center. Furthermore, noise is still present in the images.