The Development of the CIE 2000

Colour-Difference Formula:
CIEDE2000

M. R. Luo,* G. Cui, B. Rigg

Colour & Imaging Institute, University of Derby, UK

Received 4 April 2000; accepted 28 July 2000

Editor’s Note: The colour difference formula described in the Appendix to this article has been published by the CIE in
Publication 142-2001. The CIE is now developing a draft standard based on the report and would welcome comments, which
should be directed to Alan Robertson, Chair of TCI-57 (Standards in Colorimetry), National Research Council, Montreal
Road, Ottawa, Ontario KIA OR6, Canada. E-mail: [email protected]

Abstract: A colour-difference equation based on CIELAB is cations in 1976: the CIELAB and CIELUV formulae.1 The
developed. It includes not only lightness, chroma, and hue former was mainly used for the surface colour industries.
weighting functions, but also an interactive term between The latter was used for TV and illumination industries.
chroma and hue differences for improving the performance CIELAB was actually a simplified version of the ANLAB,2
for blue colours and a scaling factor for CIELAB a* scale which includes a fifth-order polynomial and requires an
for improving the performance for gray colours. Four reli- iterative procedure for calculation of ⌬E. (ANLAB was
able colour discrimination datasets based upon object co- derived to fit the Munsell colour spacing in 1944 and was
lours were accumulated and combined. The equation was adopted by the International Standardization Organization
tested together with the other advanced CIELAB based (ISO) for textile applications in 1971.) The available colour-
equations using the combined dataset and each individual difference formulae were tested by McLaren and Rigg3 in
dataset. It outperformed CMC and CIE94 by a large mar- 1976. There were not many differences among most of
gin, and predicted better than BFD and LCD. The equation them. This was mainly due to the limited sets of colour
has been officially adopted as the new CIE colour-differ- discrimination data based on surface colours and the small
ence equation. © 2001 John Wiley & Sons, Inc. Col Res Appl, 26, number of sample pairs in each set. After 1976, consider-
340 –350, 2001 able progress in colour-difference research has been made
not only in deriving new formulae, but most importantly in
Key words: color difference metrics; CIE; CIELAB; CIE94;
generating reliable colour discrimination datasets.
CIEDE2000; BFP; CMC; LCD
In 1974, McDonald4 derived a simple colour-difference
equation based on ANLAB given in Eq. (1):
INTRODUCTION
⌬E ANLAB
Colour difference research has been active over three de- ⌬E a ⫽ , (1)
1 ⫹ 0.0275C ANLAB
cades. The aim is to develop a single-number shade pass/fail
equation for evaluating the small to medium colour differ- where CANLAB is the ANLAB chroma for the standard of the
ences typically used in the surface colour industries. The pair considered.
International Commission on Illumination (CIE) recom- The above equation, including a simple modification of
mended two colour-difference formulae for industrial appli- the ⌬E calculated for differences in the chroma direction,
made a significant improvement over the ANLAB formula.
The concept is now well established and has been used to
* Correspondence to: Dr. M. Ronnier Luo, Color & Imaging Institute,
University of Derby, Kingsway House, Kingsway, Derby DE22 3HL, UK
derive many CIELAB based colour-difference equations. In
(e-mail: [email protected]) the 1970s, McDonald accumulated two datasets based upon
© 2001 John Wiley & Sons, Inc. textile thread samples: one included 640 sample pairs sur-

340 COLOR research and application

 on the shape and orientation of ellipses derived from the
data, nor on the relative size of differences (and ellipses)
within any one dataset. Finally, a consistent pattern of
ellipses was generated. Figure 1 shows these ellipses plotted
in the CIELAB a*b* diagram.
Figure 1 clearly indicates that CIELAB is a poor uniform
colour space, at least where small colour differences are
concerned. For a perfect agreement between experimental
data and CIELAB space, all ellipses should be constant size
circles. Some clear trends in Fig. 1 can be found: ellipses
close to neutral colours are the smallest; ellipses are larger
and longer when chroma is increased; most ellipses point
towards the neutral point except for those in the blue region.
Two combined datasets, BFD-Perceptibility (BFD-P) and
BFD-Acceptability (BFD-A), were formed by adjusting the
raw data between each individual dataset. They included
2776 and 1613 pairs of samples, respectively. The results
indicate that there was little difference between perceptibil-
ity and acceptability data for chromatic differences, but
FIG. 1. Luo and Rigg experimental colour discrimination lightness differences appear relatively larger for percepti-
ellipses plotted in a* b* diagram.
bility data than for acceptability data. Both datasets were
used to derive the BFD(ᐉ:c) colour difference formula.10,11
rounding 55 colour centers, and another included over 8454 Berns et al.12,13 in collaboration with the DuPont com-
pairs surrounding 600 colour centers. They were assessed 5 pany also conducted visual assessments based upon glossy
times by 8 professional colourists, and once by one observer paint samples using the pair comparison method. A dataset,
(a dyehouse manager), respectively. These were used to named RIT–DuPont, including 156 pairs (19 colour centers)
derive the JPC colour-difference equation.5–7 The formula perceptually equivalent to a near-gray anchor pair of 1 ⌬E*ab
was later modified by the members of the Colour Measure- unit was generated. The data were used to derive a relatively
ment Committee (CMC) of the Society of the Dyers and simple colour-difference equation, named CIE94, which
Colourists (SDC), due to the fact that some anomalies were was recommended for field trials by CIE in 1994.14 The
found for colours close to neutral and black. The modified chromatic ellipses for the 19 colour centers were also fitted
equation is named CMC(ᐉ:c)8 and became the ISO standard by Melgosa et al.15 and are plotted in Fig. 2. This plot shows
for textile applications in 1995. the same trends as found in Fig. 1. These figures represent-
Luo and Rigg9 accumulated most of the available exper-
imental data relating to small to medium colour differences
of surface colours. (The data accumulated included various
surface media: textile, paint, ink, etc., and two data types:
perceptibility and acceptability.) Over 120 colour discrimi-
nation ellipses were fitted from these datasets and were first
investigated in CIE x,y chromaticity diagram. It was found
that the ellipses formed a consistent pattern in terms of their
shapes and orientations, but not sizes. This was mainly
caused by different psychophysical methods and different
anchoring pairs being used in various perceptibility exper-
iments, and the different commercial tolerances used in
different acceptability experiments. [The perceptibility data
are described in terms of visual differences (⌬V), which are
proportional to the perceived colour differences. The ac-
ceptability data are described in terms of percentage accep-
tances.] Luo and Rigg then prepared over 600 pairs of wool
samples close to these colour centers, and carried out psy-
chophysical experiments using the gray scale method. Each
pair was assessed by a panel of 20 observers. The data were
used to adjust the sizes of the ellipses obtained earlier. Note
that the adjustments merely altered the visual scale used in FIG. 2. RIT–DuPont colour discrimination ellipses plotted
one dataset relative to the other datasets. This has no effect in a* b* diagram.

Volume 26, Number 5, October 2001 341

ing the two most comprehensive sets of experimental data
agree well with each other.
Kim and Nobbs16 also conducted psychophysical exper-
iments to investigate the parametric effect for evaluating
colour differences. Their experiment was based upon glossy
paint samples. The Leeds data includes 243 and 104 pairs
using gray scale and pair comparison methods, in which
each pair was assessed by 10 –12 and 12–15 observers,
respectively. The experimental results together with the
others were used to derive the Leeds Colour Difference
(LCD) equation.
In 1978, CIE published guidelines17 to coordinate re-
searchers studying colour differences. Five colour centers
were recommended for study. The guideline has been ex-
tremely useful for data gathering purpose, i.e., many re-
searchers such as Strocka et al.,18 Witt,19,20 Cheung and
Rigg,21 Berns et al.12–13 and Witt22 conducted experiments
corresponding to the 5 colour centers. The results from
different datasets agree well with each other.23 The former
3 datasets were already included in the BFD-P data. The
FIG. 3. Experimental ellipses close to the neutral axis plot-
latter dataset22 was used to study the geometrical effects ted in a* b* diagram. The BFD, Leeds, RIT–DuPont, and Witt
(magnitudes and directions of colour differences). There ellipses are plotted in black, red, green, and pink colours,
were 418 pairs, each assessed by a panel of 10 –15 observers respectively.
using the gray scale method. This dataset is denoted as the
Witt dataset in this article.
Although the CMC colour-difference formula was stan- matic ellipses in some regions of the a*b* plane to be
dardized by the ISO for textile applications, colour-differ- rotated. Hence, it can fit the experimental data much better
ence equations such as CIE94, BFD, and LCD perform than CMC and CIE94 in the blue region (see the ellipses in
equally well or better than CMC. Hence, there is a strong blue region of Figs. 1 and 2). The LCD equation has a
desire by industrialists to find a single reliable colour- different RT function.
difference equation suitable for a wide range of industries. The advanced equations have some large discrepancies
All the advanced formulae have a common feature: they among them, as well as some shortcomings. First, they have
were derived by modifying the CIELAB equation. A ge- different methods for predicting lightness differences. The
neric formula given in Eq. (2) represents all these formulae: SL functions of CMC and BFD agree reasonably well, but

冑冉 冊 冉 冊 冉 冊
they disagree with the L* scale used in the CIE94 equation.
⌬L* 2
⌬C* 2
⌬H* 2
Second, they have different methods for predicting hue
⌬E ⫽ ⫹ ⫹ ⫹ ⌬R, (2)
k LS L kC SC k HS H differences. The CIE94 and LCD equations have hue
weighting functions, SH, which are independent of hue
where angle. Furthermore, the hue-dependent functions for BFD
⌬R ⫽ R Tf共⌬C*⌬H*兲, and CMC equations are quite different. Third, the CMC and
CIE94 equations, but not the BFD and LCD equations, give
where ⌬L*, ⌬C*, and ⌬H* are the CIELAB metric light- large errors in predicting chromatic differences for saturated
ness, chroma, and hue differences, respectively, calculated blue colours. Figures 1 and 2 show that all ellipses in blue
between the standard and sample in a pair, ⌬R is an inter- region (particularly around a* ⫽ 10, b* ⫽ ⫺40) do not
active term between chroma and hue differences. The SL, point towards the origin. This disagrees with what is pre-
SC, and SH are the weighting functions for the lightness, dicted by the CMC and CIE94 equations. This is the reason
chroma, and hue components, respectively. The values cal- why a function is required to rotate these ellipses as in-
culated for these functions vary according to the positions of cluded in BFD and LCD equations. Finally, all formulae
the sample pair being considered in CIELAB colour space. wrongly predict chromatic differences between neutral co-
The kL, kC, and kH values are the parametric factors to be lours, i.e., for the majority of colour centers close to neutral,
adjusted according to different viewing parameters such as the chromatic ellipses from experimental datasets are ori-
textures, backgrounds, separations, etc., for the lightness, entated towards around 90°, not constant-diameter circles as
chroma, and hue components, respectively. For CMC and implied by all the advanced equations. This can be seen in
CIE94, an interactive term between the hue and chroma Fig. 3, which includes all ellipses close to the neutral axis
differences, RT, is set to zero. For CIE94, SL equals one. The derived from the BFD, Leeds, RIT–DuPont, and Witt data-
BFD formula has its own lightness formula derived by sets. Note that Fig. 1 includes ellipses based on acceptability
Coates et al.24 together with the RT function to allow chro- experiments. These are excluded from the BFD dataset as

342 COLOR research and application

agreed by TC1-47. For these ellipses the mean ratio of the
major- to minor-axis length is 1.7 to 1, implying a 70%
discrepancy between the equation and experimental results.
With this in mind, a CIE Technical Committee 1-47,
“Hue and Lightness” Dependent Correction to Industrial
Colour-Difference Evaluation,” was formed in 1998. It is
hoped that a generalized and reliable formula can be
achieved. The TC tasks have been divided into four areas: to
accumulate the existing reliable small to medium colour-
difference datasets, and to study the existing SL, SH, and RT
functions. The latter three tasks were assigned to different
research groups. The aim is to develop new weighting
functions by fitting available datasets and to construct a new
CIE colour-difference equation. The TC members have
agreed that the chroma weighting functions from all the FIG. 4. The ⌬E*ab/⌬V values plotted against L* scale for the
advanced equations agree well with each other. Hence, the dataset accumulated by Chou et al. The lightness weighting
one used by CIE94 is adopted, as shown in Eq. (3): functions of Eq. (4), the best fit polynomial, and CIE94 equa-
tions are also plotted using the solid, dotted, and double-
dashed lines.
S C ⫽ 1 ⫹ 0.045C*. (3)
Additionally, it was agreed that the arithmetic mean be-
tions. Figure 4 shows the fit between the experimental data
tween the standard and sample of a pair should be used for
in terms of ⌬E*ab /⌬V and three lightness difference formu-
calculating all weighting functions, rather than using the
lae: a best fit polynomial, Eq. (4), and the CIE94 (or
geometric mean or the standard of a pair. In the CIE
CIELAB) function, plotted in dotted, solid, and double-
Division 1 Meeting at Warsaw in 1999, the research group
dashed lines, respectively. For perfect agreement between
reported their findings. Based on these findings, the need to
the experimental data and CIE94 or L* scale, all points
standardize a new colour-difference equation was agreed.
should lie on a horizontal line. Obviously, this is not the
The basic structure of the equation should be the same as
case. A V or U shape function is required to fit the data. This
given in Eq. (2). The equation should fit well to the four
indicates that the L* scale gives too large ⌬L* values for
datasets: BFD-P, Leeds, RIT–DuPont, and Witt. New light-
lightness differences for dark and light samples. Equation
ness and hue-dependent weighting functions were also pro-
(4) gave quite a good fit to the data and performed almost
posed.
the same as the best fit polynomial. The main difference
between these two formulae is the predictions for lightness
DERIVING INDIVIDUAL COLOUR-DIFFERENCE differences, where L* is greater than 100, which frequently
COMPONENTS occurs for metallic coatings for angles of viewing less than
30° from the specular reflection direction of the sample. The
SL Weighting Function best fit polynomial could underestimate the lightness differ-
One of the major differences between CMC and CIE94 ences in comparison with Eq. (4) by 40% for a pair of
concerns lightness differences. It was decided that new data samples around L* of 150.
on lightness differences were required in order to determine
which formula was correct. SH Weighting Function
In the Warsaw meeting, a new SL function, given in Eq.
(4), was proposed by Nobbs25: A new hue SH function developed by Berns27 is shown in
Eq. (5):
0.015共L៮ * ⫺ 50兲 2
SL ⫽ 1 ⫹ ; (4) S H ⫽ 1 ⫹ 0.015C*T (5)
冑20 ⫹ 共L៮ * ⫺ 50兲 2
where
here L៮ * is the arithmetic mean between the L* values for the
standard and sample of the pair considered. T ⫽ 1 ⫺ 0.17 cos共h៮ ⫺ 30⬚兲 ⫹ 0.24 cos共2h៮ 兲
Recently, Chou et al.26 generated a new set of experi-
⫹ 0.32 cos共3h៮ ⫹ 6⬚兲 ⫺ 0.20 cos共4h៮ ⫺ 63⬚兲,
mental data including 280 pairs of near neutral matte and
glossy paint samples exhibiting mainly lightness differ- where h៮ is the mean of the hue angles for standard and
ences. Each pair was assessed 20 times by a panel of 14 sample in degree.
observers using the gray scale method. The results showed Equation (5) as shown in Fig. 5 (provided by Berns) was
that the lightness difference formula based upon Eq. (4) derived to fit five datasets exhibiting mainly hue differ-
gave an accurate predictions to the visual results and much ences. These include Luo–PhD (already included in the
better predictions than those from the CMC and BFD equa- BFD-P dataset), Qiao et al.’s data,28 Luo–Rigg’s ellipses,9

Volume 26, Number 5, October 2001 343

 son (104 pairs) and gray scale (203 pairs) methods. Two
factors, 0.79 and 0.93, respectively, were used to adjust
these two subsets.]
The PF/3 as given in Eq. (6) used by Guan and Luo29 was
again used as a measure of fit for deriving or testing colour-
difference formulae.

PF/3 ⫽ 100关共 ␥ ⫺ 1兲 ⫹ V AB ⫹ CV/100兴/3, (6)

where the coefficient variation (CV) and ␥ were described

by Alder et al.30 and VAB derived by Schultz.31 For a perfect
agreement between the ⌬E values predicted by a particular
equation and visual results, ⌬V, PF/3 should equal zero. A
PF/3 of 30 indicates a disagreement of about 30%.
FIG. 5. The normalized hue data plotted against CIELAB For deriving or testing colour-difference formulae using
hue angle for the Luo–Rigg, Luo–PhD, Witt, RIT–DuPont,
Qiao et al.’s datasets. The T functions of Eq. (5) are also
the COM dataset, a weighting factor was applied to each
plotted. individual dataset. This allows each dataset to be equally
weighted by having more or less the same number of pairs
(see weighting factors in Table I). This avoids the derived
RIT–DuPont, and Witt. Figure 5 also shows the data from equation from being unduly biased towards the BFD-P set,
each set plotted against CIELAB hue angle. The data were which has more pairs than those of the other datasets. For
normalized; first, by dividing the ⌬V by [Sc in Eq. (3)], and, example, each pair in the RIT–DuPont set was weighted 18
second, dividing by the average for each dataset. Thus, each times to have effectively about the same number of pairs as
dataset has a normalized average of unity. Figure 5 shows in the BFD-P set. Hence, the total number of pairs appar-
that there is very good agreement among data in different ently increases from 3657 to 11273, ⌺NiFi, where Ni and Fi
datasets, and Eq. (5) fits well to most of the data points. are the number of pairs and weighting factor for i dataset.
A “Blue dataset” was also used. This was formed by
extracting from the COM dataset 807 pairs, all with the hue
RT Function
angles between 230 –320°, and showing mainly chromatic
As mentioned earlier, the RT function is intended to differences, i.e., those pairs for which the rotation term
improve the performance of a colour-difference equation for would be significant.
fitting chromatic differences in the blue region. New func- Deriving a Function to Improve the Fitting for Blue
tions have been developed in this section by the authors. Colours. A new colour-difference equation was first estab-
Combined Experimental Dataset: COM. Table I summa- lished by using Eqs. (3)–(5) in Eq. (2) and setting RT equal
rizes the four datasets accumulated by the CIE TC 1-47: to zero. This is called the M equation. The M formula
RIT–DuPont, Witt, Leeds, and BFD-P. Each dataset is provides a base line to indicate the extent of improvement
described by number of pairs, mean ⌬E*ab, and the material by introducing different functions to rotate the ellipses in
used. BFD-P is the largest dataset including 2776 pairs and the blue region (see Figs. 1 and 2).
has many more colour differences than the others. It was The method proposed by Kuehni32 was first investigated
found to be advantageous to merge all datasets to form a by replacing the X tristimulus value by X’ ⫽ c0 X ⫺ (c0 ⫺
combined dataset (named COM) for deriving or testing 1)Z. This method does not need a separate RT function, but
colour-difference equations. A scaling factor was calculated could break the balance between the lightness, chroma, and
for each dataset to adjust the visual results (⌬V) on to a hue differences in Eq. (2), due to the change of the CIELAB
common scale, the same as that of BFD. [The Leeds dataset a* scale. Hence, it was decided to optimize the c0 together
includes two subsets: those conducted using pair compari- with kL and kC to fit the COM dataset (the kH was set to one)

TABLE I. A summary of RIT–DuPont, Witt, Leeds, BFD-P and COM datasets.

No. of Mean Weighting Adjustment
Datasets pairs ⌬Eab
* factor factor for ⌬V Material

RIT–DuPont 156 1.0 18 0.93 Glossy paint

Witt 418 1.9 7 0.43 Glossy paint
Leeds 307 1.6 9 0.79 and Glossy paint
0.93
BFD-P 2776 3.0 1 1.00 Various materials but relative scales of
individual sets adjusted using textile
samples
COM 3657 2.6 11273 — All above

344 COLOR research and application

TABLE II. Performance (PF/3) of colour-difference equation with different rotation functions.
Formula M MR1a (c0 ⫽ 1.1) MR1b (c0 ⫽ 1.2) MR2a MR2b MR2c

Combined data Set (3657 pairs)

COM 36.8 34.6 35.2 34.6 34.2 34.0
kL 1.00 1.02 1.02 1.05 1.06 1.06
kC 0.96 0.96 1.02 1.00 0.99 0.97
Optimised kL and kC 36.8 34.6 35.2 34.5 34.1 33.9
Blue dataset (807 pairs)
Blue 41.9 34.8 32.8 33.9 32.6 31.8
kL 0.85 0.87 0.93 0.99 1.01 1.02
kC 0.99 0.88 1.02 1.03 0.99 0.95
Optimised kL and kC 41.4 34.4 32.7 32.9 32.6 31.6

and the Blue dataset. It was found that c0 ⫽ 1.20 gives the R T ⫽ ⫺sin共2⌬␪ 兲 RC ,
best fit to the Blue data set and c0⫽ 1.1 gives the best fit for
the COM dataset. The equations are named MR1a and where
MR1b, respectively. (Using a c0 of 1.2 instead of 1.1 to the
⌬ ␪ ⫽ 30 exp兵⫺关共h៮ ⫺ 275⬚兲/25兴2 其
COM dataset increases the PF/3 values by 0.6 PF/3 units.)
The testing results are given in Table II, including the and
original equation (kL ⫽ kC ⫽ kH ⫽ 1) together with the
optimal equations (with the optimized kL and kC values and C*
kH ⫽ 1). The results show that both MR1a and MR1b RC ⫽ .
共2 ⫹ 0.07C*兲 3
equations gave a more accurate prediction than the M equa-
tion, and MR1a performed better than MR1b for the COM It was later realized that there is a shortcoming associated
set, and the opposite for the Blue dataset. This implies that with Eq. (7), as shown in Fig. 6. This shows that the Rc
the Kuehni’s approach is quite effective for predicting co- function in Eq. (7) predicts an increase of the rotation effect
lour differences in the blue region. from neutral (C* of zero) until reaching a maximum at
The second approach was to apply RT functions such as around C* equal to 10, and a gradual decrease to the high
those used by the LCD and BFD colour-difference equa- chroma region. By observing Figs. 1 and 2, we expect that
tions. It was found that the performances of these two RT the rotation effect should not decrease for the high chroma
functions were almost identical in terms of PF/3 measure. region. Hence, a new Rc function was developed, as given in
The LCD version given in Eq. (7) was used for further study Eq. (8):

冑
because of its simplicity. This resulted in a new colour-
difference equation, MR2a, including Eqs. (3)–(5) and (7). C* 6
The test results are also given in Table II. It can be seen that R C ⫽ 0.5 . (8)
C* ⫹ 10 6
6

the MR2a performed better than MR1a for both COM and
Blue datasets, and better than MR1b for the COM dataset, The equation is also plotted in Fig. 6 and is of a type used
but slightly worse for Blue dataset: in the CMC and BFD colour-difference equations. As can
be seen in Fig. 6, the curve corresponding to Eq. (8) starts
⌬R ⫽ R T共⌬C*⌬H*兲 (7) with a Rc value of zero at a C* value of 0.0 and ends with
a Rc value of 0.5 for C* values ranging from about 20
upwards. The power factor in Eq. (8) (6 in this case) can be
adjusted to change the slope of the curve. The three coef-
ficients in Eq. (8) (0.5, power value of 6, and a constant,
106) were optimized to fit the COM dataset. The colour-
difference equation including Eq. (8) is named MR2b. Its
performance is also given in Table II. By comparing the
PF/3 measures in Table II, it can be seen that MR2b out-
performed the MR1a, MR1b, and MR2a equations. How-
ever, a shortcoming was later found in that the MR2b
equation has a problem in calculating colour differences for
high chroma colours, i.e., as chroma increases, SC and SH
increase, and the ⌬C and ⌬H terms decrease, but the ⌬R
may not. This may result in a negative value for the ⌬E2.
FIG. 6. The RC functions for the MR2a, MR2b, and MR2c Hence, ⌬E cannot be calculated. A new equation, MR2c,
equations plotted in dashed, solid, and dot-dashed lines, was derived to overcome this problem by replacing ⌬C*
respectively. and ⌬H* by ⌬C*/ SC and ⌬H*/ SH, respectively. The latter

Volume 26, Number 5, October 2001 345

 a⬘ ⫽ a*共1 ⫹ G兲 (10)
where

冉 冑
G ⫽ 0.5 1 ⫺
C* 7
C* 7 ⫹ 25 7
冊 .

G varies from 0.5 at C* ⫽ 0, through G ⫽ 0.30 at C* ⫽ 20,

to G ⫽ 0.06 at C* ⫽ 30. The performance of the M2a and
M2b equations are tested using the COM and a Gray dataset,
in which all pairs with C* less than 10 (662 pairs) were
extracted from the COM set. These results are given in
Table III.
The results show that both the M2a and M2b equations
performed better than MR2c. M2b is preferred to M2a,
because M2b gives a better fit to the COM dataset. The
optimized kL and kC values are so close to 1.00 that there
was no need to rescale the chroma and lightness scales. The
full colour-difference equation M2b is given in Appendix 1
together with 10 pairs of worked examples given in Appen-
FIG. 7. Chromaticity discrimination ellipses corresponding dix 2.
to constant ⌬E according to the MR2c equation.

TESTING THE M2b COLOUR-DIFFERENCE

terms can maintain the balance between the ⌬C* and ⌬H*
EQUATION
terms and the ⌬C*⌬H* term, i.e., no negative value for ⌬E2.
A new RC equation given in Eq. (9) and in Fig. 6 was also The M2b colour-difference equation developed in the earlier
derived to fit the COM dataset. The test results are also sections includes not only the lightness, chroma, and hue
given in Table II. The MR2c equation performed slightly weighting functions, but also an interactive term between
better than the MR2b equation and has no problem in chroma and hue differences for improving blue colours and
calculating differences for saturated colours, as can be seen a scaling factor for the CIELAB a* scale for improving gray
from Fig. 7, where ellipses corresponding to a constant colours. Its performance is compared with five other colour-
colour difference according to the MR2c equation are plot- difference equations (CIELAB, CMC, CIE94, BFD, and
ted. It was selected as the best equation including the LCD) using the COM set together with subsets that repre-
rotation effect. sent different characteristics of colour differences: mainly

冑
lightness, mainly chroma, mainly hue, blue region, and gray
C* 7 region differences. The latter two subsets were used in the
RC ⫽ 2 (9)
C* 7 ⫹ 25 7 previous sections. The former three subsets included the
pairs of samples having 兩⌬L*/⌬E*ab兩, 兩⌬C*/⌬E*ab兩, and
A Simplified a* Function. As mentioned earlier, all ad- 兩⌬H*/⌬E*ab兩 larger than 0.9 in the COM set, respectively.
vanced CIELAB based formulae gave a poor fit to the The testing results in terms of PF/3 are summarized in Table
chromatic differences close to neutral, because they all IV. The equation performed the best in each dataset is
assume that the ellipses in the a*b* diagram are circles. As underlined to aid in comparison.
shown in Fig. 3, all experimental chromatic ellipses close to For each test, the original formula with kL ⫽ kC ⫽ kH ⫽
neutral are ellipses with an orientation around 90°. Attempts 1 together with the optimal equations with the optimized kL
were also made to improve the MR2c equation (the best
colour-difference equation so far). The obvious approach is
to rescale the a* axis. This would stretch the a* scale to TABLE III. Performance (PF/3) of colour difference
make these ellipses become circles. A scaling factor of 1.4 equations with different modified a* scales.
was obtained by minimizing the PF/3 measure to fit the Formula MR2c M2a M2b
COM dataset. This leads to another new colour-difference
equation, named M2a. Combined data set (3657 pairs)
COM 34.0 34.6 32.6
Another method was derived to allow for scaling the a* kL 1.06 0.91 1.00
axis, but with a large effect for colours close to the neutral kC 0.97 0.92 0.95
region and a smaller or no effect for higher chroma colours. Optimized 33.9 34.4 32.6
Again, the type of equation used in Eq. (8) was used to fit Gray dataset (662 pairs)
Gray 35.0 32.1 32.1
COM dataset. This leads to Eq. (10), which was integrated kL 1.24 1.02 0.98
with MR2c to become a new colour-difference equation, kC 0.95 0.86 0.85
M2b: Optimized 33.1 31.5 31.6

346 COLOR research and application

TABLE IV. Testing colour-difference equations using the COM dataset and its subsets.
CIELAB CMC CIE94 BFD LCD M2b

COM set
PF/3, kL ⫽ 1 56.3 37.9 38.1 34.9 35.2 32.6
kL 0.61 0.96 1.26 0.78 1.17 1.02
Optimized kL 52.0 37.9 36.7 33.3 34.6 32.6
Lightness difference (370 pairs)
PF/3, kL ⫽ 1 35.6 37.6 35.3 34.8 32.5 31.9
kL 0.61 0.96 1.26 0.78 1.17 1.02
Optimized kL 35.5 37.7 35.2 35.0 32.4 31.9
Chroma difference (821 pairs)
PF/3, kL ⫽ 1 53.9 33.3 33.4 31.2 31.6 30.4
kL 0.61 0.96 1.26 0.78 1.17 1.02
Optimized kL 52.9 33.3 33.8 30.8 31.7 30.4
Hue difference (577 pairs)
PF/3, kL ⫽ 1 46.8 40.2 37.6 33.3 36.8 33.9
kL 0.61 0.96 1.26 0.78 1.17 1.02
Optimized kL 46.8 40.1 37.6 33.2 36.8 33.9
Blue (807 pairs)
PF/3, kL ⫽ 1 54.3 38.3 40.0 30.0 32.3 30.5
kL 0.61 0.96 1.26 0.78 1.17 1.02
Optimized kL 49.7 38.2 40.3 29.8 31.5 30.6
Gray (662 pairs)
PF/3, kL ⫽ 1 41.6 37.8 42.3 36.7 36.8 32.1
kL 0.61 0.96 1.26 0.78 1.17 1.02
Optimized kL 57.6 37.9 37.5 35.5 34.5 32.0

and kC ⫽ kH ⫽ 1 were tested. The optimized kL value from the saturated blue region. Note that all ellipses in the blue
COM set was used for all subsets. Comparing colour-dif- region predicted by CMC and CIE94 equations only point
ference equations excluding M2b using the COM set and its towards the neutral origin; this results in large prediction
subsets, the BFD and LCD equations outperformed CMC errors.
and CIE94, with CIELAB the worst. This indicates that the All formulae were also tested using the four original
inclusion of a rotation function, RT, is effective for improv- datasets: RIT–DuPont, Witt, Leeds, and BFD-P. The
ing the prediction of the Blue subset (see Blue subset in results are given in Table V. This clearly shows that M2b
Table IV). There is not much difference between each again performs either the best or second best of all the
equation’s performance for lightness differences (see Light- equations. It is quite encouraging that it predicts even
ness subset); in fact, the CMC scale performed the worst more accurately than some formulae that were derived
due to poor predictions for some very dark lightness differ-
ences. All equations gave much more accurate predictions
to chroma differences (see Chroma subset) than CIELAB.
This implies that the SC correction is more effective than the
other corrections. For predicting hue differences, BFD per-
formed better than the others. (Its hue dependent equation
was also integrated with the other components to form
another colour-difference equation. It was found that its
performance was not better than that of the M2b equa-
tion.) It is encouraging that the M2b equation derived
here performed either the best or the second best among
all the equations investigated. In addition, it has a kL
value close to one, implying a well balance structure
between each of the individual colour-difference compo-
nents. Finally, the improvements from CMC or CIE94 to
M2b is considered to be very significant. This clearly
demonstrates that it is possible to standardize a new
improved CIE colour-difference equation for different
applications such as paint and textile.
Figure 8 plots the BFD and RIT-DuPont experimental
ellipses in Figs. 1 and 2 together with the corresponding
ellipses predicted by the M2b colour-difference equation. It FIG. 8. RIT–DuPont and BFD experimental chromaticity
can be seen that a good agreement exists between the discrimination ellipses (in red) compared to the correspond-
experimental and equation predicted ellipses, especially in ing ellipses from the M2b equation (in black).

Volume 26, Number 5, October 2001 347

TABLE V. Testing colour difference equations using 4 individual data sets.
Formula CIELAB CMC CIE94 BFD LCD M2b

RIT-DuPont 22.1 28.1 19.4 24.6 16.9 19.0

kL 0.88 0.96 1.14 0.67 1.02 0.82
Optimized kL 21.7 28.1 18.8 19.8 16.9 17.4
WITT 70.3 46.3 41.6 44.3 41.4 38.4
kL 0.43 0.68 0.95 0.59 1.42 0.80
Optimized kL 61.0 43.4 41.5 38.9 41.2 37.3
Leeds 46.7 28.1 34.0 28.0 26.6 23.1
kL 0.72 1.00 1.35 0.79 1.19 1.06
Optimized kL 43.9 28.1 31.0 25.8 25.4 23.0
BFD 56.2 39.7 42.7 34.4 39.5 37.2
kL 0.72 1.18 1.55 0.99 1.44 1.27
Optimized kL 54.9 39.1 38.9 34.4 36.7 36.1
BIT 43.4 48.8 41.1 36.0 40.8 34.1
kL 0.82 1.60 1.70 1.10 1.61 1.36
Optimized kL 42.3 43.4 32.3 35.7 33.9 31.1

from their own datasets, such as CIE94 with the RIT– and predicted better than BFD and LCD. It has been ap-
DuPont dataset and LCD with the Leeds dataset. Again, proved by the CIE and published as CIE Technical Report,
there is a large improvement from the CIE94 or CMC to “Improvement to Industrial Colour-Difference Evalua-
M2b for each individual dataset. tion.”33 It is designated as the CIE 2000 colour-difference
To some extent, it was expected that the M2b equation formula: CIEDE2000 or ⌬E00.
would perform better than the others, because it was derived
from the above datasets. Finally, it together with the other
equations was tested using a newly accumulated indepen- ACKNOWLEDGMENTS
dent dataset, BIT.23 This is based upon CRT colours in a
study of the effect of changes in viewing parameters on This work represents a close collaboration among all CIE
perceived colour differences. The experiment was divided TC1-47 members. The authors especially thank the strong
into 16 phases according to different viewing conditions: leadership of the TC Chairman, Dr. David Alman, and for
sample sizes, backgrounds, frames, widths, and colours of the development of the new lightness and hue weighting
separations between pairs of samples. For each phase, 134 functions by Dr. Jim Nobbs and Professor Roy Berns,
pairs of colours with an average of 2.7 CIELAB ⌬E units respectively. The authors would also like to thank the So-
were displayed on a CRT. Each pair was assessed 20 times ciety of Dyers and Colourists for supporting new research
by a panel of observers using the gray scale method. In total, work to clarify the new SL function.
2144 pairs are included in this dataset. The results are also
given in Table V. It is nice to see that the M2b equation
again gave the most accurate prediction of all the colour- APPENDIX 1. CIEDE2000 COLOUR-DIFFERENCE
difference equations. Hence, it has been approved by the EQUATION
CIE and published as CIE Technical Reort, “Improvement
to Industrial Colour-Difference Evaluation.”33 It is desig- The symbols used here agree with those used in CIE TC
nated as the CIE 2000 colour-difference formula: CIEDE2000 report but are different from those used in the text.
or ⌬E00.

CONCLUSION Step 1. Calculate the CIELAB L*, a*, b*, and C* as

usual:
A colour-difference equation based on CIELAB is devel-
oped following the procedures agreed by CIE TC1-47. It L* ⫽ 116f共Y/Y n兲 ⫺ 16
includes not only lightness, chroma, and hue weighting
functions, but also an interactive term between chroma and a* ⫽ 500关 f共X/X n兲 ⫺ f共Y/Y n兲兴
hue differences for improving the performance for blue
colours and a scaling factor for the CIELAB a* scale for b* ⫽ 200关 f共Y/Y n兲 ⫺ f共Z/Z n兲兴
improving the performance for gray colours. Four reliable
colour discrimination datasets based upon object colours C *ab ⫽ 冑a* 2 ⫹ b* 2,
were accumulated and combined. The equation was tested
together with the other advanced CIELAB based equations where

再 f共I兲
using the combined set and also each individual dataset. It
I1/3
for I ⬎ 0.008856
outperformed CMC and CIE94 by a considerable margin, f共I兲 ⫽ ⫽ 7.7871 ⫹ 16/116 Otherwise.

348 COLOR research and application

APPENDIX 2. The worked examples.
Pair X Y Z L⬘ a⬘ b⬘ C⬘ h⬘ G T SL SC SH RT ⌬E00

1 19.4100 28.4100 11.5766 60.2574 ⫺34.0678 36.2677 49.7590 133.21 0.0017 1.3010 1.1427 3.2946 1.9951 0.0000 1.2644
19.5525 28.6400 10.5791 60.4626 ⫺34.2333 39.4387 52.2238 130.96
2 22.4800 31.6000 38.4800 63.0109 ⫺32.6195 ⫺5.8663 33.1428 190.20 0.0490 0.9402 1.1831 2.4549 1.4560 0.0000 1.2630
22.5833 31.3700 36.7901 62.8187 ⫺31.2542 ⫺4.0864 31.5202 187.45
3 28.9950 29.5800 35.7500 61.2901 5.5669 ⫺5.3901 7.7488 315.92 0.4966 0.6952 1.1586 1.3092 1.0717 ⫺0.0032 1.8731
28.7704 29.7400 35.6045 61.4292 3.3643 ⫺4.9620 5.9950 304.14
4 4.1400 8.5400 8.0300 35.0831 ⫺44.3939 3.7933 44.5557 175.12 0.0063 1.0168 1.2148 2.9105 1.6476 0.0000 1.8645
4.4129 8.5100 8.6453 35.0232 ⫺40.3237 1.5901 40.3550 177.74
5 4.9600 3.7200 19.5900 22.7233 20.1424 ⫺46.6940 50.8532 293.33 0.0026 0.3636 1.4014 3.1597 1.2617 ⫺1.2537 2.0373
4.6651 3.8100 17.7848 23.0331 15.0118 ⫺42.5619 45.1317 289.43
6 15.6000 9.2500 5.0200 36.4612 47.9197 18.3852 51.3256 20.99 0.0013 0.9239 1.1943 3.3888 1.7357 0.0000 1.4146
15.9148 9.1500 4.3872 36.2715 50.5717 21.2231 54.8444 22.77
7 73.0000 78.0500 81.8000 90.8027 ⫺3.1244 1.4410 3.4407 155.24 0.4999 1.1546 1.6110 1.1329 1.0511 0.0000 1.4440
73.9351 78.8200 84.5156 91.1528 ⫺2.4651 0.0447 2.4655 178.96
8 73.9950 78.3200 85.3060 90.9257 ⫺0.8108 ⫺0.9208 1.2269 228.63 0.5000 1.3916 1.5930 1.0620 1.0288 0.0000 1.5381
69.1762 73.4000 79.7130 88.6381 ⫺1.3477 ⫺0.7239 1.5298 208.24
9 0.7040 0.7500 0.9720 6.7747 ⫺0.4362 ⫺2.4247 2.4636 259.80 0.4999 0.9556 1.6517 1.1057 1.0337 ⫺0.0004 0.6378
0.613873 0.6500 0.851025 5.8714 ⫺0.1477 ⫺2.2286 2.2335 266.21
10 0.2200 0.2300 0.3250 2.0776 0.1192 ⫺1.1350 1.1412 275.99 0.5000 0.7827 1.7246 1.0383 1.0100 0.0000 0.9082
0.093262 0.1000 0.145292 0.9033 ⫺0.0954 ⫺0.5514 0.5595 260.18

Note: X0 ⫽ 94.811, Y0 ⫽ 100.000 Z0 ⫽ 107.304

Step 2. Calculate aⴕ,Cⴕ and hⴕ: where

L⬘ ⫽ L* 0.015共L⬘ ⫺ 50兲 2
SL ⫽ 1 ⫹
a⬘ ⫽ 共1 ⫹ G兲a* 冑20 ⫹ 共L⬘ ⫺ 50兲 2

b⬘ ⫽ b* and

C⬘ ⫽ 冑a⬘ 2 ⫹ b⬘ 2 S C ⫽ 1 ⫹ 0.045C⬘

h⬘ ⫽ tan⫺1共b⬘/a⬘兲, and

where S H ⫽ 1 ⫹ 0.015C⬘T,

G ⫽ 0.5 1 ⫺ 冉 冑 C *ab7
C *ab7 ⫹ 25 7
冊 ,
where

T ⫽ 1 ⫺ 0.17 cos共h⬘ ⫺ 30⬚兲 ⫹ 0.24 cos共2h⬘兲

where C*ab is the arithmetic mean of the C*ab values for a ⫹ 0.32 cos共3h⬘ ⫹ 6⬚兲 ⫺ 0.20 cos共4h⬘ ⫺ 63⬚兲
pair of samples.
and

Step 3. Calculate ⌬Lⴕ, ⌬Cⴕ and ⌬Hⴕ: R T ⫽ ⫺sin共2⌬␪ 兲 RC ,

⌬L⬘ ⫽ L⬘b ⫺ L⬘s where

⌬C⬘ ⫽ C⬘b ⫺ C⬘s ⌬ ␪ ⫽ 30 exp兵⫺关共h⬘ ⫺ 275⬚兲/25兴2 其

⌬H⬘ ⫽ 2 冑C⬘bC⬘s sin 冉 冊 ⌬h⬘ and

冑
2
C៮ ⬘ 7
RC ⫽ 2 .
where C៮ ⬘ ⫹ 25 7
7

⌬h⬘ ⫽ h⬘b ⫺ h⬘s. Note that L⬘, C⬘, and h⬘ are the arithmetic means of the L⬘,
C⬘ and h⬘ values for a pair of samples. For calculating the
h⬘ value, caution needs to be taken for colours having hue
Step 4. Calculate CIEDE2000 ⌬E00: angles in different quadrants, e.g., a standard and a sample
with hue angles of 90° and 300° would have a mean value
⌬E 00 ⫽

冑冉 冊 冉 冊 冉 冊 冉 冊冉 冊
of 195°, which differs from the correct answer, 15°. This
⌬L⬘ 2
⌬C⬘ 2
⌬H⬘ 2
⌬C⬘ ⌬H⬘ can be obtained by checking the absolute difference be-
⫹ ⫹ ⫹ RT ,
k LS L k CS C k HS H k CS C k HS H tween two hue angles. If the difference is less than 180°, the

Volume 26, Number 5, October 2001 349

arithmetic mean should be used. Otherwise, 360° should be 16. Kim DH, Nobbs JH. New weighting functions for the weighted
subtracted from the larger angle, followed by calculating of CIELAB colour difference formula. Proc Colour 97 Kyoto 1997;1:
446 – 449.
the arithmetic mean. This gives 300° ⫺ 360° ⫽ ⫺60° for 17. Robertson AR. CIE guidelines for coordinated research on colour
the sample, and a mean of (90° ⫺60°)/2 ⫽ 15° in this difference evaluation. Color Res Appl 1978;3:149 –151.
sample. 18. Strocka D, Brockes A, Paffhausen W. Influence of experimental pa-
rameters on the evaluation of color-difference ellipsoids. Color Res
1. CIE. Colorimetry. 2nd Ed. CIE Publ. No. 15.2. Vienna: Central Bureau Appl 1983;8:169 –175.
of the CIE; 1986. 19. Witt K, Döring G. Parametric variations in a threshold color-difference
2. Adams EQ. x-z plans in the 1931 ICI system of colorimetry. J Opt Soc ellipsoid for green painted samples. Color Res Appl 1983;8:153–163.
Am 1942;32:168 –173. 20. Witt K. Three-dimensional threshold of color-difference perceptibility
3. McLaren K, Rigg B. The SDC recommended colour-difference for- in painted samples: variability of observers in four CIE color regions.
mula: change to CIELAB. J Soc Dyers Colour 1976;92:337–338. Color Res Appl 1987;12:128 –134.
4. McDonald R. The effect of non-uniformity in ANLAB colour space on 21. Cheung M, Rigg B. Colour-difference ellipsoids for five colour cen-
the interpretation of visual colour differences. J Soc Dyers Colour ters. Color Res Appl 1986;11:185–195.
1974;90:189 –198. 22. Witt K. Geometric relations between scales of small colour differ-
5. McDonald R. Industrial pass/fail colour matching. Part I — Prepara- ences. Color Res Appl 1999;24:78 –92.
tion of visual colour matching data. J Soc Dyers Colour 1980;96:372– 23. Cui G, Luo MR, Rigg B, Li W. Colour difference evaluation using
376. CRT colours. Part I—Data Gathering and testing colour difference
6. McDonald R. Industrial pass/fail colour matching. Part II — Methods formulae. Color Res Appl 2001:26:394 – 402.
of fitting tolerance ellipsoids. J Soc Dyers Colour 1980;96:418 – 433. 24. Coates E, Fong KY, Rigg B. Uniform lightness scales. J Soc Dyers
7. McDonald R. Industrial pass/fail colour matching. Part III — Devel- Colour 1981;97:179 –183.
opment of a pass/fail formula for use with instrumental measurement 25. Nobbs LJH. Private communication.
of colour difference. J Soc Dyers Colour 1980;96:486 – 496. 26. Chou W, Lin H, Luo MR, Westland S, Rigg B, Nobbs J. The perfor-
8. Clarke FJJ, McDonald R, Rigg B. Modification to the JPC79 colour- mance of lightness difference formulae. Col Tech 2001;117:19 –29.
difference formula. J Soc Dyers Colour 1984;100:128 –132 and 281– 27. Berns RS. Derivation of a hue-angle dependent, hue-difference
282. weighting function for CIEDE2000, Proc Colour 2001. Rochester
9. Luo MR, Rigg B. Chromaticity-discrimination ellipses for surface 2001; in print.
colours. Color Res Appl 1986;11:25– 42. 28. Qiao Y, Berns RS, Reniff L, Montag E. Visual determination of hue
10. Luo MR, Rigg B. BFD(ᐉ:c) colour-difference formula. Part I — suprathreshold color-difference tolerances. Color Res Appl 1998;23:
Development of the formula. J Soc Dyers Colour 1987;103:86 –94. 302–313.
11. Luo MR, Rigg B. BFD(ᐉ:c) colour-difference formula. Part II — 29. Guan SS, Luo MR. Investigation of parametric effects using small
Performance of the formula. J Soc Dyers Colour 1987;103:126 –132. colour differences. Color Res Appl 1999;24:331–343.
12. Alman DH, Berns RS, Snyder GD, Larsen WA. Performance testing of 30. Alder C, Chaing KP, Chong TF, Coates E, Khalili AA, Rigg B.
color-difference metrics using a color tolerance dataset. Color Res Uniform chromaticity scales — new experimental data. J Soc Dyers
Appl 1989;14:139 –151. Colour 1982;98:14 –20.
13. Berns RS, Alman DH, Reniff L, Snyder GD, Balonon–Rosen MR. 31. Schultz W. The usefulness of colour-differences formulae for fixing
Visual determination of suprathreshold color-difference tolerances us- colour tolerances. In: Color metrics. Soesterberg: AIC/Holland; 1972.
ing probit analysis. Color Res Appl 1991;16:297–316. p 245–265.
14. CIE. Technical report: Industrial colour-difference evaluation. CIE 32. Kuehni G. Towards an improved uniform color space. Color Res Appl
Pub. No. 116. Vienna: Central Bureau of the CIE; 1995. 1999;24:253–265.
15. Melgosa M, Hita E, Poza AJ, Alman DH, Berns RS. Suprathreshold 33. CIE Technical Report: Improvement to Industrial Colour-Difference
color-difference ellipsoids for surface colors. Color Res Appl 1997; Evaluation. CIE Pub No 142-2001. Vienna: Central Bureau of the CIE;
22:148 –155. 2001.

350 COLOR research and application