Page 189 - Computational Colour Science Using MATLAB
P. 189

176                   MULTISPECTRAL IMAGING

                    xlabel(’wavelength’)
                    title(’(a)’)
                    subplot(2,2,2)
                    plot(w,mspec,’k-’)
                    xlabel(’wavelength’)
                    title(’(b)’)
                    subplot(2,2,3)
                    plot(w,v,’k-’)
                    xlabel(’wavelength’)
                    title(’(c)’)
                    subplot(2,2,4)
                    plot(w,v1,’k-’)
                    xlabel(’wavelength’)
                    title(’(d)’)

                 The lower left (c) and lower right (d) panes of Figure 10.6 show the first three
               basis functions computed from the raw set of reflectance spectra and from a
               centred set of spectra (where the mean is first subtracted), respectively. Although
               the two sets of basis functions look quite different if we correct for the arbitrary
               sign of the functions it can be seen that there are only small differences between
               the two sets of basis functions (Figure 10.7).




               10.5.2 Representation of reflectance spectra in a linear model

               The computation of the basis functions using svds allows us to write

                    P ¼ Ba,                                                     ð10.15Þ
               where P is the w6n matrix of reflectance spectra, B is the w6m matrix of basis
               functions, and a is the m6n matrix of coefficients, where n is the number of
               samples, w is the number of wavelength intervals at which the samples are
               represented and m is the number of basis functions in the linear model. The
               coefficient matrix a thus allows each reflectance spectrum to be represented by
               just m coefficients. The central goal of PCA is to reduce the dimensionality of a
               data set whilst retaining as much as possible of the variation present in the data
               set (Jolliffe, 1986). It is relatively straightforward to compute the coefficient
               matrix a by rearranging Equation (10.16),

                    a ¼ B P,                                                    ð10.16Þ
                         þ
               where B denotes the pseudoinverse of the matrix of basis functions B. We note,
                      +
               however, that if the basis functions are orthonormal, then Equation (10.16) is
               equivalent to Equation (10.17),
   184   185   186   187   188   189   190   191   192   193   194