Page 47 -

2.2 Feature S~ace Metrics 33

Figure 2.12. Examples of two-class discriminants with pattern similarity measured
according to a Mahalanobis metric: (a) Linear; (b) Quadratic.

2.3 The Covariance Matrix

Consider the first two classes of cork stoppers from the dataset described in
Appendix A (Cork Stoppers.xls). Figure 2.13 shows the scatter plot of these two
classes, o, and cy, using features N (number of defects) and PRTlO=PRT/IO (total
perimeter of the defects in tens of pixels).

I2O 7

Figure 2.13. Scatter plot for two classes of cork stoppers with features N, PRT10.

The scatter plot shown in Figure 2.13 uses the same scale for the coordinate
axes. This allows the evaluation under equal conditions of the features'
contributions to the Euclidian distance measures, the ones that we are accustomed