Page 151 - Computational Colour Science Using MATLAB

138 CHARACTERIZATION OF CAMERAS

Figure 8.3 Effect of number of terms (m) in the polynomial model on training and testing
performance (median colour difference)

It is evident that the performance of the best neural network and polynomial
models produces a test error of about 2 CIELAB units. It is not surprising that
the two systems should provide equivalent performance. Training the neural
networks can be quite time consuming, however, and there are many parameters
to determine, such as the number of hidden units, the transfer functions for the
units in the network, the learning rule, the parameters of the learning rule, and so
on and so forth. Furthermore, each time the network is trained from different
random initial weights a different transform is achieved. The results shown in

Figure 8.4 Effect of number of hidden units (n) in the neural-network model on training and
testing performance (median colour difference)