The Scarcity of Universal Colour Names
Gunilla Borgefors
Centre for Image Analysis, Dept. of Information Technology, Uppsala University, Sweden
Keywords: Colour Names, Basic Colour Terms, Colour Perception, Deep Learning, Image Retrieval, Image
Annotation.
Abstract: There is a trend in Computer Vision to use over twenty colour names for image annotation, retrieval and to
train deep learning networks to name unknown colours for human use. This paper will show that there is
little consistency of colour naming between languages and even between individuals speaking the same
language. Experiments will be cited that show that your mother tongue influences how your brain processes
colour. It will also be pointed out that the eleven so called basic colours in English are not universal and
cannot be applied to other languages. The conclusion is that only the six Hering primary colours, possibly
with simple qualifications, are the only ones you should use if you aim for universal usage of your systems.
That is: black, white, red, green, blue, and yellow.
1 INTRODUCTION
Under ideal conditions, humans can perceive over
four million different colours. But how many can we
consistently name?
This question is important when computer vision
and pattern recognition interacts with humans and
colour is involved. Examples are using deep learning
(or other methods) to name unknown colours or,
more frequently, to use colour names for image
annotation and retrieval. A seminal paper is (van de
Weijer et al., 2009) that uses up to 22 English colour
names for these purposes. This paper has been much
cited and the many colour names in it are often used.
However, the colour names people use are not
universal and depend on many circumstances, e.g.,
the culture and language in which the person grow
up and her interests and hobbies. I will show that
there is very little consensus on how to name the
parts of the human colour gamut. What is azure to
one may be simply blue to another and be the same
colour as grass to a third. There is even clear evi-
dence that your mother tongue influences how your
brain processes colour, even when no naming is
involved.
In view of the non-consensus of colour names,
the trend of using many colour names in computer
vision and pattern recognition becomes problematic.
How meaningful are the deeply learned names
for people who were not themselves first trained to
“correctly” name the colours the researchers chose?
How well will images be retrieved if users are allow-
ed to freely use their own colour names or do not
understand the imposed English names? These are
the questions this paper aims to answer.
When discussing colour names most linguists use
the Munsell colour chips, Fig. 1. Colour hues are
divided into 40 steps, denoted 1-40 from red to red,
and for each hue there are eight lightnesses denoted
B-I. Saturation is always maximal. A greyscale in
ten steps, A-J, is also included. There are thus 330
chips.
Each colour name has a focus, i.e., the most typi-
cal Munsell chip for that colour; and a range, i.e. all
chips that can be called that colour.
Section 2 will briefly describe human colour
vision, historical and current colour linguistics, and
give an insight in personal colour names. Section 3
will describe three experiments that show that your
colour naming system influences how your brain
responds to colour stimuli. Finally, Section 4 will
contain conclusions and recommendations.
2 COLOUR NAMING
2.1 Human Colour Vision
Humans have three colour opsins in the retina,
S(hort), M(edium), and L(ong), where length refers
496
Borgefors, G.
The Scarcity of Universal Colour Names.
DOI: 10.5220/0006649004960502
In Proceedings of the 7th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2018), pages 496-502
ISBN: 978-989-758-276-9
Copyright © 2018 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
to light wave-length. S, M, and L cones are most
sensitive to purple, green, and greenish yellow,
respectively. The signals from the three opsins are
combined in the retina before they are sent on to the
brain in three channels: the M+L channel that
measures “lightness;” the M-L channel that mea-
sures “redness;” and the S-(M+L) channel that
measures “blueness.”
Ewald Hering studied human colour vision and
came up with six “Urfarben” (primary colours),
(Hering, 1878). He based these on his knowledge of
human perception and our “ghost” colour after
images. After having stared for a time at a red square
on a white background, we briefly see a ghostly
green square when the red one is suddenly removed.
The same is true for blue/yellow and white/black.
These three pairs are closely linked to the three
colour channels in the optic nerve, even though they
had not been discovered at that time.
2.2 Early Colour Linguistics
The study of colour names across time and cultures
started in the second half of the 19
th
century (Big-
gam, 2012). The first publication was (Gladstone
1858) (written while William Gladstone took a break
from British politics). He was intrigued by the scar-
city of colour terms, and the strangeness of those
used, in Homer’s works (e.g., green honey, violet
sheep). Lazarus Geiger, how knew very many anci-
ent and modern languages, made the same obser-
vation for Indian Vedic Poems in Sanskrit; Old
Testament in Hebrew; Quran in Arabic; and Sagas in
Icelandic (Geiger, 1869). The general conclusion,
made possible by the recent theory of evolution, was
that full colour vision had evolved very recently in
human history.
At this same time linguists, anthropologists, and
missionaries studied the languages of “primitive”
tribes and discovered that many of them had few and
confusing (to the Europeans) colour concepts, just as
the ancients. Therefore it was concluded that they
had not developed full colour vision either.
After a train accident in 1875 Sweden, that was
caused by a colour-blind driver mistaking the red
stop signal for green, Fritiof Holmgren (Holmgren,
1877), devised a simple test for colour blindness,
using differently coloured yarns. The test became
widely used, not only to improve traffic safety, but
also to determine that “primitive” people had perfect
colour vision.
The conclusion was that colour naming
completely is random and only depends on culture.
Research in colour linguistics stopped for a long
time. However, all humans that have all three colour
opsins have very similar colour perception. Should
that not lead to similarities in how the colour space
is divided?
2.3 Basic Colour Terms
In 1969 the area of colour linguistics restarted with
(Berlin and Kay, 1969). Their study was based on
many, mostly North and South American, languages.
A key concept is the Basic Colour Terms (BCT)
present in a language. A BCT should fulfil many
conditions, the most important being:
Meaning is not understood from itself (yellow vs.
lemon)
Cannot be contained in a larger category (red vs.
scarlet)
Can be used for everything (yellow vs. blonde)
Consensus among native speakers
Adapts to grammar (greener vs. “avocadoier”)
High frequency in speech and writing
Not a recent loan word
Short response time for naming
The very few languages with only two BCTs only
distinguish dark/cold colours from light/warm
colours. The Dani in New Guinea use such a lan-
guage, where mili is used for black, dark and cold
colours with focus in dark blue or dark green; and
mola is used for white, light and warm colours with
focus in pink or dark red, (Heider, 1972).
A much earlier study (Almqvist, 1883) of the
Chukchis of Siberia discovered that they had three
BCTs (although the term was not used then): nukin,
focus black, nidlikin, focus white, and tschetlju,
focus red.
Based on the conditions above, place names from
a road atlas (Motormännen, 2011) and plant names
from a flora (Lid, 1974) I consider my own language
Swedish to have eight BCTs. In place names the
colours black, green, and red are most common,
while for plants names they are yellow, white, and
blue. Grey and brown are also significantly present
in both categories. All other colours are very rare.
The Americans Berlin and Kay took English as
the standard for all languages and identified eleven
BCTs. In addition to the eight Swedish ones they list
pink, orange, and purple, see Fig. 1 again for English
BCT foci. All languages investigated so far has from
two to twelve BCTs eleven is not universal.
Regularities between languages that cannot be
explained if colour naming is completely random
were described in (Berlin and Kay, 1969), but some
of the conclusions have been superseded by later
studies including more languages, see the next Sub-
section.
The Scarcity of Universal Colour Names
497
2.4 The World Colour Survey
After the restart of colour linguistics researchers
went all over the world collecting colour naming
systems from as many languages as possible using
the Munsell chips in a standardized way. In 2009 the
World Color Survey was published (Kay et al.,
2009) (the same Paul Kay 40 years later), presenting
the results for 110, mostly pre-industrial, languages
from all over the world. For each language the focus
and range for all BCTs, as roughly agreed by a
number of informants, are marked on the chip chart.
Note that in the following I use English colour
names, like black, not as BCTs, but as rough indi-
cations of the colour ranges of the BCTs of different
languages.
Languages are divided into five stages, I-V,
depending on the number of BCTs. Stage I have two
BCTs and stage V six or more.
There are no Stage I languages in the World
Colour Survey, but several Stage II. The amazing
discovery is that all of them divide the chips in
roughly the same way, into black, white, and
red, just like the Chukchi. An example is shown in
Fig. 2. If colour naming was completely random,
this would not be the case.
Languages with four BCTs divide colour space
in different ways, but the divisions are still not ran-
dom. Except for a very few cases, all languages
follow one of five “paths,” A-E, see Fig. 3.
The most common Stage III is III
A
which have
white and red, but divides black into black and
grue.Including both green and blue colours in a
single BCT is so common that linguists have invent-
ed the term grue to denote it.
The other common possibility for Stage III,
III
BC
, is to divide red into red and yellow and
keep all dark, cold colours together. Indications
show that Proto-Indo-European, that was spoken on
the steppes north of the Black and Caspian Seas
about 10,000 years ago and is the ancestor of almost
all European and many Middle East and Indian lan-
guages of today, was Stage III
BC
(Biggam, 2012).
There are also a few Stage III
DE
languages where
the four BCTs are white, black, red, and “every-
thing else,” see Fig. 4.
In Stage IV
A
yellow is split from red. In Stage
IV
B
black splits into black and grue, while in
Stage IV
C
black is divided into black and green.
This is how the Vikings talked: all dark and blue
colours were called “blå,” i.e., blue. Thus they called
Africa “Blueland” where the “blue men” lived!
Figure 3: The five paths of colour naming that describes
almost all of the very many languages investigated. See
the text.
Stage IV
D
is IV
A
again, while IV
E
divides the
“everything else” into blue and yellow+green.
Finally all Stages join in V, where all the Hering
primaries are BCTs. With more BCTs there is so
little regularity that all attempts of finding “paths”
have been abandoned.
It should perhaps be noted that grey is a special
case, which can be present as a BCT in earlier
Stages than V.
Thus it does seem that human vision has an
influence on colour naming after all, as in all known
languages with enough BCTs the Hering primaries
are BCTs. Also, the borders between BCT ranges
tend to be in about the same places, when they are
present at all.
2.5 Blue
As already mentioned, many languages place green
and blue colours into one BCT, denoted grue. This is
the case for all Celtic languages, e.g., Welsh, Irish,
and Breton and many African ones, like Bété and
Zulu. Until very recently Japanese was also a grue
language, although it seems at present to be moving
towards separation. Of course green and blue can be
distinguished also in grue languages, by, e.g., saying
“grue like grass” and “grue like sky.”
A few languages with twelve BCTs exist, to the
English-speaking linguists’ surprise. The most cited
example is Russian, which divides English BCT
blue into синий (siniy) for dark blue and голубой
(goluboy) for light blue. As we will see later, this
divide actually influences Russian brain organisa-
tion. Linguists seem not completely convinced, but
Italian blu (dark blue) and azzurro (light blue)
should perhaps also be considered BCTs? Another
twelve BCT language is Hungarian that separates
English BCT red into dark red and light red.
The paper that inspired this one (van de Weijer et
al., 2009) lists blue, cyan, turquoise, and azure as
useful blue terms among their 22 colour names.
ICPRAM 2018 - 7th International Conference on Pattern Recognition Applications and Methods
498
(Strangely not indigo, that is traditionally one of the
colours of the rainbow.) Only blue is an English
BCT while the other three are included in the blue
BCT range. Having one colour term that includes
three others, and then considering all four separate is
both illogical and confusing.
2.6 Personal Colour Names
In the previous Subsection differences between lan-
guages were discussed. In this one the focus is on
differences between individuals. The colour names
used vary considerably within the same language
depending on age, sex, education, being mono- or
multi-lingual, urban or rural... So, how many colour
names does a person usually use? And how many
colours can they remember?
One way of sorting colour names is dividing
them into four groups: 1: BCT (green), 2: qualified
BCT (dark green), 3: qualified fancy (spruce green),
and 4: fancy (avocado), (Biggam, 2012).
It is often said that women are better at colours
than men. A 1977 American study (Rich, 1977)
using 25 Munsell chips showed that men and nuns
mostly use categories 1-2, while other women
mostly use categories 3-4. Other studies have
showed similar results. Just like the scholars in the
19
th
century thought the ancients colour-blind, this
has resulted in the myth that women have better
colour vision than men. This is as false as the earlier
myth. That colour blindness is more common in men
and a few rare women do have superior colour
vision (Jameson et al., 2001) does not affect the
waste majority of both sexes that has normal colour
vision.
Two more recent studies of young people have
showed no sex differences for university students
and teenagers in USA (Smith et al., 1995, Machen,
2002), so this difference between the sexes may be
disappearing. Earlier sex differences are thought to
reflect women’s greater interest in colours for hob-
bies and appearance.
An interesting study is (Derefeldt and Swartling,
1995) that addresses colour naming and colour
remembering. The purpose was to investigate how
many colours a person can remember correctly, if
they are allowed to name the colours themselves. In
the learning session the subjects were shown a set of
35 colour chips on a normalized colour display and
asked to name and remember each colour. In the
identification session only one colour was shown on
the screen and the subjects asked to name it. The
subjects could then correctly identify 30-35 colours,
with a median of 32.5.
This was a study where subjects were allowed
about half an hour for training. Other studies have
showed that with intense training it is possible to
remember many more colours (Derefeldt, 2007), but
it could be argued that this study shows the approxi-
mate number of remembered colours for the average
person. If everybody agreed on the same thirty
colours the over twenty in (van de Weijer et al.,
2009) would be reasonable, but this is not the case.
Also interesting are the names used by the sub-
jects in the study. It was done in Swedish, so they
are translated here and in the report. The only un-
qualified BCT (group 1) used was brown. Qualified
BCTs (2) were also rare, but a few like “dark yel-
low” and “medium grey” were used. The most
common category was qualified fancy (3), e.g., “flag
blue,” “pigeon blue,” “thunder blue,” and “midnight
blue.” The oddest was “Elvis green” (from the cover
of an Elvis record). There where also some fancy
names (4), such as “cerise,” “plum,” and “jade.”
Some names were obviously invented during the
test, showing that the subjects did not have enough
colour names in their vocabularies for the 35 chips.
Another observation is that the authors Derefeldt
and Swartling use the English BCTs without com-
ment even though they are not suitable for Swedish.
The most obvious problem is the authors’ attempts
to squeeze the many different colour terms for
“blue-red” used by the subjects (showing there is no
suitable Swedish BCT here) into the English BCT
purple. There were also problems with English BCT
pink, which one subject called “white-red.” Even
though this is a single case, further investigations
would probably show that many other modern
languages do not fit in the English Procrustes bed.
3 EXPERMENTS
3.1 Russian Blues
In this experiment (Winawer et al., 2007) English
and Russian subjects were shown a big blue square
with two smaller blue squares beneath it, see Fig.5
left. The task was to, as quickly as possible,
determine if the left or right small square was equal
to the big one. For English subjects, for whom all
colours were covered by the BCT blue, the decision
time was linearly dependent on the distance between
the colours: the more different the colours the faster.
For Russians, for whom the lighter blues are in BCT
goluboy, while the darker ones are in BCT siniy, the
results were different. They were faster when the
The Scarcity of Universal Colour Names
499
colours were in different BCTs even if the colour
distance was the same as within a BCT.
The test was repeated, but this time the subjects
had to repeat out laud a string of random numbers,
thus occupying the brain’s linguistic centre. This
time, also the Russians’ decision times were depen-
dent only on colour distance. The linguistic classifi-
cation of colours does determine how the colours are
perceived, even when no names are involved.
Figure 5: The set-up for the “Russian blues” (left) and the
“Brain divided” (right) experiments. From (Winawer et
al., 2007) and (Gilbert et al., 2006).
3.2 A Brain Divided
Most language processing is done in the left hemi-
sphere of the brain. Therefore, (Gilbert et al., 2006)
devised a test to see if there is a difference between
the hemispheres regarding colour perception.
As they had English subjects, they used the
green-blue BCT border instead of the siniy-goluboy
one, but otherwise the test was similar to the previ-
ous one. A ring of small green or blue squares, see
Fig. 5 right, were shown to one hemisphere of the
brain (this is not very difficult to arrange, as the left
half of the retinas of both eyes communicate with
the right hemisphere and vice versa). In this way,
they could investigate if language has an influence
on colour handling even when using only one lan-
guage. One square was different and the task was
again to determine if it was on the left or the right
side of the ring.
The results were clear. When the left hemisphere
saw the colour patches, the response time was short-
er when the two colours were on different sides of
the blue-green border, while no such difference was
noticeable for the right hemisphere. Again, proof
that your mother tongue influences how your brain
processes colour.
3.3 In Focus or Out
In the previous tests response time was used as a
measure of brain activity. In the experiment by (Tan
et al., 2008) brain activity was measured directly
using Magnetic Resonance Imaging of the subjects’
brains. The language used was Mandarin Chinese.
The experiment used six colours: red, green, and
blue, that are BCTs in Mandarin, close to their foci
and thus easy to name; and “brown,” “grey,” and
“beige” that are not Mandarin BCTs and difficult to
name, see Fig. 6.
First, two squares were shown for a split second
and the task was to determine if they were equal or
not by pressing a button. No language task was thus
involved. For the easy to name BCTs two small
areas in the cerebral cortex of the left hemisphere
were activated, while for the non-BCTs those areas
were inactive.
In the second test, the subject was shown one
colour and asked to say the name out loud. This
time, the two previously identified small areas in the
cortex lit up brightly, thus being very active, for all
colours. They are apparently the areas in the brain
most involved in colour naming. The test shows they
are involved in colour processing of BCTs, but not
other colours, even though no colour naming task is
going on!
The conclusion of the three experiments is that
speakers of different languages probably do perceive
colours slightly differently, after all. Gladstone and
his contemporaries did have a point, not because of
different colour vision, but because different cultures
and languages wires our brains differently.
Figure 6: The colours used in the Chinese MRI brain acti-
vation test. From (Tan et al., 2008).
4 CONCLUSIONS
I have showed that colour foci and ranges vary con-
siderably between cultures and languages, even
though the divisions between BCT ranges are similar
when they occur al all. I have also shown the colour
names we use vary a lot between individuals even in
same culture. The (Derefeldt and Swartling, 1995)
ICPRAM 2018 - 7th International Conference on Pattern Recognition Applications and Methods
500
study shows as a meta-result the dangers of using the
eleven English BCTs as universals.
The experiments cited shows that the colour
naming system we grew up with even has con-
sequences for brain organisation, thus influencing
our colour processing and perception.
When using colour names in computer vision
and pattern recognition applications involving per-
sons not specifically trained to recognize and name
the colours used, it is thus best to stay with very few
colours. The Hering primaries seem safe enough.
It should also be remembered that although a
trained person can remember and identify up to 500
colours, a stressed human can identify only three or
four colours (Derefeldt, 2007). The latter discovery
has, e.g., influenced the designs of displays in
fighter aircraft.
It is, by the way, probably no coincidence that
the basic heraldic colours used over the centuries in
Europe and elsewhere or (gold/yellow), argent
(silver/white), azure (blue), gules (red), sable (black)
and vert (green) are exactly the Hering primaries.
Experience had shown that these colours were easy
to distinguish, especially since only metal on colour
or vice versa was allowed.
Qualifying the six primaries using lightness and
saturation is probably also safe, e.g., “light green”
and “dark green,” or “vivid red” and “dull red.”
Repeating the questions from the Introduction:
How meaningful is it to use many colour names in
applications intended for general, untrained, users?
It will not be clear for everybody which Munsell
chips can be called “azure,” especially if you mother
tongue is not English and is English azure the
same as Italian azzurro? Which images will be re-
trieved using “flag blue?” Woad blue as the Swedish
or Ukrainian flag or indigo blue as the British or
French?
I suggest results can be disappointing and frus-
trating if the reality of the non-standardization of
colour names is not taken into account.
REFERENCES
Almquist, E., 1883. Studies of the colour sense of the
Chukchis. Brockhaus, Leipzig.
Berlin, B., Kay, P., 1969. Basic color terms: Their Univer-
sality and Evolution. University of California Press.
Biggam, C.P., 2012. The Semantics of Colour A histori-
cal approach, Cambridge University Press.
Derefeldt, G., 2007. Personal communication.
Derefeldt, G., Swartling, T., 1995. Colour concept retrie-
val by free colour naming. Identification of up to 30
colours without training. Displays, 16 (2), pp. 69-77.
Geiger, L. 1869. Der ursprung der sprache. Verlag der J.
G. Cottaschen buchhandlung, Stuttgart.
Gilbert, A.L., Regier, T., Kay, P., Ivry, R.B., 2006: Whorf
hypothesis is supported in the right visual field but not
the left. PNAS, 103 (2), pp. 489-494.
Gladstone, W. 1858. Homer’s Perception and use of
colour. In Studies of Homer and the Homeric Age,
Oxford U. Press.
Hering, E. 1878. Zur Lehre vom Lichtsinne. Akademie der
Wissenschaften in Wien.
Heider, E.R., 1972. Probabilities, Sampling, and Ethno-
graphic Method: The Case of Dani Color Names. Man
7 (3), pp. 448-466.
Holmgren, F. 1877. On colour blindness and its relation-
ship to railway traffic and the marine. Berling.
Jameson, K.A., Highnote, S.M., Wasserman, L.M., 2001.
Richer color experience in observers with multiple
photopigment opsin genes. Psychonomic Bulletin &
Review 8 (2), pp. 244261.
Kay, P., Berlin, B., Maffi, L., Merrifield, W.R., Cook, R.,
2009. The World Color Survey, CSLI Publ., Stanford,
CA, USA.
Lid, J., 1974. Norsk og svensk flora. Norske samlaget,
Oslo.
Machen, V., 2002. Color Naming by Boys and Girls.
Perceptual and Motor Skills 94, pp. 348-350.
Motormännen 2011. Svensk Vägatlas. Nordstedts,
Stockholm.
Munsell, A.H., 1915. Atlas of the Munsell Color System.
WadsworthHowland Press, USA.
Mylonas, D., MacDonald, L., 2016. Augmenting Basic
Colour Terms in English. Color Research and Appli-
cation, 41 (1) pp. 32-43.
Rich, E., 1977. Sex-Related Differences in Colour Voca-
bulary. Language and Speech 20 (4), pp. 404-409.
Smith, J.J., Furbee, L., Maynard, K., Quick, S., Ross, L.,
1995. Salience Counts: A Domain Analysis of English
Colour Terms. J. of Linguistic Anthropology 5 (2), pp.
203-216.
Tan, L.H., Chan, A.H.D., Kay, P., Khong, P.-L., Yip,
L.K.C., Luke, K.-K., 2008. Language affects patterns
of brain activation associated with perceptual decision.
PNAS, 105 (10), pp. 4004-4009.
van de Weijer, J., Schmid, C., Verbeek, J., Larlus D, 2009.
Learning Color Names for Real-World Applications,
IEEE T. on Image Processing, 18 (7), pp. 1512-1523.
Winawer, J., Witthoft, N., Frank, M.C., Wu, L., Wade,
A.R., Boroditsky, L., 2007. Russian Blues reveal
effects of language on colour discrimination. PNAS
104 (19), pp. 7780-7785.
The Scarcity of Universal Colour Names
501
Figure 1: The Munsell colour chips with the approximate foci of the English Basic Colour Terms (x).
Figure 2: Stage II language Bété with three BCTs. It is spoken by half a million people in the Ivory Coast.
Figure 4: Stage III
DE
language Karanjá with four BCTs, spoken by about 3000 people in central Brazil.
zεli
zεli
kpᴐ
fεε
lyby
ãrè
ura
lyby
ICPRAM 2018 - 7th International Conference on Pattern Recognition Applications and Methods
502