Recognition of Meter Values - Hierarchical Neural Networks for Image Interpretation - page 149

(a) (b)
Fig. 7.15. Digit preprocessing for problematic meter value examples: (a) successful digit
segmentation; (b) segmentation failure.

connected to the 120 input pixels. Ten context units, which are set to the outputs of
the block classifier for the other digit, are also located in the second layer. The third
layer consists of 10 output units that are fully connected to these context units as
well as to the hidden units. They signal the digit's class in a 1-out-of-10 encoding.
The hidden units and the outputs are

P
-units that compute a weighted sum of their
inputs, followed by a sigmoidal transfer function.
The network is trained to produce the desired outputs using gradient descent on
the squared output error. Training is done until the performance on the test set does
not improve any more. A separate classifier is trained for the left and the right digit.
After training, the performance on the training set is almost perfect. While all
left digits of the 4,351 training examples can be recognized, only two of the right
digits are substituted. They can be rejected easily using the difference between the
activities of the most active and the second most active output as confidence.