Recognition of Meter Values - Hierarchical Neural Networks for Image Interpretation - page 145

4 6 9 ?

9 2 4 6

4 4 9 2

4 6 4 ?

7 0 ? 0

4 1 8 ?

(a) (b)
Fig. 7.11. Block recognition examples from the test set. The inputs to the hierarchical block
classifier, its outputs, and the interpretation of the output are shown for: (a) examples that
were recognized successfully; (b) examples for which recognition failed.

which recognition failed. Analysis of the problematic examples reveals that recogni-
tion failure is mainly caused by missing image parts or by failure to center the digits
of interest during preprocessing. Such errors in digit segmentation occur more often
when the region of interest includes foreground structures in addition to the digits,
as in the example in the second and the third row of the figure.
One can observe that for most ambiguous examples, the network is able to indi-
cate its uncertainty by producing outputs that deviate from the desired 1-out-of-10
pattern. This makes it possible to compute a meaningful classification confidence
as follows. For both digits, the difference between the maximal output activity and
the second largest one is taken as confidence. Since the actives belong to the inter-
val [0,1], the digit confidence has a value of one when a single output is one and
all other outputs have zero activity. If more than one output has high activity or all
output activities are low, the digit confidence has a low value.
Both digit confidences are combined into a block confidence by taking the aver-
age. If one digit confidence is below a reject parameter ρ , or the block confidence is
below the block reject parameter ρ = 1 − (1 − ρ ) 2 which is more strict, the example
is rejected.
Figure 7.12 summarizes the test performance of the hierarchical block classifier.
About 2% of the examples are substituted when the reject parameter ρ = 0 is used
and all outputs are accepted. By rejecting 2.4% of the examples, half of the substi-
tutions can be avoided. To reduce the substitution rate further, a larger fraction of
the examples must be rejected.