A Colour Harmony Model for

Two-Colour Combinations

Li-Chen Ou,* M. Ronnier Luo

Department of Colour & Polymer Chemistry, University of Leeds, Leeds LS2 9JT, United Kingdom

Received 14 May 2004; revised 8 July 2005; accepted 29 October 2005

Abstract: This study investigates harmony in two-colour an accurate prediction.13–17 These studies can be divided
combinations in order to develop a quantitative model. A into two categories: those based on the orderly arrangement
total of 1431 colour pairs were used as stimuli in a psycho- of colours and those based on the interrelationship between
physical experiment for the visual assessment of harmony. colours.
These colour pairs were generated using 54 colours se- Studies based on orderly arrangement of colours included
lected systematically from CIELAB colour space. During those by Ostwald,1 Munsell,2 Itten,3 and Nemcsics.4 The
the experiment, observers were presented with colour pairs general idea behind these studies was that colours could
displayed individually against a medium gray background harmonize only when they were selected systematically
on a cathode ray tube monitor in a darkened room. Colour from a hue circle or from a specific “path” within an ordered
harmony was assessed for each colour pair using a 10- colour space. This is only an assumption, however, and
category scale ranging from “extremely harmonious” to seems to be true only if the selected colours are presented to
“extremely disharmonious.” The experimental results the viewer in an ascending or descending sequence of hue,
showed a general pattern of two-colour harmony, from lightness, or chroma. As Granville14 pointed out, “it is the
which a quantitative model was developed and principles smooth visual progression through colour space that gives
for creating harmony were derived. This model was tested the harmonious results, not that the colours bear some
using an independent psychophysical data set and the re- unique harmonious relationship to each other.”
sults showed satisfactory performance for model prediction. The studies based on the interrelationship of colours
The study also discusses critical issues including the defi-
included those by Goethe,5 Chevreul,6 Moon and Spen-
nition of colour harmony, the relationship between harmony
cer,7–9 Albers,10 and Chuang and Ou.11 The general idea
and pleasantness, and the relationship between harmony
was that colours could harmonize only when they were
and order in colour. © 2006 Wiley Periodicals, Inc. Col Res Appl, 31,
complementary or analogous (similar) in ether hue, light-
191–204, 2006; Published online in Wiley InterScience (www.interscience.
ness, or chroma. While the definitions of “complementary”
wiley.com). DOI 10.1002/col.20208
and “analogous” varied from study to study, they were
Key words: colour harmony; colour psychology; colour determined by considering the effect of colour contrast
order; colour pleasantness; colour combinations (in terms of colour difference) between each constituent
colour in a combination. The theories in this category
INTRODUCTION have been tested experimentally, but none of them was
acceptable in providing an accurate prediction of colour
Colour harmony has been of enormous significance in art harmony.15–17
and design. While the basic mechanisms of colour percep- As a result, some artists claimed that there were no laws
tion have been extensively studied, the semantic and aes- of colour harmony15 and that colour harmony was important
thetic aspects of colour—the very phenomena that give only for scientists.18 However, the poor performance of
colour its psychological power—are still poorly understood. existing theories was in fact due to the difficulties that the
There is a long tradition of research into colour harmo- earlier studies encountered. These difficulties included a
ny,1–11 but none of the studies was acceptable either in wide disagreement over the definition of colour harmony,
explaining the concept of colour harmony12 or in providing a large number of factors that may influence colour
harmony, and a lack of modern colour-science techniques
*Correspondence to: L. Ou (e-mail: [email protected]) for the earlier studies. Discussion of these issues is given
© 2006 Wiley Periodicals, Inc. below.

Volume 31, Number 3, June 2006 191

Definition of Colour Harmony preference,23–25 these techniques played an essential role in
colour measurement, colour specification, and data analysis.
There was wide disagreement over the definition of colour
Therefore, the present study was conducted based on these
harmony in the earlier studies. Key words in these defini-
techniques.
tions included “completeness,”5 “complementary/analo-
On the basis of the discussion above, this study hypoth-
gous,”6 “order,”1 “balance,”2 and “unambiguous.”7 While
esizes the following:
this diversity provides colour designers with more possibil-
ities for generating potentially harmonious colour schemes, ● there is a strong link between colour harmony and the
it makes it more difficult for scientists wishing to study
emotion “pleasantness;”
colour harmony. A clear, universal definition of harmony is ● colour harmony can be investigated using psychophysical
a fundamental part of any such research. Therefore, after
methods;
investigating a variety of existing definitions, the present ● there is a general pattern of colour harmony, which can be
study adopted that given by Judd and Wyszecki19:
quantified in terms of colour appearance attributes for
“When two or more colours seen in neighbouring areas
each constituent colour in a combination.
produce a pleasing effect, they are said to produce a colour
harmony.”
This definition implies a strong link between harmony METHODS
and the emotion “pleasantness” in colour. It agrees with
The aim of the study is to develop a quantitative model for
Granville’s later remarks, which describe colour harmony as
two-colour harmony. To achieve this, visual assessment of
the “colour usage that pleases people.”14 More recently,
harmony was carried out and the resulting psychophysical
Burchett20 suggested a similar definition of colour harmony:
data were collected for use in modeling. Based on the
“when two or more colours are brought together to produce
hypotheses described above, the experimental settings and
a satisfying affective response, they are said to be harmon-
procedures were determined, as described in the following.
ised.” Note that in these definitions, the two key words
pleasing and satisfying are both emotional terms. According
to Russell,21 pleasantness is an emotion of high pleasure and Colour Stimuli
medium arousal; satisfaction is an emotion of high pleasure
In the experiment, 1431 colour pairs were used as stimuli,
and low arousal. This suggests that colour harmony is
each consisting of two square colour patches placed side by
related strongly with the emotion of pleasure. Thus, the
side. These colour pairs included all possible two-colour
present study hypothesizes that despite the various ways of
combinations generated by 54 colours, which were sampled
generating potentially harmonious colour schemes, such as
systematically from CIELAB colour space.26 These colours
complementary and analogous, harmonious combinations
consisted of 49 chromatic colours and 5 achromatic colours.
will always give viewers a pleasing impression. It is further
The chromatic colours were made up of 7 hues: red, orange,
hypothesized that this impression can be measured using
yellow, green, cyan, blue, and purple, as shown in Fig. 1a;
psychophysical methods. Thus, the definition given by Judd
each hue consisted of seven tones: vivid, pale, dull, dark,
and Wyszecki was adopted in this study for assessing har-
light-grayish, grayish, and dark-grayish, as shown in Fig.
mony.
1b. The 5 achromatic colours included white, light gray,
medium gray, dark gray, and black.
Factors Affecting Colour Harmony These colour samples well covered Berlin and Kay’s 11
basic colour terms27: black, white, gray, red, orange, yellow,
It has been suggested 12,19,20,22 that colour harmony can be green, blue, purple, pink, and brown. Note that except for
influenced by many factors such as shape, size, texture, the white, black, and the 7 vivid colours, all these colour
number of colours, and the relative positions of colours in a samples fall into three equal-lightness and three equal-
combination. As the first part of a comprehensive research chroma categories, as illustrated in Fig. 1b. This has the
project of colour harmony, the present study focuses on advantage that the experimental data can be analyzed in a
simplified conditions; it uses square colour patches as the systematic fashion.
stimuli presented against a uniform gray background on a Each colour stimulus was presented against a medium-
cathode ray tube (CRT) monitor in a darkened room. While gray background (L* ⫽ 50) on a CRT monitor using the
the findings of the present study may not necessarily apply GOG model28 to transform CIE tristimulus values into RGB
to conditions that deviate from this work, they will serve as primaries for display. Table I shows the specifications of
a fundamental basis for future research into colour harmony. each colour stimulus, obtained from measurement results
using a Minolta CS-1000 telespectroradiometer.
Modern Colour-Science Techniques
Observers
Modern colour-science techniques such as colour measure-
ment, uniform colour spaces, and colour difference formu- Seventeen observers naive to the present study, including 11
lae were unavailable for earlier studies of colour harmony. males and 6 females, participated in the experiment. All of
According to recent research into colour emotion and colour the observers were Chinese students at the University of

192 COLOR research and application

 FIG. 1. The 54 colour samples in (a) the CIELAB colour space and (b) the equal-hue plane.

Derby, UK. Among them were 4 from the School of Art & indicating the degree of harmony for each colour pair; the
Design and 13 from the Colour & Imaging Institute. Each higher the score, the more harmonious the colour pair appeared
observer has passed Ishihara’s Tests for Colour Deficiency.29 to the observers. These scores were analyzed using CIELAB.

Experimental Procedures Inter-observer and Intra-observer Agreement

During the experiment, each observer was presented with To examine the consistency of observer responses, experi-
colour stimuli on a CRT monitor at a viewing distance of 50 mental data were evaluated in terms of inter-observer and
cm and a visual field size of 10° for each colour stimulus. intra-observer agreement. The inter-observer agreement is
As shown in Fig. 2, each colour pair was displayed in the concerned with how well the observers agreed with each
center of the screen with 10 numbered buttons arranged in other. This agreement was determined by averaging the root
a line underneath. These buttons were divided into two mean squared (RMS) between each observer’s data and the
sides, i.e., “harmonious” and “disharmonious.” Each side overall mean:

冑 冘共 x ⫺ x៮ 兲
consisted of five categories labeled using numbers from 1 to 2
5, representing “just perceptibly,” “quite,” “moderately,” i i
i
“very,” and “extremely,” respectively (see the Appendix). RMS ⫽ , (1)
By pressing one of these buttons, observers indicated their N
judgements as to how harmonious the colour pair appeared. where xi represents the category number given by an ob-
All of the colour pairs were presented individually in a server for stimulus i, x៮ i represents the mean category num-
random sequence. In each pair, the left–right position of ber of all observers for stimulus i, and N is the number of
each constituent colour was also randomized. stimuli.
The experiment was divided into six sessions, each pre- As a result, the agreement was slightly better among females
senting 239 colour pairs (except for Session 6, which pre- (with an RMS value of 2.05) than among males (2.35).
sented only 237 pairs). In each session, observers assessed Intra-observer agreement, on the other hand, is concerned
harmony, each taking approximately 15 min. To examine with the repeatability of each single observer’s colour har-
repeatability of observer responses, Session 1 was dupli- mony judgements. This agreement was determined by av-
cated once after all six sessions had been finished. This eraging the RMS for each observer’s experimental results
replicated session is called Session 7. For each observer, the between Sessions 1 and 7. This was done by replacing x៮ i of
whole experiment took about 3 weeks to complete. Eq. (1) with yi where xi and yi are category numbers for
stimulus i obtained in Sessions 1 and 7, respectively. The
results show that the intra-observer agreement was slightly
RESULTS
better among females (with an RMS value of 2.18) than
Torgerson’s Law of Categorical Judgement30 was used in this among males (2.57).
study to transform the experimental data onto an interval scale. These results were then compared with the authors’ pre-
The resulting scale values are called colour harmony scores, vious studies into colour emotion and colour prefer-

Volume 31, Number 3, June 2006 193

TABLE I. Specifications of the 54 colour samples used in this study.
Sample Tone Hue L* a* b* C* h

1 Red 35.0 46.4 23.2 51.8 27

2 Orange 56.7 30.6 52.1 60.4 60
3 Yellow 79.2 2.6 61.3 61.4 88
4 Vivid Green 34.1 ⫺25.8 17.0 30.8 147
5 Cyan 34.6 ⫺17.8 ⫺8.9 19.9 206
6 Blue 35.8 ⫺2.4 ⫺33.4 33.5 266
7 Purple 35.7 36.9 ⫺23.6 43.8 327
8 Red 79.8 21.9 10.9 24.5 26
9 Orange 79.6 13.1 22.2 25.8 59
10 Yellow 79.4 1.4 25.2 25.3 87
11 Pale Green 79.5 ⫺19.4 12.6 23.1 147
12 Cyan 80.0 ⫺13.3 ⫺5.9 14.5 204
13 Blue 80.6 ⫺1.9 ⫺24.3 24.4 265
14 Purple 80.3 19.5 ⫺12.1 23.0 328
15 Red 56.9 21.8 10.6 24.2 26
16 Orange 56.5 12.8 21.9 25.3 60
17 Yellow 56.4 1.0 25.1 25.1 88
18 Dull Green 56.6 ⫺19.5 12.6 23.2 147
19 Cyan 57.0 ⫺12.9 ⫺6.3 14.3 206
20 Blue 57.6 ⫺2.0 ⫺24.4 24.5 265
21 Purple 57.3 19.3 ⫺12.3 22.9 327
22 Red 34.4 21.0 10.4 23.4 26
23 Orange 34.0 12.1 21.3 24.5 60
24 Yellow 34.0 1.1 24.3 24.3 87
25 Dark Green 34.0 ⫺18.6 12.3 22.3 146
26 Cyan 34.4 ⫺12.6 ⫺5.8 13.9 205
27 Blue 35.0 ⫺1.4 ⫺24.1 24.2 267
28 Purple 34.8 18.4 ⫺11.8 21.9 327
29 Red 79.8 11.1 5.8 12.5 28
30 Orange 79.7 6.7 11.3 13.1 60
31 Yellow 79.6 0.9 12.9 13.0 86
32 Light grayish Green 79.6 ⫺9.7 6.6 11.7 146
33 Cyan 79.9 ⫺6.8 ⫺2.8 7.3 203
34 Blue 80.0 ⫺0.9 ⫺12.0 12.0 266
35 Purple 79.8 9.8 ⫺6.1 11.6 328
36 Red 56.9 11.0 5.3 12.2 26
37 Orange 56.6 6.2 11.1 12.7 61
38 Yellow 56.6 0.8 12.4 12.4 87
39 Grayish Green 56.6 ⫺10.0 6.5 11.9 147
40 Cyan 56.8 ⫺6.6 ⫺3.3 7.3 207
41 Blue 57.1 ⫺0.9 ⫺12.3 12.4 266
42 Purple 57.1 9.4 ⫺5.9 11.2 328
43 Red 34.1 10.5 5.3 11.7 27
44 Orange 34.1 6.1 10.7 12.3 60
45 Yellow 34.0 0.6 12.3 12.3 87
46 Dark Grayish Green 34.0 ⫺9.6 6.3 11.5 147
47 Cyan 34.2 ⫺6.3 ⫺2.9 7.0 204
48 Blue 34.6 ⫺0.9 ⫺12.0 12.0 266
49 Purple 34.6 9.2 ⫺5.6 10.8 329
50 Black 2.3 ⫺0.1 ⫺0.4 0.4 262
51 Dark Gray 34.2 ⫺0.2 0.2 0.3 124
52 Achromatic Mid Gray 56.9 ⫺0.1 0.3 0.3 115
53 Light Gray 80.0 3.2 ⫺0.9 3.4 345
54 White 100.0 ⫺0.1 0.3 0.3 106

ence.23–25 The comparison results showed that all of these Testing Existing Colour Harmony Principles
studies, including the present research, had the same level of
Some of the existing colour harmony principles were tested
consistency of observer responses for both inter-observer
using the present experimental data. These principles in-
and intra-observer agreements. Note that in these studies,
cluded the following:
the number of observers participating in each experiment
ranged from 17 to 31. This suggests that the consistency of
visual responses in such studies does not vary significantly ● Complementary hue1–7,10,11— colours can harmonize if
once there are sufficient observers (in this case approxi- they are complementary in hue. In the present study, the
mately 17). following hue pairs were regarded as being complementary:

194 COLOR research and application

 opposite principle, perhaps called unequal lightness, would
have a CD value of 0.60 (i.e., 1 ⫺ 0.40 ⫽ 0.60), which was
acceptable in terms of harmony prediction in this test. As a
result, equal hue and unequal lightness were regarded as
perhaps the only principles that could predict two-colour
harmony to an acceptable extent, with CD values of 0.61
and 0.60, respectively.

MODELING COLOUR HARMONY

As shown in Table II, a number of existing colour harmony

principles were tested using the present experimental data.
In fact, many other “likely principles” were also examined,
such as complementary lightness/chroma, neighbouring
hue, and equal saturation. Among these principles, however,
FIG. 2. Screen layout of the experiment. only equal hue and unequal lightness showed acceptable
performance for harmony prediction. Note that the CD
values of both principles were only around 0.60, indicating
red– cyan, yellow– blue, and green–purple (see Table I). The that the two principles still cannot accurately predict two-
orange colours were not tested here because there was no colour harmony.
complementary hue available for orange in this study. It was not until the method of “bubble charts” 23 was
● Equal hue 1,2,4,6,7— colours can harmonize if they share brought into data analysis that the present study started to
the same hue. determine whether a general pattern of colour harmony can
● Equal lightness1,2— colours can harmonize if they share be uncovered. Using this method, the study identified three
the same lightness. factors of colour harmony, namely chromatic difference,
● Equal chroma1,2— colours can harmonize if they share lightness effect, and hue effect.
the same chroma.
● Equal tone 1,2,6— colours can harmonize if they share the
same lightness and chroma. Chromatic Difference
The first factor identified is labeled chromatic difference,
The measure of correct decision (CD) was used to quan- which is a mixture of chroma and hue differences. This
tify each principle’s performance: factor was revealed using the method of bubble charts, as
冘c
i
i
shown in Figs. 3a to d. In these diagrams, bubbles show the
locations of the colour samples used in this study in
CD ⫽ , (2) CIELAB colour space. The size of each bubble indicates the
N
colour harmony score for a colour pair generated by a
where ci represents the percentage of observers whose har- colour sample and one of the following four green colours:
mony judgements for colour pair i (in terms of the decision vivid, pale, dull, and dark greens. These four greens are
between “harmonious” and “disharmonious”) agreed with a colour samples 4, 11, 18, and 25, as shown in Table I. For
given colour harmony principle; N is the number of colour each of the green colours, three cross sections are presented
pairs. to show results for different lightness levels. There are four
The CD values range from 0 to 1. A value of 0 indicates types of bubbles in these diagrams: (1) gray, (2) white with
that no observer agreed with a specific principle; a value of a fine outline, (3) white with a thick outline, and (4) white
1 indicates that all observers agreed with this principle; a with a broken outline. Gray bubbles represent colour pairs
value of 0.50 indicates that half of observers agreed with the that have positive values for their colour harmony scores.
principle. The test results are shown in Table II. Among the White bubbles with fine outlines represent colour pairs that
five principles, equal hue shows the highest CD value, 0.61, have negative values. As mentioned earlier, for both bubble
suggesting that equal hue is the most likely principle to types, the bubble size (area) represents the absolute value of
create two-colour harmony. The equal lightness principle a colour harmony score; the larger the bubble size, the
had the lowest CD value, 0.40, suggesting that equal light- higher the value is in terms of the magnitude disregarding
ness is the least likely to do so. This also suggests that the the sign of the score. White bubbles with thick outlines

TABLE II. Test results of existing colour harmony principles in terms of CD values.
Complementary hue Equal hue Equal lightness Equal chroma Equal tone

CD value 0.53 0.61 0.40 0.56 0.53

Volume 31, Number 3, June 2006 195

FIG. 3. Bubble charts for colour pairs generated from one of the 54 colour samples and one of the following green colours:
(a) vivid green, (b) pale green, (c) dull green, or (d) dark green. For each green colour, three cross sections are presented to
indicate different lightness levels.

represent the locations of the four green colours. White the observers. Note that in the other two lightness sections,
bubbles with broken outlines represent the corresponding a few gray bubbles are presented near the corresponding
locations of each green colour shown at different lightness locations of the vivid green. This suggests that this vivid
sections. green can harmonize with a colour if they have different
In Fig. 3a, the vivid green is located in the low-lightness lightness but similar hue and chroma. This pattern was also
section, where only white bubbles having fine outlines are identified for all the other green colours, as shown in Figs.
shown at the same lightness level. This means that any 3b to d. It was even found for all the other colour pairs used
colour pair generated by this vivid green and a low-lightness in this study. This pattern had two components, one con-
colour tended to have a negative value for the colour har- cerned with the colour difference in lightness and the other
mony score, i.e., these colour pairs were disharmonious to the difference in chroma and hue. The former will be

196 COLOR research and application

FIG. 4. (a) The relationship between colour harmony score and chromatic difference. (b) Mean colour harmony scores with
95% confidence intervals. The curve is determined by Eq. (3).

discussed later in the next section. The latter is called in this number of gray bubbles with bigger sizes seen in the high-
study the effect of chromatic difference. This effect indi- lightness section. This suggests two tendencies for two-
cates that the colour harmony score tends to decrease as the colour harmony:
chromatic difference (a mixture of chroma difference and
hue difference) becomes larger between the constituent co- ● small lightness difference between the constituent colours
lours. This can be illustrated in Figs. 4a and b, where the in a colour pair tended to reduce the harmony in that pair
curve is determined by the following equation: and
● high lightness values of the constituent colours tend to
H C ⫽ 0.04 ⫹ 0.53tanh共0.8 ⫺ 0.045⌬C兲, (3)
enhance the harmony.
in which
The former is labeled the effect of lightness difference
1
⌬C ⫽ 关共⌬H *ab兲2 ⫹ 共⌬C*ab/1.46兲2 兴2 , (⌬L), as previously mentioned. The latter is labeled the
effect of lightness sum (Lsum). This lightness sum effect was
where HC represents colour harmony score; ⌬H*ab and ⌬C*ab also identified in the results shown in Fig. 3c. In this
are CIELAB colour difference values in hue and chroma, diagram, the gray bubbles in the high-lightness section are
respectively. greater in both their area and their number than in the
The effect of chromatic difference agreed with a conven- low-lightness section.
tional colour harmony theory called the equal hue principle. A somewhat clearer relationship between harmony and
This principle asserts that colours can harmonize if they lightness sum is shown in Fig. 6a, where the colour har-
share the same hue 1,2,4,6,7. mony score is plotted against lightness sum. The diagram
shows that the harmony score tends to increase with light-
ness sum. This trend can be represented by the curve shown
Lightness: Difference and Sum
in Fig. 6b, which was defined using the equation
In addition to the effect of chromatic difference (⌬C), the
results shown in Figs. 3a to d also suggest another colour H Lsum ⫽ 0.28 ⫹ 0.54tanh共 ⫺ 3.88 ⫹ 0.029Lsum) (4)
harmony factor, the effect of lightness difference (⌬L), as in which
previously mentioned. This factor indicates that the har-
mony of a colour pair tends to decline if the constituent L sum ⫽ L*1 ⫹ L*2 ,
colours share the same lightness. However, according to a where HLsum represents colour harmony score; L*1 and L*2
further analysis, this effect did not show a significant impact
on harmony as did the effect of chromatic difference. As
illustrated in Fig. 5, colour harmony score is plotted against
lightness difference. This diagram depicts wide scattering,
indicating no clear connection between lightness difference
and harmony. This implies there are still many factors other
than lightness difference that can affect harmony.
Another lightness-related factor was also identified in this
study and was labeled lightness sum (Lsum). This factor was
again revealed by means of the bubble charts method. As
shown in Fig. 3a, no gray bubble is presented in the low-
lightness section and only a few gray bubbles are shown in FIG. 5. Colour harmony score plotted against lightness
the medium-lightness section. However, there are a larger difference.

Volume 31, Number 3, June 2006 197

FIG. 6. (a) The relationship between colour harmony score and lightness sum. (b) Mean colour harmony scores with 95%
confidence intervals. The curve is defined by Eq. (4).

are lightness values of the constituent colours in a colour diction can be solved. The resulting Eq. (6) had significantly
pair. better predictive performance, as shown in Figs. 9a and b.
To ensure validity of the effect of lightness sum, mean
colour harmony scores of the following nine colour combi- H L ⫽ H Lsum ⫹ H⌬L (6)
nation types were examined: LL, LM, LH, ML, MM, MH, In the literature, the effect of lightness sum (Lsum) was
HL, HM, and HH. For each combination type, L represents never mentioned. The effect of lightness difference (⌬L), on
a low-lightness colour; M a medium-lightness colour; and H the other hand, agreed to some extent with findings by
a high-lightness colour. According to Eq. (4), these combi- Moon and Spencer7 and Chuang and Ou.11 The two studies
nation types should exhibit the following trend: HH ⬎ both claimed that small lightness difference led to colour
HM ⬎ HL, MH ⬎ MM ⬎ ML, and LH ⬎ LM ⬎ LL, as disharmony.
shown in Fig. 7a, where the greater-than sign stands for
“being more harmonious than”. However, the test results
showed a slightly different trend. As illustrated in Fig. 7b, Hue Effect
HH has a lower mean score than HM, and MM has a lower In the literature, hue has been the main focus of colour
mean score than ML. harmony studies.1–11 According to the experimental results,
The results shown in Fig. 7b indicate that Eq. (4) over- as shown earlier, hue had a significant impact on two-colour
predicts when the two constituent colours are of the same harmony in terms of the hue difference, i.e., the less the hue
lightness. This problem echoes the effect of lightness dif- difference (and likewise for chroma difference) between the
ference, i.e., a small lightness difference between constitu- constituent colours in a colour pair, the more likely it is that
ent colours in a pair reduces the harmony of that pair. This the colour pair appears harmonious. The present section
effect can be illustrated by Fig. 8, which is defined as describes a further investigation into the effect of hue on
H ⌬L ⫽ 0.14 ⫹ 0.15tanh共 ⫺ 2 ⫹ 0.2⌬L兲, (5) harmony, where the hue angle of each constituent colour
was treated as the main variable.
where As shown in Fig. 10a, the mean colour harmony scores
for each of the seven hues used in this study are plotted: red,
⌬L ⫽ 兩L* 1 ⫺ L* 2兩.
orange, yellow, green, cyan, blue, and purple (see Table I).
By combining Eqs. (4) and (5), the problem of overpre- In this diagram, “blue colour pairs” (i.e., the colour pairs

FIG. 7. (a) Predicted results from Eq. (4). (b) Visual results for mean colour harmony scores of colour pairs generated from
three lightness levels: high (H), medium (M), and low (L).

198 COLOR research and application

 in Fig. 10a was found to apply to the entire set of colour
samples except for yellows (with a hue angle of around
90°). The yellows of light-grayish, pale, and vivid tones
were all found to have extraordinarily high values for their
mean scores. Dark-grayish and dark yellows, on the other
hand, were both found to have extraordinarily low values.
Note that light-grayish, pale, and vivid were all high-light-
ness tones; dark-grayish and dark were both low-lightness
tones. This means that high-lightness yellows tended to
FIG. 8. A graphical representation of Eq. (5). harmonize with another colour in a two-colour combination
and that low-lightness yellows (i.e., khaki colours) tended
not to harmonize. This trend is labeled the “yellow effect”
containing a blue as a constituent colour) have the highest and is denoted by EY. It can be represented by an extreme
mean score and “red colour pairs” have the lowest. In value distribution of the lightness and hue angle, as defined
addition, the mean score tends to become higher as the hue in the equation
angle increases from 0 to 270°, after which it decreases until
360°. This trend can be illustrated by the curve shown in the E Y ⫽ 关共0.22L* ⫺ 12.8兲/10兴exp兵共90⬚ ⫺ hab 兲/10
same diagram. This curve is defined by the equation
⫺ exp关共90⬚ ⫺ hab 兲/10兴其, (8)
H S ⫽ ⫺ 0.08 ⫺ 0.14sin共hab ⫹ 50⬚兲 ⫺ 0.07sin共2hab ⫹ 90⬚兲,
where L* is the lightness value and hab is the CIELAB hue
(7) angle.
where hab is the CIELAB hue angle for a constituent colour This equation is illustrated in Fig. 11, where the peak
in a colour pair. height at hue angle 90° is determined by L*, i.e., the larger
This trend was further analyzed by dividing the results the L* value the higher curve peak. This equation was then
shown in Fig. 10a into the seven tones: vivid, pale, dull, combined with Eq. (7), as given below:
dark, light-grayish, grayish, and dark-grayish, as defined in
Table I. The results are shown in Fig. 10b, where the trend

FIG. 10. (a) Mean colour harmony scores for colour pairs
FIG. 9. Performance of model prediction by (a) Eq. (4) and that contain a colour of a specific hue. (b) The results shown
(b) Eq. (6). in (a) separated by seven tones.

Volume 31, Number 3, June 2006 199

 E C ⫽ 0.5 ⫹ 0.5 tanh共 ⫺ 2 ⫹ 0.5C*ab 兲

H S ⫽ ⫺ 0.08 ⫺ 0.14sin共hab ⫹ 50⬚兲 ⫺ 0.07sin共2hab

⫹ 90⬚兲

E Y ⫽ 关共0.22L* ⫺ 12.8兲/10兴exp兵共90⬚ ⫺ hab 兲/10

⫺ exp关共90⬚ ⫺ hab 兲/10兴其 .
The predicted values CH range from ⫺1.24 to 1.39 on an
interval scale. The larger the CH value, the more harmoni-
FIG. 11. The yellow effect (EY), defined by Eq. (8). The ous the colour pair would appear. In Fig. 12, the predicted
curve-peak height changes dynamically according to the colour harmony values are plotted against the visual results.
lightness value of the given colour. The model explains 71% of the total variance of the visual
results (R2 ⫽ 0.71).

H SY ⫽ EC(HS ⫹ EY). (9)

Significance of the Factors
where EC ⫽ 0.5 ⫹ 0.5tanh (⫺2 ⫹ 0.5 C*ab), serving as a
correction equation to illustrate that the hue effect (i.e., HS To see whether all three factors in this model are significant
⫹ EY) becomes less significant as the chroma decreases. in determining colour harmony values, each factor was
This equation predicts the mean colour harmony score for examined by the F-test. First, three incomplete models were
a single colour at a hue angle of hab and a lightness value of created by removing each factor from the original model.
L*. By combining the HSY values for both constituent The three incomplete models were labeled “HC excluded,”
colours of a colour pair, the mean colour harmony score for “HL excluded” and “HH excluded.” The colour harmony
that colour pair is predicted, as given in the equation values predicted by each incomplete model were then com-
pared with the visual results in terms of a statistical measure
H H ⫽ HSY1 ⫹ HSY2, (10) called variance. This measure was determined in this study
where HSY1 and HSY2 are the harmony values predicted by by calculating the square of RMS [see Eq. (1)] between the
Eq. (9) for the two constituent colours in a colour pair. predicted values and the visual results. Three variance val-
ues were obtained for each incomplete model. A variance
value was also calculated for the original model.
A Model of Colour Harmony By dividing each of the three variance values for the
The three colour harmony factors, chromatic effect (HC), incomplete models by the one for the original model, three
lightness effect (HL), and hue effect (HH), as defined in Eqs. F-test statistics were determined: 1.46 for HC excluded, 2.86
(3), (6), and(10), respectively, are assumed to be indepen- for HL excluded, and 1.74 for HH excluded. The three
dent of each other. This means that each factor is correlated statistics were all greater than the critical value 1.09 at a
linearly with colour harmony value if and only if the other significance level of 0.05. This indicates that all of the three
two are held constant. Therefore, the three factors can be factors, HC, HL and HH, have significant contribution to the
combined additively to form a two-colour harmony model, model given in Eq. (11). Note that among the three incom-
as shown in the following: plete models, HL excluded has the greatest F-test statistic,
2.86, suggesting that HL is the most dominant factor in the
CH ⫽ H C ⫹ HL ⫹ HH, (11) model. The F-test results are summarized in Table III.
where
H C ⫽ 0.04 ⫹ 0.53tanh共0.8 ⫺ 0.045⌬C兲
1
⌬C ⫽ 关共⌬H *ab兲 2 ⫹ 共⌬C *ab/1.46兲 2兴 2
H L ⫽ HLsum ⫹ H⌬L
H Lsum ⫽ 0.28 ⫹ 0.54tanh共 ⫺ 3.88 ⫹ 0.029Lsum)
in which Lsum⫽L*1⫹L*2
H ⌬L ⫽ 0.14 ⫹ 0.15tanh共 ⫺ 2 ⫹ 0.2⌬L兲
in which ⌬L ⫽ 兩L*1 ⫺ L*2 兩
H H ⫽ HSY1 ⫹ HSY2
FIG. 12. The model explains 71% of the total variance of
H SY ⫽ EC*共HS⫹EY) the visual results.

200 COLOR research and application

Testing the Model Performance TABLE IV. Colour samples used in Gurura’s study31
To ensure the general applicability of the model stated in Test colour L* C* h
Eq. (11), it was tested using an independent psychophysical
Black 32.4 1.5 85
data set accumulated by Gurura et al.31 This data set was Blue 51.1 40.0 272
obtained from Gurura’s visual assessment of colour har- Brown 49.5 23.7 59
mony using 30 colour pairs as the stimuli. These pairs were Cream 99.5 17.0 87
Gold 85.9 38.8 89
generated randomly from 16 colour samples and took the Green 66.3 44.7 157
form of paper cards. Each colour pair was presented under Gray 66.4 5.3 97
a D50 simulator in a viewing cabinet. For each colour pair, Lime 78.5 61.2 123
Orange 70.4 71.2 44
the constituent colours were presented side by side against Pink 84.5 37.6 6
a gray background with L* ⫽ 60. Table IV shows the Plum 41.5 27.0 9
CIELAB specifications of the 16 colour samples. The 30 Purple 42.0 27.7 297
Red 53.4 62.0 30
colour pairs used in Gurura’s experiment are summarized in Silver 87.4 3.1 136
Table V. White 96.2 2.1 23
The experiment was conducted with 12 observers; 7 were Yellow 90.8 79.5 85
from China and 5 were from the following countries: Aus-
tralia, Japan, Korean, Sri Lanka, and Zimbabwe. All of the ence) between the constituent colours in a colour pair may
observers were either students or staff members at the reduce the harmony of that pair.
University of Derby. During the experiment, each observer (d) Hue effect. Among various hues, blue is the one most
was asked to judge the 30 colour pairs using a five-category likely to create harmony in a two-colour combination; red is
scale: “strongly harmonious,” “slightly harmonious,” “neu- the least likely to do so. In addition, bright yellows more
tral,” “slightly disharmonious,” and “strongly disharmoni- often create harmony in two-colour combinations than dark
ous.” yellows (i.e., khaki colours).
As shown in Fig. 13, Gurura’s experimental results Figure 14 provides a simple illustration of what these
agreed well with the harmony values predicted by Eq. (11), principles mean. Suppose we are looking for a colour to be
with a correlation coefficient of 0.75. This suggests that the combined with a cyan in order to generate a harmonious
model stated in Eq. (11) can predict colour harmony accu- colour pair. As shown in Fig. 14, this cyan colour is located
rately under conditions that deviate from the current study. on the equal-lightness plane at L* ⫽ 50. According to
The differences between the experimental settings of the Principle (a), the colours lying on the vertical line that goes
two studies included the colour samples chosen, the display through this cyan colour (i.e., the line containing the blue
media (CRT colours versus object colours), light sources and the red arrows) are most likely to create harmony when
(D65 versus D50), and observers’ cultural backgrounds combined with this cyan. According to Principle (b), the
(Chinese versus multicultural). colours located in the direction of the blue arrow are better
choices than those following the red arrow. According to
Principle (c), the colours within the range of the yellow
Interpretations of the Model
arrows should be avoided as they are too close to the
The model stated in Eq. (11) can be interpreted as per the example cyan colour. Thus, colours on the line of the blue
following two-colour harmony principles: arrow and above the yellow-arrow region are the preferred
(a) Equal-hue and equal-chroma. Any two colours vary- solutions. This line is called a “path of harmony,” illustrat-
ing only in lightness tend to appear harmonious when com- ing that the most likely colours to harmonize with a given
bined together. colour are those having equal hue (called the dominant hue
(b) High lightness. The higher the lightness value of each in this study) and equal chroma but having a lightness value
constituent colour in a colour pair, the more likely it is that higher than that colour (about 15 ⌬E*ab higher). Note that
they will appear harmonious. Principle (d) is not used in this example, as it applies only
(c) Unequal lightness values. Small lightness variations when a colour-pair comparison is made between different
(i.e., less than around 15 units of CIELAB colour differ- dominant hues.

TABLE V. Colour pairs used in Gurura’s experiment31

TABLE III. The F test results for the three incomplete
models, HC excluded, HL excluded, and HH excluded. plum/blue yellow/white green/plum
white/red orange/cream blue/cream
pink/gray pink/green white/pink
Original HC HL HH
black/lime gray/purple plum/gray
model excluded excluded excluded
brown/blue cream/yellow blue/red
pink/plum blue/white gray/black
Variance 0.07 0.10 0.20 0.12
yellow/gray orange/green green/white
F test statistic 1.46 2.86 1.74
pink/purple cream/plum brown/cream
white/cream gray/brown plum/red
Critical value: 1.09 (␣ ⫽ 0.05)
red/yellow lime/white white/orange
Number of degrees of freedom: 1430

Volume 31, Number 3, June 2006 201

FIG. 13. Test results of the model stated in Eq. (11) using FIG. 14. A graphical representation of the model in CIELAB
Gurura’s psychophysical data.31 colour space.

DISCUSSION definition given by Judd and Wyszecki is adopted. Using

that definition, this study seeks to reveal the aesthetic aspect
In this study, a quantitative model for two-colour harmony of colour harmony with support of psychophysical evi-
was developed. This model was tested using an independent dence. Having said that, the above claim is still a hypothesis
data set and the results showed a satisfactory predictive and so must be further verified in future work. This will be
performance by the model. Note that the experimental set- conducted using physiological and neurological methods
tings for the test data differed from those used for the model together with emotion-measuring instruments such as facial
data in terms of colour samples, display media (CRT co- muscle movement sensors and brain functioning recorders.
lours versus object colours), viewing conditions (D65 ver-
sus D50), and observers’ cultural backgrounds (Chinese
Harmony and Order in Colour
versus multicultural). This suggests that there is a general
pattern of two-colour harmony, which may be consistent Order has long been the predominant approach to under-
across different media, viewing conditions, and cultures. standing colour harmony.12 According to conventional co-
This, however, must be further verified in future work. lour harmony theories—such as those by Ostwald,1 Mun-
A number of issues still need to be addressed here with sell,2 Itten,3 and Nemcsics4— colour order refers to any
regard to the validity of the methodology used for colour systematic selection of colours from an ordered colour
harmony research, as the following describes. space. In the present study, the 54 colour samples represent
a systematic selection of colours from CIELAB, as depicted
in Figs. 1a and b. According to these theories, colour pairs
Harmony versus Pleasantness
generated by these 54 colours should all, or at least the
The first and perhaps most debatable issue is the definition majority, have obtained positive colour harmony scores.
of colour harmony. The current study adopts the definition However, the experimental results showed an approxi-
given by Judd and Wyszecki19: “when two or more colours mately normal distribution of harmony scores, as illustrated
seen in neighbouring areas produce a pleasing effect, they in Fig. 15. This indicates that a systematic selection of
are said to produce a colour harmony.” As already men- colours does not necessarily lead to harmony.
tioned earlier, this definition implies a link between colour
harmony and “pleasing,” or pleasantness. The question is:
would this definition lead the observers to judge pleasant-
ness (which is a measure of emotion) rather than harmony
(which has been said to be a measure of aesthetics)? Before
answering this question, we must understand what it means
to say that colour harmony is a “measure of aesthetics.”
Many existing theories have defined colour harmony using
aesthetic terms such as complementary,6 order,1 balance,2
and unambiguity.7 In fact, these are not definitions of colour
harmony, but colour selection methods for generating po-
tentially harmonious colour schemes. In the present study,
we have assumed that regardless of which colour selection
method is applied, a harmonious combination can always FIG. 15. Frequency distribution of colour harmony scores
give viewers a pleasing impression. This is the reason the obtained from this experiment.

202 COLOR research and application

 Another definition of colour order was any selection of colour shapes, textures, and relative position of each
colours from a specific “path” within a colour space, such as constituent colour;
the principles of equal hue, equal lightness, or equal ● investigating the influences of culture, age, and gender on
chroma. In the present study, equal hue was the only exist- colour harmony;
ing colour harmony principle that showed good predictive ● colour harmony for combinations of more than two co-
performance, with a CD value of 0.61 (see Table II). lours;
According to the present study, however, four principles ● colour harmony for real-world objects under various con-
of two-colour harmony were developed, as described pre- texts.
viously under Interpretations of the Model. Based on these
principles, a “path of harmony” was identified (Fig. 14),
APPENDIX (INSTRUCTIONS TO OBSERVERS)
showing that the most likely colours to harmonize with any
given colour are those having equal hue and equal chroma In each experimental session, you will be presented with
but with a higher lightness value (about 15 ⌬E*ab higher). colour pairs on a computer monitor and your task is to
This path can also be seen as a “colour order” for two- assess the degree of harmony for each colour pair using the
colour harmony, although it seems to be too complicated in scale below.
comparison with conventional principles.

Validity of the Proposed Model

The model stated in Eq. (11) has been tested using an
independent data set31 and the results showed satisfactory
performance for model prediction of two-colour harmony.
Note, however, that the model was developed using uniform
colour patches placed side by side against a medium gray
background as the stimuli and that the test data set was also FIG. A1. Scale used by observers for harmony or dishar-
based on uniform colour patches with a medium gray back- mony.
ground. Therefore, this model may not necessarily apply to
all possible colour combination types, shapes, textures, and
background colours. Furthermore, this model is not appli- The numbers shown in this scale represent equal intervals
cable to combinations of more than two colours. More of “harmony” or “disharmony.” For instance, the difference
specifically, the model is applicable only to harmony in the in harmony between numbers 1 and 2 should be the same as
area of the boundary between two uniform colours, as between 3 and 4.
depicted in Fig. 16. Having said that, this model forms a Note that colour harmony is defined in this experiment as
foundation for advanced colour harmony models to be built follows:
from in our future work, which will include: When a colour combination produces a pleasing effect,
that combination is said to be harmonious, i.e., those colours
● investigating the relationships between colour harmony go together.
and emotions;
● comparing colour harmony judgments between different ACKNOWLEDGMENTS
forms of stimuli, such as colour patches, complex images,
and real-world objects; The authors thank Dr Guihua Cui (at the Department of
● exploring the effects of appearance and configurations on Colour & Polymer Chemistry, University of Leeds) for
colour harmony, such as different background colours, developing the experimental software for use in this study.
Special thanks go to Dr Peter A. Rhodes (in the same
department) for valuable comments and suggestions relating
to this work.

1. Ostwald W. The color primer. From the German edition “Die Farben-
fibel” of 1916. New York: Van Nostrand Reinhold; 1969.
2. Munsell AH. A grammar of color. From the original version of 1921.
New York: Van Nostrand Reinhold; 1969.
3. Itten J. The art of color. Translated by Ernst van Haagen from German
edition “Kunst der Farbe”. New York: Van Nostrand Reinhold; 1961.
4. Nemcsics A. Colour dynamics: environmental colour design. New
York: Ellis Horwood; 1993.
5. Goethe JW. Theory of colours. Translated by C. L. Eastlake (1840)
from German edition “Farbenlehre” of 1810. Introduction by D. B.
Judd. MIT Press; 1970.
FIG. 16. The model applies to harmony in the area of a 6. Chevreul ME. The principles of harmony and contrast of colors.
boundary between two uniform colours. Translated by C. Martel (1854) from the first French edition of 1839.

Volume 31, Number 3, June 2006 203

 Introduction and explanatory notes by F. Birren. New York: Van 21. Russell JA. A circumplex model of affect. J Pers Soc Psychol 1980;
Nostrand Reinhold; 1981. 39:1161–1178.
7. Moon P, Spencer DE. Geometric formulation of classical color har- 22. Hård A, Sivik L. A theory of colors in combination: a descriptive
mony. J Opt Soc Am 1944;34:46 –59. model related to the NCS color-order system. Color Res Appl 2001;
8. Moon P, Spencer DE. Area in color harmony. J Opt Soc Am 1944; 26:4 –28.
34:93–103. 23. Ou L, Luo MR, Woodcock A Wright A. A study of colour emotion and
9. Moon P, Spencer DE. Aesthetic measure applied to color harmony. J colour preference, part I: colour emotions for single colours. Color Res
Opt Soc Am 1944;34:234 –242. Appl 2004;29:232–240.
10. Albers J. Interaction of color. Revised edition. Yale University; 1975. 24. Ou L, Luo MR, Woodcock A Wright A. A study of colour emotion and
11. Chuang M, Ou L. Influence of a holistic color interval on color colour preference, part II: colour emotions for two-colour combina-
harmony. Color Res Appl 2001;26:29 –39. tions. Color Res Appl 2004;29:292–298.
12. Burchett KE. Color harmony. Color Res Appl 2002;27:28 –31. 25. Ou L, Luo MR, Woodcock A Wright A. A study of colour emotion and
13. Eysenck HJ. A critical and experimental study of color preferences. colour preference, part III: colour preference modeling. Color Res
Am J Psychol 1941;54:385–394. Appl 2004;29:381–389.
14. Granville WC. Color harmony: what is it? Color Res Appl 1987;12: 26. Commission Internationale de l’Éclairage (CIE). Technical Re-
196 –201. port 15:2004: Colorimetry, 3rd ed. Vienna: CIE Central Bureau;
15. Pope A. Notes on the problem of color harmony and the geometry of 2004.
color space. J Opt Soc Am 1944;34:759 –765. 27. Berlin B, Kay P. Basic color terms: their universality and evolution.
16. Granger GW. Aesthetic measure applied to color harmony: an exper- Berkeley: University of California Press; 1969.
imental test. J Gen Psychol 1955;52:205–212. 28. Commission Internationale de l’Éclairage (CIE). Technical Report
17. Yeh C, Chuang M. A study on the influence of color interval and area 122-1996: The relationship between digital and colorimetric data for
factor on color harmony of object colors. In: Proceedings of the Third computer-controlled CRT displays. Vienna: CIE Central Bureau;
Asia Design Conference, China–Japan–Korea Design Symposium, 1996.
1998. pp 813– 820. 29. Ishihara S. Ishihara’s tests for colour deficiency. 38 Plates Edition.
18. Birren F. Science and art, objective and subjective. Color Res Appl Tokyo: Kanehara Trading; 2003.
1985;10:180 –186. 30. Torgerson WS. Theory and methods of scaling. New York: John Wiley
19. Judd DB, Wyszecki G. Color in business, science and industry, 3rd ed, & Sons; 1958.
New York: John Wiley & Sons; 1975. 31. Gurura H, MacDonald LW, Dalke H. Background: an essential factor
20. Burchett KE. Color harmony attributes. Color Res Appl 1991;16:275– in colour harmony. In: Proceedings of the Interim Meeting of the
278. International Color Association (AIC), 2004. pp 239 –242.

204 COLOR research and application