Page 218 - Classification Parameter Estimation & State Estimation An Engg Approach Using MATLAB
P. 218

LINEAR FEATURE EXTRACTION                                    207

            a two-dimensional measurement space with samples from two classes.
            The covariance matrices of both classes are depicted as ellipses.
                                                             T
            Figure 6.8(b) shows the result of the operation    1/2 V . The operation
            V T  corresponds to a rotation of the coordinate system such that the
            ellipse of class ! 1 lines up with the axes. The operation    1/2  corres-
            ponds to a scaling of the axes such that the ellipse of ! 1 degenerates into
            a circle. The figure also shows the resulting covariance matrix belonging
            to class ! 2 .
                                             T
              The result of the operation    1/2 V on z is that the covariance matrix
            associated with ! 1 becomes I and the covariance matrix associated with
                             T
            ! 2 becomes    1/2 V C 2 V   1/2 . The Bhattacharyya distance in the trans-
            formed domain is:


                                           2                   3
                                1       1           1 2  T    1 2
                                  T        6  jI þ   V C 2 V  j 7
                                2              q                       ð6:38Þ
                                        2        ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi5
                       J BHAT ð  V zÞ¼ ln4
                                                              1
                                                    1
                                                      T
                                             2 N  j  V C 2 V  j


                                                              2
                                                    2
            The second step consists of decorrelation with respect to ! 2 . Suppose
            that U and   are matrices containing the eigenvectors and eigenvalues of
                                           T
            the  covariance  matrix     1/2 V C 2 V   1/2 .  Then,  the  operation
                     T
             T
            U    1/2 V decorrelates the covariance matrix with respect to class ! 2 .
            The covariance matrices belonging to the classes ! 1 and ! 2 transform
                  T
            into U IU ¼ I and  , respectively. Figure 6.8(c) illustrates the decorrel-
            ation. Note that the covariance matrix of ! 1 (being white) is not affected
                                         T
            by the orthonormal operation U .
              The matrix   is a diagonal matrix. The diagonal elements are denoted
                                                        T
                                                 T
            
 i ¼   i, i . In the transformed domain U    1/2 V z, the Bhattacharyya
            distance is:
                                    "       #     N 1
                                 1
                         1
                     T
                           T
              J BHAT ðU   V zÞ¼ ln    jI þ  j   1  X  ln  1 p ffiffiffiffi  1  ð6:39Þ

                         2
                                          ffiffiffiffiffiffi ¼
                                                            
 i þ p
                                 2   2 N  p j j  2  i¼0  2        ffiffiffiffi
                                                                  
 i
            The expression shows that in the transformed domain the contribution
            to the Bhattacharyya distance of any element is independent. The con-
            tribution of the i-th element is:
                                   1   1 p       1

                                     ln     ffiffiffiffi                       ð6:40Þ
                                            
 i þ p
                                   2   2          ffiffiffiffi
                                                  
 i
   213   214   215   216   217   218   219   220   221   222   223