Recognition of Meter Values - Hierarchical Neural Networks for Image Interpretation - page 136

(a) (b) (c)
Fig. 7.2. Pitney Bowes Model M postage meter: (a) drawing by inventor Arthur H. Pitney; (b)
original postage meter from circa 1920; (c) original Model A mailing machine, which housed
the postage meter (pictures adapted from [10]).
tered, representing the largest single source of postage revenue. More than a million
postage meters are being used by large and small quantity U. S. mailers in business,
industry, and other categories.
During automatic mail processing, not only the letter's address is read for sort-
ing, but also the stamps are recognized. This is necessary in order to compare the
postage value of the stamp to the weight of the letter for checking if the postage is
sufficient. It is desirable to apply the same check to metered letters.
For a successful recognition of meter values, first, the meter stamps must be
detected. Next, the exact location of the meter value must be determined. Finally,
the value must be read.
In the following, a system is described that covers only the last step, the ac-
tual recognition of the isolated meter value. The system is trainable to recognize
the entire meter value, without prior digit segmentation. It is based on the Neural
Abstraction Pyramid architecture. If this block classifier cannot make a confident
decision, single digit classifiers are combined with its results.
7.2 Swedish Post Database
For the following experiments, a database of Swedish Post meter marks is used
which was collected by Siemens ElectroCom Postautomation GmbH. It contains
5,471 examples that were assigned randomly to a training set of size 4,372 and a
test set of size 1,099. Figure 7.3 shows some example images from this dataset. As
can be seen, the recognition of the meter value is not an easy task. The images are
of low resolution and low contrast. Typically, digits have a size of only 10 × 4 pixels.
High variance of print, lighting, and background complicate recognition further.
Frequently, the meter values are difficult to read even for humans.
On the other hand, the meter values are not arbitrary combinations of digits, but
come from a set of standard postage values. Table 7.1 shows the 16 most frequent
values that account for 99.2% of the dataset. The meter values are not uniformly
distributed. The five most frequent values cover almost 90% of all examples. Fur-