d a low
nents of
for later
rmation
original
(3.1)
r space
visible
2 stable
n some
s. Then
by CIE
h other.
mposite
able the
squares
>sulting
of the
images
Stefan Diener
image CIE XYZ CIE Luv CIE Lab lilola HSI YO,C,
1 1.436 0.950 0.965 0.850 0.822 0.978
2 1.076 0.982 0.984 0.960 0.968 1.031
3 1.124 0.890 0.997 1.017 0.984 0.987
4 1.026 0.936 0.893 1.127 1.106 0.912
5 1.107 0.971 1.024 0.965 0.976 0.957
6 1.302 0.939 0.952 0.856 1.015 0.937
7 1.274 0.961 0.954 0.951 0.917 0.943
8 0.797 1.030 1.029 1.056 1.036 1.052
9 0.816 0.944 0.942 1.178 0.964 1,155
mean 1.106 0.956 0.971 0.996 0.976 0.995
RMSE 0.213 0.038 0.043 0.112 0.079 0.075
Figure 4. RMSE of the grey values representing the matching accuracy relative to the mean occurred errors.
It seams that CIE Luv provides the best matching accuracy due to the smallest mean RMSE (0.956) and the smallest
divergence from that value (0.038). Second best is CIE Lab. I;bI;, HSI and YC,C, are in the midfield, but CIE XYZ is
the most inaccurate method.
The performance of the colour space transformations is very important for daily work. Figure 5 shows the mean
computation times for an 1024x1024 pixels composite in seconds relative to CIE XYZ.
Y CO:
1.001
CIE Lab lilola HSI
1.992 0.999 1.298
CIE XYZ
1.000
CIE Luv
1.439
rel. to CIE XYZ
Figure 5. Computation time relative to CIE XYZ for a complete 1024x1024 pixels colour composite
from a 256x256 pixels colour image and a 1024x1024 pixels panchromatic image.
As expected the three linear colour space transformations are the fastest. HSI is almost 30% slower, CIE Luv 44% and
CIE Lab consumes the double time of CIE XYZ. All transformations are implemented using floating-point operations,
so it is surely possible to speed up the current computation time.
4 CONCLUSIONS
In section 2 we discussed that one has to determine functional coherence between the radiometric behaviour of the CCD
and all influencing factors for the DMC normalisation approach. To get these functional coherences we carried out
several tests at Z/I Imaging to qualify functions and parameters for each covered characteristic. We investigated the
behaviour due to temperature changes, properties of lens and aperture, influences by TDI shifting and the special
sensivity of each CCD element itself. It is shown that one has only to store one normalisation look up table which
consists of a dark and a bright image.
Section 3 introduces the common workflow for the colour composite generation. The short discussion about the
interpolation methods showed why bicubic interpolation is used in necessary all interpolation steps.
Thereafter six wide-spread colour space transformations were compared. A first ranking was made after the optical
inspection of generated composite images, although this wasn't an easy task for the observers. The RMSE of the
difference images and the achievable matching accuracy gave a second objective ranking. Last but not least the required
computation times were compared.
So the CIE Lab colour space transformation produces high quality composites, but the required computation time
doesn't fit our requirements. The related CIE Luv is faster and reserves the quality. So we suggest to use this colour
space transformation because it's a good compromise between image quality and computation time. If the latter is the
most important item Y C,C, will be the best choice.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part Bl. Amsterdam 2000. 87
