You are on page 1of 7

Testing CIELAB-Based Color-

Difference Formulas

Manuel Melgosa
Departamento de Óptica, Facultad de Ciencias, Universidad de Granada. 18071—Granada (Spain)

Received 29 September 1998; accepted 17 March 1999

Abstract: The CMC, BFD, and CIE94 color-difference for- weighted by two different types of variables: the weighting
mulas have been compared throughout their weighting functions (designated by the letter W), which are intended
functions to the CIELAB components DL*, DC*, DH*, and to improve the perceptual uniformity of CIELAB; and the
from their performance with respect to several wide data- parametric factors (designated by the letter k), which are
sets from old and recent literature. Predicting the magni- meant to account for the influence of specific experimental
tude of perceived color differences, a statistically significant conditions on perceived color differences. In addition, in the
improvement upon CIELAB should be recognized for these case of the BFD formula,4 there is an additional crossed
three formulas, in particular for CIE94. © 2000 John Wiley & term in Eq. (1) to account for the rotation of some experi-
Sons, Inc. Col Res Appl, 25, 49 –55, 2000 mental discrimination ellipses in the plane a*b* (color
Key words: color-difference formulas; CIELAB; CIE94 discrimination ellipses predicted by CIE94 or CMC are
radially oriented in this plane).
The CMC and BFD formulas have been successfully used
INTRODUCTION by different industries, mainly in the United Kingdom, as a
It is generally assumed that, although much more research better alternative to the CIELAB formula. Specifically, in
needs to be done in the field of industrial color-difference textile pass/fail color-difference work, the Colour Measure-
evaluation, in the last few years significant advances have ment Committee has recently recommended again the use of
been made through the use of several CIELAB-based for- the CMC formula.6 The CIE94 formula was the result of the
mulas. As is well known, using CIELAB the color-differ- analyses performed by the CIE TC 1-29 seeking a simple
ences are easily computed as Euclidean distances in this CIELAB-based formula with a significant improvement in
space or, better yet, from an intuitive point of view, as the the prediction of perceived color-differences, in particular
result of the differences of the three color attributes (light- under experimental conditions typical of those commonly
ness, chroma and hue differences: DL*, DC*, DH*). How- used in industrial applications. In addition to the CMC,
ever, when the CIELAB space was recommended in 1976,1 BFD, and CIE94 formulas, the use of other unpublished
it was indicated that “in different practical applications it formulas such as those from Marks & Spencer or Datacolor
may be necessary to use different weightings for DL*, also should be mentioned.7 The active research in this field
DC*, and DH*.” 2 This indication was followed, among has led to, among other things, the publication of new CIE
others, by the CMC,3 BFD,4 and CIE945 color-difference guidelines8 updating those given in 1978,9 and to the ap-
formulas, which can be considered CIELAB-based formu- proval of the CIE TC 1-47 “Hue and lightness-dependent
las and have the next common structure: correction to industrial color-difference evaluation,” at-

FS D S D S D G
DL* 2
DC* 2
DH* 2 1/ 2 tempting to improve the performance of the recently rec-
DE* 5 1 1 . (1) ommended CIE94 formula.
k LW L k CW C k HW h
For a long time, the use of multiple color-difference
In this equation, the lightness, chroma, and hue differences formulas for industrial applications was a major shortcom-
are always computed from CIELAB, and each one is ing, which was successfully solved by the recommendation
of CIELAB and CIELUV made in 1976. In particular, ever
since this recommendation, the use of the CIELAB space
Correspondence to: Manuel Melgosa; e-mail: mmelgosa@goliat.ugr.es
Contract grant sponsor: Ministerio de Educación y Ciencia (España)
has been widely accepted.10,11 As pointed out by Witt,12
Contract grant number: Research Project PB96-1454 from the current use of different CIELAB-based formulas,
© 2000 John Wiley & Sons, Inc. we should consider whether we are now returning to the

Volume 25, Number 1, February 2000 CCC 0361-2317/00/010049-07 49


FIG. 1. Dependences of the weighting functions for lightness, chroma, and hue for different color-difference formulas. A
common origin with CIELAB has been adopted in each plot by multiplying by the scaling factors given in Table I.

past with a high diversity of formulas, which could be a new CIE94 formulas, as shown in Fig. 1. In this figure, seeking
potential source of misunderstanding in industrial applica- a clearer comparison of the dependences shown by each one
tions. In addition, as pointed out by Alman et al.,13 it might of the three weighting functions of these formulas, we have
be thought that pragmatic considerations dictate that revi- placed them at the same origins by multiplying by appro-
sion of the color-difference metric standards should occur priate scaling factors given in Table I. The weighting func-
only if the improvement is real (statistically significant) and tion for lightness (upper left) is an increasing function of
substantial enough to counterweight the expenses of soft- lightness for CMC and BFD, but there is no such correction
ware modification and personnel retraining. for CIE94. Relatively similar increasing weighting func-
For these reasons, in this article we report some compar- tions for chroma are proposed by the three CIELAB-based
isons of CIELAB and the CIELAB-based formulas CMC, formulas analyzed here (upper right). Finally, for the
BFD, and CIE94. In the first part of the next section, we weighting functions for hue, in general, there are both hue
compare the different weighting functions proposed by and chroma dependences. The chroma dependence (lower
these formulas, showing that they cannot be considered right) again shows relatively similar increasing functions for
identical, although they have a similar structure [Eq. (1)]. the three formulas, but not the hue dependence (lower left):
Next, we analyze the performance of these formulas, to- there are two clear minimums for CMC, three minimums for
gether with CIELUV and CIELAB, with respect to different BFD, and no hue dependence for CIE94. It should be said
wide groups of experimental data selected from old and
recent literature. Finally, the statistical significance of the
improvements achieved by these formulas have also been TABLE I. Scaling factors used to multiply the different
tested for these experimental data. weighting functions and put them in Fig. 1 at the same
origin than CIELAB, for comparative purposes.
RESULTS AND DISCUSSION Wh vs. Hue Wh vs. C*
WL WC (C* 5 50) (h 5 45°)
Comparison of Weighting Functions
CMC 1.96 1.43 0.57 1.45
Different weighting functions for lightness (W L ), chroma BFD 4.18 1.80 0.74 1.81
(W C ) and hue (W h ) are proposed by the CMC, BFD, and CIE94 1.00 1.00 0.57 1.00

50 COLOR research and application


FIG. 2. Ratio of the weighting functions for chroma and hue
FIG. 3. Relationship between the weighting function for
for different formulas—CIE94, CMC (h 5 45° and 135°) and
chroma proposed by CIE94 and those proposed by the
BFD (h 5 45° and 135°)—against chroma.
CMC and BFD formulas, as a function of the chroma.

that the most important qualitative discrepancies between


these three CIELAB-based formulas concern the usefulness Figure 3 shows a relationship between CIE94 and CMC
of a lightness and a hue-dependence correction. These two or BFD, through the ratio of their weighting functions for
questions are currently being studied by the CIE TC 1-47. chroma W C . The values on the y-axis of this plot can be
In Fig. 1, the W L function for BFD (upper left) has been interpreted as the CMC or BFD color-differences corre-
computed after transforming the L BFD logarithmic scale; sponding to pairs of samples with a common color-differ-
assuming that the increments can be treated as differentials, ence of 1.0 CIE94 units, having only a chroma component
(Y/Y O ) is greater than 0.008856 and (1.5/Y O ) is negligi- (i.e., DL* 5 DH* 5 0), and with different average
ble.14 In this figure, a medium chroma (C* 5 50) has been chroma. As shown in this example, a constant color-differ-
selected for the plot of W h against hue-angle (lower left), ence of 1.0 CIE94 units is equal to CMC color-differences
and a 45° hue-angle for the one of W h against C* (lower in a relatively wide range (approximately 1.1–1.6 CMC
right). This last hue-angle value has been selected, because units). Of course, this is a particular example for pairs of
CMC and BFD have nearly the same value of W h at C* 5 samples with only chroma differences; there is no general
50 (before applying the scaling factors leading to Figure 1), analytical relationship between the color differences com-
which, in addition, is close to 1.0. puted from CIE94 and CMC or BFD, as usually happens for
From the values on the y-axes of the plots shown in Fig. any other pair of color-difference formulas.
1, it should be noted that in general the corrections to the Finally, Fig. 4 shows the differences between the radially
three components of the CIELAB color-differences are of
different magnitudes: for example, the weighting function
for chroma differences (upper right) seems to be the most
important correction to CIELAB for any of our three for-
mulas, while the hue dependence in the weighting function
for hue in CMC or BFD (lower left) seems to be a minor,
although not negligible, correction factor. In addition, we
also observe that the slope of the weighting function for
lightness (upper left) is substantially higher for BFD than
for CMC, as a consequence of the common origin used in
this plot (the scaling factor used for BFD was 4.18, as
shown in Table I).
Although from Fig. 1 some of the weighting functions of
our CIELAB-based formulas appear to be relatively similar,
it should be noted that, in general, these formulas are
different, as illustrated by the next two examples.
Figure 2 shows the ratio between the weighting functions
for chroma and hue for different formulas, which is always
FIG. 4. Rotation of the major axis of the ellipses predicted
greater than 1.0. Although this ratio is an increasing func- by BFD with respect to the a*b* radial orientation, as a
tion of chroma for CIE94, however, for either CMC or BFD, function of the hue-angle (a medium chroma C* 5 50 has
it is nearly constant from a relatively low value of chroma. been assumed).

Volume 25, Number 1, February 2000 51


FIG. 5. Performance of different color-difference formulas for several experimental datasets from old and recent literature,
from the values of our parameter of isotropy S2% (see text). Lower values indicate better performance.

oriented ellipses predicted by CIE94 or CMC in a*b*, and interesting samples in practical applications. The method
the rotated ellipses predicted by BFD, as a consequence of employed to compare the performance of the different for-
its additional crossed term. The orientations of the BFD mulas analyzed here has been previously applied by us23–25
ellipses have been computed from the CIELAB metric and follows the steps described below.
coefficients associated with this formula.15 The highest ro- Since the pioneering work carried out by MacAdam in
tations of the BFD ellipses are in the blue and red zones, and 1942,17 contours of equally perceived color-differences DV
their magnitude is slight for the medium chroma selected around a given center are usually reported in the literature as
here (C* 5 50), although it affects the performance of the ellipses or ellipsoids in a color space. It would be desirable
formula, as shown below. for computed color-differences DE i between the center of
any of such ellipses (or ellipsoids) and points regularly
distributed on the periphery of the ellipse (or the surface of
Comparison with Experimental Results
the ellipsoid) also to have a constant value. Thus, in a
From the previous results, it should be concluded that the preliminary step, to evaluate the performance of a given
three CIELAB-based formulas considered here are not iden- formula, we compute for each ellipse/ellipsoid the mean and
tical. Additional evidence for the differences between these the coefficient of variation of the DE i color-differences,
formulas could be obtained by plotting the projections of which can be designated as ^DE& and S%, respectively.
some of their a*, b*, L* predicted ellipsoids (with semi- Next, for a set of experimental ellipses or ellipsoids, we
axes W L , W C , W h , assuming unit parametric factors), as compute the normalized standard deviation of ^DE& and the
illustrated by Berns.16 Thus, perhaps the key question normalized arithmetic mean of S%, designated as S1% and
should be to compare the performance of these CIELAB- S2%, respectively. Parameter S1% gives a measurement of
based formulas with respect to different experimental data- the differences between several regions of the color space
sets on color discrimination obtained under different exper- (that is a measurement of the “homogeneity”), while param-
imental conditions. eter S2% indicates the local uniformity in different direc-
For this purpose, we have selected four experimental tions around a color center (that is, a measurement of the
datasets obtained with visual colorimeters and reported in “isotropy”). Lower values of these parameters indicate bet-
classical literature,17–20 together with two recent outstanding ter performance of the formula used, the ideal values being
datasets from object colors,21,22 which constitute the most zero in both cases. In particular, low values of parameter

52 COLOR research and application


FIG. 6. Performance of different color-difference formulas for several experimental datasets from old and recent literature,
from the values of our parameter of homogeneity S1% (see text). Lower values indicate better performance.

S2% constitute the most important indicator for the good- CIELAB, in particular for object colors (second row), as it
ness of the color-difference formulas tested. might be expected, bearing in mind that they were used to
Figure 5 shows the values of parameter S2% from dif- develop these formulas. (3) CMC and CIE94 give similar
ferent color-difference formulas and four wide datasets. results, with the exception of the RIT-DuPont dataset,22
Classical results from old literature17–20 are shown in the which was the main dataset used for the development of
first row, distinguishing chromaticity (upper left) and color- CIE94. (4) The good results found by BFD for chromaticity
differences (upper right). This distinction between the re- differences (upper left) do not hold for color differences
sults obtained from ellipses and ellipsoids comes from the (upper right), perhaps as a consequence of the high slope in
significant differences found between them in a previous the weighting function for lightness adopted by this for-
work.24 Bearing in mind that there are different experimen- mula, as mentioned before.
tal conditions and visual scales in different experiments, in Conclusions rather similar to the previous ones can be
general it is not appropriate to merge the results from found from the values of our parameter S1%, as shown by
different experiments, as we have done in the first row of Fig. 6. Specifically, improvement upon CIELAB can be
Fig. 5. However, for our current comparative purposes, this seen again for the three CIELAB-based formulas, in partic-
is not misleading, because an initial normalization was ular for CIE94.
applied before computing S2% and S1% for each ellipse or
ellipsoid, and, in addition, it can be seen that the main
Statistical Significance of the Differences
conclusions from these plots are consistent with those found
from the individual experiments, which have been reported By applying the U-test,27 we have checked when our main
in previous works.23–25 parameter S2% has significantly different values in the four
In our opinion, the main conclusions from Fig. 5 can be previous datasets. This nonparametric test was preferred,
summarized as follows: (1) In general, CIELAB improves because no assumptions about the underlying distributions
CIELUV, although a quite similar performance can be seen are necessary. The formulas being statistically different at a
for chromaticity ellipses (upper left), which can be consid- 90% confidence level are connected by asterisks in Fig. 7.
ered to be in agreement with the 1976 recommendation for Among the main conclusions from Fig. 7, it can be noted
both formulas and previous works.26 (2) The three that CIE94 is significantly different than CIELAB for all the
CIELAB-based formulas (CMC, BFD, and CIE94) improve datasets (and the same can be stated for CMC and BFD for

Volume 25, Number 1, February 2000 53


FIG. 7. Results of the nonparametric U-test from the values of our parameter S2%. The asterisks indicate for each dataset
the formulas significantly different at a confidence level of 90%.

three of the four wide datasets analyzed in this figure). In 1. Robertson AR. Historical development of CIE recommended color
addition, we also observe that CIE94 and CMC are not difference equations. Col Res Appl 1990;15:167–170.
2. CIE Pub. 15.2, Colorimetry. 2nd Ed. p 33 (Note 9). Vienna: CIE
significantly different for any of our datasets at a 90%
Central Bureau; 1986.
confidence level (and even for any of the experimental 3. Clarke FJJ, McDonald R, Rigg B. Modification to the JPC79 colour-
works considered in the group of classical literature17–20). difference formula. J Soc Dyers Col 1984;100:128 –132.
BFD is significantly different than CIE94, except for the 4. Luo MR, Rigg B. BFD(,c) colour difference formula. Part I—Devel-
Luo and Rigg ellipses,21 probably because most of the opment of the formula. J Soc Dyers Col 1987;103:86 –94.
ellipses in this dataset were used for the development of 5. CIE Pub. 116-1995, Industrial colour-difference evaluation. Vienna:
CIE Central Bureau; 1995.
BFD. It should also be noted that CIELAB and CIELUV are
6. McDonald R, Smith KJ. CIE94 —A new colour-difference formula. J
significantly different for the two datasets having ellipsoids, Soc Dyers Col 1995;111:376 –379.
but not for those having ellipses. For future developments 7. McDonald R. European practices and philosophy in industrial colour-
and testing of color-difference formulas, the use of pairs of difference evaluation. Col Res Appl 1990;15:249 –260.
samples with an appropriate tridimensional distribution 8. Witt K. CIE guidelines for coordinated future work on industrial
around each center (such as made, for example, with those colour-difference evaluation. Col Res Appl 1995;20:399 – 403.
9. Robertson AR. CIE guidelines for coordinated research on colour-
used in the design of the RIT-DuPont experiment22), should
difference evaluation. Col Res Appl 1978;3:149 –151.
be emphasized.24 10. Kuehni RG. Industrial color difference: progress and problems. Col
Res Appl 1990;15:261–265.
11. Komatsubara H, Kawakami G. Research on the application of color-
CONCLUSION
difference formulas to industry. Stud Col 1988;35:6 –20.
In summary, the last few years have, in our opinion, seen 12. Witt K. Some recent developments in the evaluation of small color
differences. Proc 1998 OSA Ann Meet, Invited lecture SuF2; 1998. p
significant improvements in the prediction of the magnitude
56.
of perceived color-differences, by the use of several 13. Alman DH, Berns RS, Snyder GD, Larsen WA. Performance testing of
CIELAB-based formulas, in particular CIE94. Of course, color-difference metrics using a color tolerance dataset. Col Res Appl
much more research needs to be done within this topic, for 1989;14:139 –151.
example, by analyzing the influence of specific parametric 14. Melgosa M, Hita E, Poza AJ, Pérez MM. The weighting function for
factors, such as those affecting the optimal weighting func- lightness in the CIE94 color-difference model. Col Res Appl 1996;21:
tion for lightness differences.14 In any case, we believe that 347–352.
15. Melgosa M, Pérez MM, Hita E. Chromaticity-discrimination thresh-
the current structure of the CIE94 formula will be useful, at olds with aperture and object colors: experimental results and predic-
least in the near future, without disregarding that the current tions of some recent color-difference formulas. Appl Opt 1996;35:
use of the CIELAB space as the starting point could be 176 –187.
surpassed by advances in the knowledge concerning color 16. Berns RS. Deriving industrial tolerances from pass-fail and colorimet-
appearance and the physiology of our visual system. ric data. Col Res Appl 1996;21:459 – 472.
17. MacAdam DL. Visual sensitivities to color differences in daylight. J
Opt Soc Am 1942;32:247–274.
ACKNOWLEDGMENTS 18. Brown WRJ, MacAdam DL. Visual sensitivities to combined chroma-
ticity and luminance differences. J Opt Soc Am 1949;39:808 – 834.
The author is grateful to Mr. David Nesbitt for translating 19. Brown WRJ. Color discrimination of twelve observers. J Opt Soc Am
the original manuscript into English. 1957;47:137–143.

54 COLOR research and application


20. Wyszecki G, Fielder G. New color-matching ellipses. J Opt Soc Am metrics tested with an accurate color-difference tolerance dataset. Appl
1971;61:1135–1152. Opt 1994;33:8069 – 8077.
21. Luo MR, Rigg B. Chromaticity-discrimination ellipses for surface 25. Melgosa M, Pérez MM, Jiménez del Barco L, Hita E. A test of
colours. Col Res Appl 1986;11:25– 42. TC 1-29 color-difference formula using some classical datasets
22. Berns RS, Alman DH, Reniff L, Snyder GD, Balonon–Rosen MR. (Preliminary Report). Proc Eur Opt Soc Ann Meet. Univ Zaragoza
Visual determination of suprathreshold color-difference tolerances us- y Sociedad Española de Óptica, Zaragoza (España); 1993. p
ing probit analysis. Col Res Appl 1991;16:297–316. 37–38.
23. Melgosa M, Hita E, Romero J, Jiménez del Barco L. Some classical 26. Pointer MR. A comparison of the CIE 1976 colour spaces. Col Res
color differences calculated with new formulas. J Opt Soc Am A Appl 1981;6:108 –118.
1992;9:1247–1254. 27. Kruskal WH, Walis WA. Use of ranks in one-criterion variance
24. Melgosa M, Quesada JJ, Hita E. Uniformity of some recent color analysis. J Am Stat Assoc 1952;47:583– 621.

Volume 25, Number 1, February 2000 55

You might also like