Recognition of Meter Values - Hierarchical Neural Networks for Image Interpretation - page 146

3.5

flat classifier
hierarchical classifier

ρ recognized substituted rejected

3

0 1077 (98.0%) 22 (2.0%) 0

2.5

0.25 1061 (96.5%) 11 (1.0%) 27 (2.4%)

2

0.62 1023 (93.1%) 6 (0.54%) 70 (6.4%)

1.5

0.70 979 (89.1%) 2 (0.18%) 118 (10.7%)

1

0.73 927 (84.3%) 1 (0.09%) 171 (15.6%)

0.5

0.87 816 (74.2%) 0 283 (25.8%)

0

0 5 10 15 20 25

(a) (b)

Fig. 7.12. Performance of the meter value block classifier on the test set: (a) flat classifier
compared to hierarchical classifier; (b) recognition as a function of the reject parameter ρ for
the hierarchical classifier.

rejects %

For comparison, several fully connected three-layered feed-forward neural net-
works with sigmoidal activation functions were trained on the same data. The net-
works had 16, 32, 64, 128, or 256 hidden units. They were trained using gradient
descent on the squared output error until the test set performance did not improve
any more. The recognition performance of the best flat network, which had 32 hid-
den units, is also shown in Figure 7.12(a). It substitutes 35 (3.18%) of the 1,099 test
examples in the zero-reject case. For higher reject rates the flat network is outper-
formed by the hierarchical network as well.
Since the rejections necessary for reliable recognition reduce the acceptance rate
of the classifier, the next section describes a second recognition system that tries to
verify the examples rejected by the hierarchical block classifier.

7.5 Digit Recognition

Because the block recognition system described in the previous section is not a
perfect classifier, it is complemented by a digit recognition system as illustrated
in Figure 7.13. A separate digit classifier is used for the left and the right digit of
interest since they have different a-priori class distributions and are embedded in
different context. Both digit classifiers receive the output of the block classifier for
the other digit as contextual input in addition to the preprocessed digit.
The digit recognizers are queried only if the block classifier is not confident
enough and rejects an example. Digit recognition consists of three steps: digit pre-
processing, digit classification, and combination of the digit outputs with the results
of the block classifier.

7.5.1 Digit Preprocessing

The image of the preprocessed meter value cannot be given directly to the digit
classifier. Some digit-specific preprocessing is necessary to facilitate recognition.