Invited Paper

The CIE 2000 colour difference formula: CIEDE2000

M. R. Luo, Colour & Imaging Institute, University of Derby

ABSTRACT

The CIE Technical Committee TC 1-47 Hue and Lightness Dependent Correction to Industrial Colour Difference
Evaluation was established in October 1998 and its aim was to improve the performance of the C1E94 colour-difference
formula. As a result of close collaboration between the TC members, the CIE 2000 colour difference formula, CIEDE2000,
was developed within two years. This paper describes the development of this formula.

Keywords: Colour difference formula, CIELAB, CMC, C1E94, BFD, LCD, CIEDE2000, Lightness, chroma, hue weighting
functions and parametric factors.

1. INTRODUCTION

Colour difference is one of the important research fields in colorimetry with the goal of developing a single number pass/fail
formula for industrial applications, i.e. the same colour difference predicted by the formula for pairs of samples located in
different colour regions gives the same perceptual differences. Many colour scientists have striven to achieve this since the
earliest colour difference formulae in the 30s. Undoubtedly one of the major contributions of the CIE in this century is the
development of the CIE 2000 colour difference formula (CJEDE2000). The work was led by Alman, the chairman of CIE
TC1-47 Hue and Lightness Dependent Correction to Industrial Colour Difference Evaluation. The first TC meeting was
held at Baltimore in 1998. Various tasks were identified and assigned to TC members. It was agreed that the work should
2
focus on the modification of the C1ELAB formula following the lines of C1E94 and rather than on the development
of a uniform colour space. The final structure of the formula is given in Equation (1).

AE = !(v' (AC' (AH' (AC'Yvi'

I 2 2 2

00 Il I +R
I +1 I +1 I I I (1)
kLSL ) ) kHSH ) T(kS IkHSH)
There are four enhancements to the CIE94formula: a modification of the CIELAB a* scale, which resulted in the new zC'
and AL!' terms in Equation (1), a new lightness-difference weighting function (SL), for which AL' = L1L, a hue-chroma
interactive term (RT) and a new hue-difference dependent function (S11). The Sc function is the same as that in C1E94 and the
k and kH are the parameters for taking into account parametric effects.

This paper describes the development of CIEDE2000 and explains the reasons for these enhancements. Although it was an
iterative process, the following sections are arranged according to the sequence of development. (The symbols used in
Equation (1) and following equations correspond to those in the final CIE report. Each weighting function in the
CIEDE2000 is calculated using the arithmetic mean of the coordinates of the standard and sample to avoid different results
when they were exchanged.)

2. DATASETS AND MEASURE OF FIT

The first task for developing a colour difference formula was to determine what experimental datasets were to be included.
,
Four datasets considered to be the most reliable were accumulated: RIT-DuPont , Witt Leeds and BFD-Perceptibility 6
including 156, 418, 307 and 2776 colour-difference pairs. They have a mean iE*j, of 2.7. These datasets were merged to
form a combined dataset, called COM. This was considered to be advantageous because fitting a formula to one dataset
rather than fitting a formula to each individual dataset results in a single formula. The visual differences (zlV) in each
individual dataset were adjusted to have the same visual scale as that of the BFD-Perceptibility dataset. Finally, the pairs in
each dataset were weighted 18, 7, 9 and 1 times for the RIT-DuPont, Witt, Leeds and BFD-Perceptibility datasets
respectively. This was done to avoid too much weight for the BFD-Perceptibility dataset when developing a formula.

The PF/3 given in Equation (2) was used by Guan and Luo a measure of fit for deriving colour difference formulae.

9th Congress of the International Colour Association, Proceedings

554 of SPIE Vol. 4421 (2002) © 2002 SPIE · 0277-786X/02/$15.00

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 05/18/2015 Terms of Use: http://spiedl.org/terms

 PF/3 = 100 [(l)+VAB + CV/100]13 (2)
where the coefficient variation (CV) and y were described by Alder et al 8 and VAB derived by Schultz . For a perfect
agreement between the tiE values predicted by a particular formula and visual results, zlV, PF/3 should equal zero. A PF/3
of 30 indicates a disagreement of about 30%.

3. THE DEVELOPMENT OF THE LIGHTNESS WEIGHTING FUNCTION (SL)

Figure 1(a) shows five lightness weighting functions from CMC, BED, LCD , CIELAB (or C1E94) and CIEDE2000 plotted
against L*. It can be seen that all functions except CIELAB or (C1E94) indicate that the L* scale over-predicts the lightness
differences for the lighter samples. For predicting lightness differences for the darker samples, there are large differences
between these functions. The CMC and BED SL functions indicate an under-prediction but the LCD function suggests that
no correction to L* scale is needed. Furthermore, the CIEDE2000 function gave a diametrically opposite prediction.
Nobbs ° at Leeds University investigated the lightness weighting functions (SL). He selected the mainly lightness-difference
pairs from the four datasets and plotted AE*j/z1V value of each pair against L* scale. He found that all datasets except BFD-
Perceptibility show a distinct trend having a minimum value of AE*ab MV at about L* =50 and the values rise to a maximum
for both the very dark and very light regions. For the BFD-Perceptibility data, all functions fit reasonably well to the data
due to a large spread of data.

At a later stage, the results from a research project supported by the Colour Measurement Committee of the Society of the
Dyers and Colouritsts were reported by Chou et al . Two hundred eighty pairs of near neutral matt and glossy paint
samples exhibiting mainly lightness differences were assessed. The results showed that the lightness weighting function
based upon Equation (3) gave an accurate prediction to the visual results. Figure 1(b) shows the fit between the
experimental data in terms of z1E',,, MV and 3 functions: a best fit quadratic equation, Equation (3) and the C1E94 (or
CIELAB) function plotted in dotted, solid and double-dashed lines respectively. The figure clearly shows a V or U shape
function is required to fit the data. Equation (3) was finally chosen due to the fact that the best fit quadratic equation could
under-estimate the lightness differences comparing with Equation (3) by 40% for a pair of samples around L* of 150, which
may well occur for metallic coatings.

SL_l+ (3)
20 + (L'_5o)

2.5

/
1.5
E LtE
AV
11 SF0
CIELAB (C1E94)

CIELAB(C1E94)
0.5
CMC
- - CMC
— —LCD
CIEDE2000
— BFD

0 10 20 30 40 50 60 70 80 90 100 0 30 40 60 80 100
L L*
(a) (b)
Figure 1. (a) Various lightness weighting functions plotted against L* scale. (b) The ziE(CIEL4B) /LIV values plotted against L* scale for
SDC/CMC dataset.

Proc. SPIE Vol. 4421 555

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 05/18/2015 Terms of Use: http://spiedl.org/terms

 4. THE DEVELOPMENT OF THE CHROMATIC ROTATION FUNCTION (RT)
Luo and Rigg 6 found that from their experimental ellipses plotted on C1ELAB a *b* diagram (see Figure 2), all ellipses
around blue region all do not follow the pattern of the others as pointing toward the origin. This resulted in the development
of BFD formula including a hue and chroma interactive term
120
(see Equation (1)). This concept was also employed by Kim b*
and Nobbs to develop their LCD formula by including an
100 )
equation simpler than the BFD fifth-order cosine equation.
Kuehni
12
was also investigated this by replacing the X 0
80
tristimulus value by X' = c0X-(c0-1)Z.
60 0
All three types of equations were optimised to fit the COM
000
dataset. Furthermore, a sub-dataset (Blue) selected from the
COM including all pairs with hue angles between 230° and
40
cç
cO
8l
0
;,
320° was also used to test the performance of each optimised 20 Oç 0
C
equation. Finally, a modified LCD hue and chroma interactive a*
o 0
term as shown in Equation (4) was adopted due to its testing
performance and robustness.
0

-20 0
a cO0Q0!
oOo
I

0
-40
RT sin( 2A0 )R
where -60
I ii— \ 12 1 -60 -40 -20 0 20 40 60 80 100 120
AO=30exp-[h'—275°)I25j I (4)
—6
I Figure 2. Luo and Rigg experimental colour discrimination
R = 0. —6 ellipses plotted in a* b* diagram
'4C' +106

5. THE DEVELOPMENT OF THE NEW REDNESS-GREENNESS SCALE (a')

It is also known that all CIELAB modified formulae give a poor fit to the chromatic differences close to neutral, because
they all assume that the ellipses in a *b * diagram are circles. As shown in Figure 2, all experimental chromatic ellipses close
to neutral are ellipses with an orientation around 90°. Attempts were also made to improve this shortcoming. The obvious
approach is to re-scale the a axis. This would stretch the a 'K scale to make these neutral ellipses become circles. A scaling
factor of 1.4 was obtained by minimising the PF/3 measure to fit the COM data set. Another method was derived to allow
for scaling the a 'K axis but with a large effect for colours close to the neutral region and a smaller or no effect for higher
chroma colours. This leads to Equation (5) which was later included in the CIEDE2000 equation.
a'=a*(l+G) (5)
where

G =0.5 1— /_7
VC' +25

6. THE DEVELOPMENT OF THE HUE WEIGHTING FUNCTION (SH)

A new hue SH function developed by Berns 12 is given in Equation (6). Equation (6) was derived to fit 5 datasets exhibiting
mainly hue differences. These include Luo-PhD (already included in the BFD-P data set), Qiao et al data 13 Luo-Rigg
ellipses and RIT-DuPont and Witt. Figure 3 shows the data from each set plotted against CJELL4B hue angle. The data were
normalised so that each dataset has a normalised average of unity. It can be seen that there is good agreement between data
in different datasets and Equation (6) fits well to most of the data points.

556 Proc. SPIE Vol. 4421

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 05/18/2015 Terms of Use: http://spiedl.org/terms

 SH =1+O.O15CT (6)
where
T =1_O.17cos(h)_3O0)+O.24cos(2h))+O.32cos(3h+60)_O.2Ocos(4h_630)

2.5

I 1.5
w

E
z0

0.5

60 120 180 240 300 360

CIELAB hue angle, degrees

Figure 3. The normalised hue data plotted against CIELAB hue angle for the five datasets
studied. The T function of Equation (6) is also plotted.

7. THE PERFORMANCE OF CIEDE2000

Although the PF/3 measure (Equation (2)) was used in the development of the CIEDE2000, it does not indicate whether or
not there is any statistically significant difference between the two formulae compared in predicting a particular dataset.
Alman 14 proposed a method with the following testing hypothesis:

H0: Ve,A = Ve,B

HA: Ve,A not equal Ve,B
Fresr Ve,A/ Ve,B
Fcrit = F(dfA, dfB, 0.95) (critical F value)
Reject H0 if Ftest> Fcrit OR if Ftest < l/Fcrit
Ve A = (Av, — (a0 + a1AE ))2
where —2)

VB =(Av, —(a2 +a3f))2J%2)

where Fcrit value can be calculated or found from statistics textbooks; Ve,A and VeB represent the residual error variances
after scaling correction for the full and reduced CIEDE2000 formulae with dfA and dfB error degrees of freedom
respectively. The a0 and a1 coefficients are obtained from the least-square fit between the AE from the full CIEDE2000
formula (the combination of the Equations (1) to (6) except Equation (2)); the a2 and a3 coefficients are obtained from the
least-square fit between the AE' from the reduced CIEDE2000 formula. Finally, N is the number of pairs in the dataset.
The first test was made to compare the full CIEDE2000 formula against each reduced version. For example, considering
whether SL should be calculated from Equation (3), the reduced version assumed that kL SL = 1. For testing each version
against the full formula, two types of datasets were used. One was the combined data set, COM, and the other was a subset
from the COM dataset. The subsets can be divided into two groups: individual component differences and selected colour
regions. The former includes subsets consisting of essentially lightness differences, essentially chroma differences, and
essentially hue differences. For example the lightness differences were selected from the COM dataset by only including
those pairs where the value of JOUIAL*I/AE*ab was larger than 90%. Similarly JOOIZIC*I/LIE*0b was larger than 90% for

Proc. SPIE Vol. 4421 557

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 05/18/2015 Terms of Use: http://spiedl.org/terms

 chroma differences. The latter group includes all pairs located in the 2300< h< 320 °sector for the 'blue' differences, and
c*<1o for the grey regions respectively. The testing results are summarised in Table 3 in terms of Ftest values. The
underlined bold values indicate a significantly better performance of CIEDE2000 compared to the reduced version at the
99% confidence level, and at the 95% level for those in bold. The other values show that there is no significant difference
between the performance of the two formulae tested.
Table 1 The Ftest values for comparing CJEDE2000 with its reduced versions

Dataset No. of Degree of kLSL=l kcSc=l kHSH1 T1 RTO a'

Pairs Freedom
COM 1 1273 3655 0.889 0.472 0.770 1 .001 0.872 0.916
Lightness 1730 348 0.879 1.012 0.989 1.002 1.001 0.993
Chroma 2476 801 1.043 0.521 0.922 0.982 0.795 0.927
Hue 1438 486 1.007 0.931 0.739 0.858 0.985 0.942
Blue 2523 773 0.966 0.573 0.707 1.019 0.678 0.821

The results show that (except for the T function) inclusion of each term in the formula gives a highly significant (99% level)
improvement in the performance of formula. They are arranged as Sc, 5m Sb RT and a ' according to the degree of
improvement. Although the T function is judged to be an insignificant improvement using the COM dataset, it is still a
significant improvement at the 95% level using the Hue subset. The similar confidence level of improvement can also be
found for the a ' component using the Grey subset.
Table 2 The Ftest values for comparing CJEDE2000 with the other colour difference formulae

Data set No. of Degree of CIELAB CMC CIE9 4 BFD LCD

Pairs Freedom
COM 11273 3655 0.488 0.763 1.027 0.891
Lightness 1730 348 0.878 0.837 0.877 0.947 1.011
Chroma 2476 801 0.471 0.819 0.818 0.928 0.894
Hue 1438 486 0.544 0.757 0.839 0.943 0.841
Blue 2523 773 0.417 0.715 0.629 1.082 0.817
Grey 2436 637 38 0.863 j4 0.960

The same testing method was applied to compare CIEDE2000 with CIELAB, C1E94, CMC, BFD and LCD formulae. The
results are given in Table 2 in terms of Ftest values for each comparison. Again, for the underlined bold values, there is a
significant superiority of CIEDE2000 at the 99% confidence level, and at the 95% level for those in bold.

For the COM dataset, CIEDE2000 outperformed the other formulae except for BFD the improvement always being
significant at the 99% level. For the Chroma subset, it is surprising that C1EDE2000 performed significantly better than
C1E94 (they have the same 5c function) For the Blue subset, the formula again performed significantly better than the other
formulae except for BFD. For the Grey subset, it still predicted more accurately than C1E94. The present results clearly
showed that the CIEDE2000 formula performed much better than the current standards: C1E94 and CMC formulae.

8. CONCLUSIONS

This paper summarises the work involved in the development of CIE 2000 colour difference formula: C1EDE2000. Its
performance was thoroughly evaluated by comparing with its reduced versions and the other formulae. The results show
that all the components developed showed a significant enhancement to the CIELAB formula. In addition, this formula
should be confidently used for industrial applications because it outperformed the current standards: CMC and C1E94 by a
considerable margin.

ACKNOWLEDGEMENTS
The author would like to thank D. Alman for his outstanding leadership of CIE TC1-47, J. Nobbs and R. Berns for
contributing their work, B. Rigg for providing advice, W. Chou for conducting experiment and the other TC members for
their strong encouragement. Also, the author would like to thank the support from the Society of Dyers and Colourists

558 Proc. SPIE Vol. 4421

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 05/18/2015 Terms of Use: http://spiedl.org/terms

 (SDC), UK, for sponsoring colour-difference research, and finally, G. Cui for his involvement at every stage of the
development work.

REFERENCES

1. CIE. Technical report: Industrial colour-difference evaluation. CIE Pub. No. 1 16. Vienna, Austria: Central Bureau
oftheClE; 1995.
2. F. J. J. Clarke, R. McDonald, and B. Rigg, Modification to the JPC79 colour-difference formula. J Soc Dyers Go!
1984;100:128-132 and 28 1-282.
3. R. S. Berns, D. H. Alman, L. Reniff, G. D. Snyder, and M. R. Balonon-Rosen, Visual Determination of
Suprathreshold Color-Difference Tolerances Using Probit Analysis. Go! Res App! 1991 ;16:297-316.
4. K. Witt, Geometric Relations between Scales of Small Colour Differences. Go! Res App! 1999;24:78-92.
5. D. H. Kim, and J. H. Nobbs, New weighting functions for the weighted CIELAB colour difference formula. Proc
Colour 97 Kyoto 1997;1 :446-449.
6. M. R. Luo, and B. Rigg, BFD(!:c) colour-difference formula. Part I —Development of the formula. J Soc Dyers
Gol 1987;103:86-94. M. R. Luo, and B. Rigg, BFD(!:c) Colour-difference formula, Part Il—Performance of the
formula. J Soc Dyers Go! 1987;103:126-132.
7. 5. 5. Guan, and M. R. Luo, Investigation of parametric effects using small colour differences. Gol Res Appi
1999;24:331-343.
8. C. Alder, K. P. Chaing, T. F. Chong, E. Coates, A. A. Khalili and B. Rigg, Uniform chromaticity scales - new
experimental data., J. Soc. Dyers Gol., 98, 14-20 (1982).
9. W. Schultz, The Usefulness of Colour-Differences Formulae for Fixing Colour Tolerances, Color Metrics,
Soesterberg, AIC/Holland, 245-265 (1972).
10. J. H. Nobbs, Lightness difference weighting term 5L' unpublished CIE TC 1-47 report (1999).
11. R. G. Kuehni, Towards an improved uniform color space, Go!. Res. App!. 24, 253-265 (1999).
12. R. S. Berns, Derivation of hue-angle dependent, hue difference weighting function for CIEDE2000, ibid.
13. Y. Qiao, R. S. Berns, L. Reniff, and E. Montag, Visual Determination of Hue Suprathreshold Color-Difference
Tolerances. Go! Res App! 1998;23:302-3 13.
14. D. H. Alman, Performance comparison for full and reduced color-difference models, unpublished CIE TC 1-47
report (2000).

Proc. SPIE Vol. 4421 559

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 05/18/2015 Terms of Use: http://spiedl.org/terms