Recognition of facial expressions - Knowledge Based Systems ...

numeric information. Usually, a neuron is set to handle values within a given interval, i.e.
[-0.5, +0.5] or [0, 1]. Since there is a limitation in representing numeric values, any other
value that is outside the interval can be mapped to a value in the interval (Figure 10).
Figure 10. Mapping any value so as to be encoded by a single neuron
In this case of encoding, the previous process of discretization applied on the value of the
model parameters is no longer necessary. Every neural unit can manage to keep any value
of one parameter without any other intervention. By using a single neuron for encoding
the value of a parameter, the structure of the ANN becomes simpler and the
computational effort is less. The network’s structure is as presented below:
In the case the network is used for recognizing the presence of Action Units, the output
layer would require 22 neurons, one for each AU. Moreover, in the case the ANN would
be assumed to recognize facial expressions, the output layer would consist only of few
bits, enough for encoding an index to be associated to one of the six different basic facial
expressions.
The second method of encoding parameter value in an ANN is to use discretization
mechanism before and to encode any value by using a small group of neurons.
For a set of 10 parameters, each parameter is encoded as a group of three neurons that is
able to encode a value big enough to represent the maximum number of the classes for
each discrete parameter. The experiments conducted required a number of 5 or 7 distinct
values per parameter. The second layer is that of the hidden neurons. For the experiments
- 46 -