Page 201 - Classification Parameter Estimation & State Estimation An Engg Approach Using MATLAB
P. 201

190                         FEATURE EXTRACTION AND SELECTION

              A performance measure more suited to express the separability of
            classes is the ratio between interclass and intraclass distance:

                                    J INTER  traceðS b Þ
                                          ¼                             ð6:9Þ
                                    J INTRA  traceðS w Þ

            This measure possesses some of the desired properties of a performance
            measure. In Figure 6.2, the numerator, trace(S b ), measures the area of
            the ellipse associated with S b . As such, it measures the fluctuations of the
            conditional expectations around the overall expectation, i.e. the fluctu-
            ations of the ‘signal’. The denominator, trace(S w ), measures the area of
            the ellipse associated with S w . As such, it measures the fluctuations due
            to noise. Therefore, trace(S b )/trace(S w ) can be regarded as a ‘signal-
            to-noise ratio’.
              Unfortunately, the measure of (6.9) oversees the fact that the ellipse
            associated with the noise can be quite large, but without having a large
            intersection with the ellipse associated with the signal. A large S w can be
            quite harmless for the separability of the training set. One way to correct
            this defect is to transform the measurement space such that the within-
            scattering becomes white, i.e. S w ¼ I. For that purpose, we apply a linear
            operation to all measurement vectors yielding feature vectors y ¼ Az n .
                                                                     n
            In the transformed space the within- and between-scatter matrices
                                    T
                         T
            become: AS w A and AS b A , respectively. The matrix A is chosen such
                     T
            that AS w A ¼ I.
                                                                T
              The matrix A can be found by factorization: S w ¼ V V where L is a
            diagonal matrix containing the eigenvalues of S w , and V a unitary matrix
            containing the corresponding eigenvectors; see appendix B.5. With this
                                                T
            factorization it follows that A ¼    1/2 V . An illustration of the process
                                                    T
            is depicted in Figure 6.2. The operation V performs a rotation that
            aligns the axes of S w . It decorrelates the noise. The operation    1/2  scales
            the axes. The normalized within-scatter matrix corresponds with a circle
            with unit radius. In this transformed space, the area of the ellipse
            associated with the between-scatter is a useable performance measure:

                                                  1  T     1


                            J INTER=INTRA  ¼ traceð  V S b V  Þ
                                                  2
                                                           2
                                                   1  T
                                        ¼ traceðV  V S b Þ             ð6:10Þ
                                                 1
                                        ¼ traceðS S b Þ
                                                w
            This performance measure is called the inter/intra distance. It meets all
            of our requirements stated above.
   196   197   198   199   200   201   202   203   204   205   206