Page 168 - Computational Colour Science Using MATLAB

IMPLEMENTATIONS AND EXAMPLES 155

end
xyzm = r2xyz(meas, 400, 700, ’d65___64’);
xyzp = r2xyz(pred, 400, 700, ’d65___64’);
labm = xyz2lab(xyzm, ’d65___64’);
labp = xyz2lab(xyzp, ’d65___64’);
thisDE = cmcde(labm, labp)

where c1 and c2 are the target area coverages for the inks printed first and
second, respectively, solid1 and solid2 are the solid ink colours for the two inks,
p1 and p2 are the respective tone-reproduction curves and overlap is a 1631
matrix containing the reflectance of the overlap region. The entries of overlap
may be measured from a print of the solid colour of one ink over another or it
could be predicted using a Kubelka–Munk (or some other) model.

9.5.2 Continuous-tone printer

An example of the characterization of a continuous-tone printer is provided from
Sueeprasan (2003) who used a third-order masking model to characterize a
Kodak Color Proofer 9000A printer that is based upon dye-sublimation
technology. The data collected by Sueeprasan is used here to directly compare
characterization models based upon neural networks and polynomial trans-
forms. Two sets of data were obtained: a set of training data contained 729
colours and a set of test data contained 144 colours. For both sets the RGB
inputs to the printer driver were available and the CIE XYZ values for illuminant
D65 were measured (Figures 9.6 and 9.7).
In this example characterization, mappings are developed between the printer
RGB values and the CIE XYZ values. For many practical situations a mapping
that predicts RGB values based upon XYZ values would be more useful. To test
such a mapping requires that the RGB values predicted by the mapping are
physically reproduced using the imaging device and the CIE vales measured and
compared with the target values. This testing procedure is necessary in many
cases but is time consuming. For this example, the advantage of the RGB !XYZ
mapping is that a single test set can be created and used to evaluate various
characterization methods. The XYZ values of the test set are known. Evaluation
of the methods using the test set is achieved simply by comparing the XYZ values
from the model with the actual XYZ values that were measured for the test set.
A characterization was first developed using a neural network and the training
samples. The input to the network was a 36729 array of RGB values. The
original range of RGB values was [0, 255] but they were scaled to be in the range
[0, 1] before presenting them to the network. The output to the network was a
36729 array of L*a*b* values and these were also scaled to be approximately in
the range [0.1, 0.9]. It is important to take into account the range of the output