Cultural Heritage Applications: Archaeological Ceramics and Building Restoration - On Statistical Pattern Recognition in Independent Component Analysis Mixture Modelling - page 134

Fig. 6.3 Measurement device employed in ultrasonic signal acquisition. A detail of the

ultrasound transducer case is included

quantization (LVQ), and multilayer perceptron (MLP) [ 16 ]. As well the k-nearest
neighbour (kNN) was tested 0. Table 6.1 shows the overall percentage of classi-
fication accuracy achieved by the different Mixca variants.
Table 6.2 shows the overall percentage of classification accuracy achieved by
the other different methods implemented. Note that different values of the fitting
variables required in each method (e.g., the value k in kNN) were tested and the
results shown are the best ones obtained.
The best performance in classification was obtained using Mixca at PSS ratio of
1 (total probabilistic supervision), achieving a classification accuracy of 83 %,
which is much better than the rest of supervised methods (LDA, RBF, LVQ, MLP,
kNN). As the PSS ratio is reduced, the performance of Mixca gets worse. How-
ever, for PSS ratio 0.8, Mixca is still the best one with classification accuracy of
79 %. For PSS ratio 0.6, only LDA gives a slightly better result. This confirms the
convenience of not assuming any parametric model of the underlying probability
density as is assumed in LDA and in the parametric Mixca variants. Besides, other
supervised non-parametric methods (RBF, LVQ, MLP, kNN) cannot compete with
Mixca since it is a hybrid method with an implicit parametric model (ICA), which
allows of a training set of relatively small size.