Information Technology Reference
In-Depth Information
Fig. 7.3.
Some examples of the Swedish Post database (converted to grayscale). While some
meter values are clearly readable, others are challenging even for experienced human ob-
servers.
Table 7.1.
Most frequent meter values of the Swedish Post database. The 16 values shown
account for 99.2% of the dataset.
4.60
4.40
9.20
3.90
4.10
3.80
8.80
7.00
3.50
6.00
8.20
5.00
13.00
12.00
7.80
8.00
#
3176
534
522
341
319
98
91
82
78
64
42
19
18
16
14
14
%
58.1
9.8
9.5
6.2
5.8
1.8
1.7
1.5
1.4
1.2
.77
.35
.33
.29
.26
.26
thermore, one can observe that it suffices to read the two digits next to the point
separator in order to uniquely identify the meter value.
In addition to the RGB-image and the meter value, an automatically determined
rectangular region is given for each example that should contain the digits belonging
to the meter value and nothing else.
7.3 Preprocessing
Before an example can be given to the block recognizer, some preprocessing is
needed to make its task easier. The goal of the preprocessing is to reduce the variance
of the examples by color filtering and by increasing the image contrast, such that the
print becomes black and the background becomes white, and by normalizing slant
and position of the meter value.
7.3.1 Filtering
Segmenting the print from the background is not an easy task, since the image qual-
ity of the Swedish Post database is quite low. Noise should be discarded, while at the
Search WWH ::
Custom Search