Page 171 - Computational Colour Science Using MATLAB

158 CHARACTERIZATION OF PRINTERS

Figure 9.8 Typical learning behaviour of a neural network trained to map from RGB to
L*a*b*
After training the network performance was tested using the MATLAB
command

poutput = sim(net,input);

which generates the predicted output matrix poutput for the matrix input given
the state of the network net. The target and predicted output values were re-
scaled to the original values of the CIELAB space and the colour difference was
computed for each of the 729 samples. The median CIELAB colour difference
was 3.47 (maximum 12.89). This error is referred to as the training or
memorization error. The testing or generalization error was computed by
using the sim command with the training input matrix containing the 144
samples and the median colour difference was found to be 3.16 (maximum
11.87). There is no reason why, as in this case, the testing error may not be less
than the training error. However, the testing error should always be used as the
indication of the network’s ability to predict output for input vectors that were
not used during the training process.
The number of units in the hidden layer was varied and the above process
repeated. Table 9.1 lists the performances that were obtained. The performance
of the neural network was compared with that of a third-order masking or