Again, random 75% (251 texture images) of the available
data samples are used for training and the remaining 25% (84
texture images) for testing, the difference being that the data
used for SOMs training is statistically pre-processed with
PCA. Similarly to the previous case, the number of training
epochs, the number of neurons, the SOM’s topology, and the
number of PCs used for the training are varied. Each subexperiment
is performed 50 times and the minimal, maximal,
and the mean accuracy for these runs are recorded. The results
are given in Table 7 and Table 8, and illustrated in Fig. 6. The
sample confusion matrix (Table 9) and the sensitivity and
specificity metrics (Table 10) show that the classifier is
having difficulties recognising some of the Corkstone samples,
but its overall performance is very good, having over 90%
accuracy for the other classes.