Color Space Transformations

Philippe Colantoni and Al

2004

1 Introduction
This document defines several color concepts and all the mathematic relations
used in ColorSpace. The first version of this document has been built 3 years
ago using several documents and unfortunately I did not keep all the references.
If you find in this document something you write, send me an email and I will
include your name in the acknowledgment section.

2 Generality
2.1 What is the difference between device dependent and
device independent color space?
A device dependent color space is a color space where the resultant color depends
on the equipment and the set-up used to produce it. For example the color
produced using pixel values of rgb = (250,134,67) will be altered as you vary
the brightness and contrast on your display. In the same way if you change
the red, green and blue phosphors of your monitor will have slightly different
characteristics and the color produced will change. Thus RGB is a color space
that is dependent on the system being used, it is device dependent. A device
independent color space is one where the coordinates used to specify the color
will produce the same color wherever they are applied. An example of a device
independent color space is the CIE L∗ a∗ b∗ color space (known as CIELAB and
based on the human visual system).

Another way to define a device dependency is to imagine an RGB cube

within a color space representing all possible colors (for example a CIE based
color space). We define a color by its values on the three axes, however the exact
color will depend on the position of the cube within the perceptual color space,
i.e. move the cube (by changing the set-up) and the color will change. Some
device dependent color spaces have their position within CIE space defined.
They are known as device calibrated color spaces and are a kind of half way house
between dependent and independent color spaces. For example, a graphic file
that contains colorimetric information, i.e. the white point, transfer functions,

1
and phosphor chromaticities, would enable device dependent RGB data to be
modified for whatever device was being used - i.e. calibrated to specific devices.

2.2 What is a color gamut ?

A color gamut is the area enclosed by a color space in three dimensions. It
is usual to represent the gamut of a color reproduction system graphically as
the range of colors available in some device independent color space. Often the
gamut will be represented in only two dimensions.

2.3 What is the CIE System ?

The CIE has defined a system that classifies color according to the HVS (the
human visual system). Using this system we can specify any color in terms of
its CIE coordinates.

The CIE system works by weighting the spectral power distribution of an

object in terms of three color matching functions. These functions are the sen-
sitivities of a standard observer to light at different wavelengths. The weighting
is performed over the visual spectrum, from around 360nm to 830nm in set
intervals. However, the illuminant, the lighting and the viewing geometry are
carefully defined, since these all affect the appearance of a particular color. This
process produces three CIE tristimulus values, XYZ, which are the building
blocks from which many color measurements are made.

2.4 Gamma and linearity

Many image processing operations, and also color space transforms that in-
volve device independent color spaces, like the CIE system based ones, must
be performed in a linear luminance domain. By this we really mean that the
relationship between pixel values specified in software and the luminance of a
specific area on the CRT display must be known. In most cases CRT have
a non-linear response. The luminance of a CRT is generally modeled using a
power function with an exponent, e.g. gamma, somewhere between 2.2 (NTSC
and SMPTE specifications) and 2.8. This relationship is given as follows:

luminance ∼ voltageγ

Where luminance and voltage are normalized. In order to display image

information as linear luminance we need to modify the voltages sent to the CRT.
This process stems from television systems where the camera and receiver had
different transfer functions (which, unless corrected, would cause problems with
tone reproduction). The modification applied is known as gamma correction
and is given below:

N ewV oltage = OldV oltage(1/γ)

2
 (both voltages are normalized and γ is the value of the exponent of the
power function that most closely models the luminance-voltage relationship of
the display being used.)
For a color computer system we can replace the voltages by the pixel values
selected, this of course assumes that your graphics card converts digital values
to analogue voltages in a linear way. (For precision work you should check this).
The color relationships are:

R = a.(R0 )γ + b G = a.(G0 )γ + b B = a.(B 0 )γ + b

where R0 , G0 , and B 0 are the normalized input RGB pixel values and R, G,
and B are the normalized gamma corrected signals sent to the graphics card.
The values of the constants a and b compensate for the overall system gain and
system offset respectively (essentially gain is contrast and offset is intensity).
For basic applications the value of a, b and γ can be assumed to be consistent
between color channels, however for precise applications they must be measured
for each channel separately.

3 Tristimulus values
3.1 The concept of the tristimulus values
The light that reaches the retina is absorbed by three different pigments that
differ in their absorption spectra. Relative absorption spectra of the short-
wavelength cone (blue) s(λ), middle-wavelength cone (green) m(λ) and long-
wavelength cone (red) l(λ) can be see figure 1.

Figure 1: Human cones (and rods) absorption spectra

If one pigment absorbs a photon which leads to its photoisomeration the in-
formation about the wavelength of the photon is lost (principle of univariance).
Lights of different wavelengths are able to produce the same degree of isomera-
tions (if their intensities are adjusted properly) and consequently produce equal
sensations. The probability of isomeration of one pigment is not correlated with

3
the probability of isomeration of another pigment. The probability of isomer-
ation is just determined by the wavelengths of the incident photons. If we do
not think about the influences of the spatial and temporal effects that influence
perception, the sensation of color is determined by the number of isomerations
in the three types of pigments. Therefore colors can be described by just three
numbers, the tristimulus values, independent of their spectral compositions that
lead to these three numbers.

3.2 The Tristimulus Values

The tristimulus values T for a complex light I(λ) (light that is not monochro-
matic) can be calculated for the specific primaries P with their corresponding
color matching functions Pi (λ):
Z Z Z
T1 = P1 (λ).I(λ).dλ T2 = P3 (λ).I(λ).dλ T3 = P2 (λ).I(λ).dλ
λ λ λ

3.3 Consequences
1. The amount of excitation of the three pigment types for a complex light
stimulus I(λ) can be calculated:

Z Z Z
Sexc = s(λ).I(λ).dλ Mexc = m(λ).I(λ).dλ Lexc = l(λ).I(λ).dλ
λ λ λ

2. The amount of excitation of each pigment type to a stimulus P (λ) can be

calculated:
Z Z Z
Sexc = s(λ).P (λ).dλ Mexc = m(λ).P (λ).dλ Lexc = l(λ).P (λ).dλ
λ λ λ

Each primary has a defined overlap in the absorption spectra of the three
pigments and consequently leads to a defined sensation. An increase in
intensity of one primary reduces to the multiplication with a scalar for
each pigment. Now, because of the principle of univariance, we can add
the influences of the three primaries to the resulting excitations of the
three pigment types.
3. Because of 2, there exists a linear transformation between the tristimulus
values of a set of primaries and the color space formed by the isomerations
of the cone pigments.
Tristimulus values describe the whole sensation of a color. There exist a lot
of other possibilities to describe the sensation of color. For example it is possible
to use something equivalent to cylinder coordinates where a color is expressed
by hue, saturation and luminance. If the luminance or the absolute intensity of a
color is not of interest then a color can be expressed in chromaticity coordinates.

4
3.4 Chromaticity Coordinates
In order to calculate the tristimulus values T of a light stimulus due to a set of
primaries we need to know the spectral shape of the color matching functions
and of the stimulus. The tristimulus values are calculated by the integration of
the product of the color matching function and the stimulus over the wavelength.
The tristimulus values describe the sensation of the stimulus due to the set of
primaries including the absolute intensities of the three primaries needed to
match the stimulus. Of course the luminance of that stimulus could be varied
without changing the hue and saturation of the stimulus. This is reflected in the
chromaticity coordinates c that form a two dimensional space thus luminance
is ignored.

T1 T2 T3
c1 = c2 = c3 =
T1 + T2 + T3 T1 + T2 + T3 T1 + T2 + T3

It is not necessary to mention the third coordinate because:

c1 + c2 + c3 = 1

3.5 Spectrum locus

We can read the tristimulus values for the spectral colors as the values of the
color matching functions. For complex light stimuli we would have to integrate.
After that we can calculate the chromaticity coordinates out of the tristimulus
values. In CIE space the tristimulus values are called X, Y and Z, the chromatic-
ity coordinates are called x and y. The curve of the chromaticity coordinates of
the spectral colors is called the spectrum locus (fig. 2).
The straight line connecting the blue part of the spectrum with the red part
of the spectrum does not belong to the spectral colors, but it can be mixed out
of the spectral colors just as all colors inside the spectrum locus.

4 Color spaces definitions

Color is the perceptual result of light in the visible region of the spectrum, hav-
ing wavelengths in the region of 380 nm to 780 nm. The human retina has three
types of color photoreceptor cells cone, which respond to incident radiation with
somewhat different spectral response curves. Because there are exactly three
types of color photoreceptor, three numerical components are necessary and
theoretically sufficient to describe a color.

Because we get color information from image files which contain only RGB
values we have only to know for each color space the RGB to the color space
transformation formulae.

5
 Figure 2: CIE 1931 xyY chromaticy diagram

4.1 Computer Graphic Color Spaces

Traditionally color spaces used in computer graphics have been designed for
specific devices: e.g. RGB for CRT displays and CMY for printers. They are
typically device dependent.

4.1.1 Computer RGB color space

This is the color space produced on a CRT display when pixel values are applied
to a graphic card or by a CCD sensor (or similar). RGB space may be displayed
as a cube based on the three axis corresponding to red, green and blue (see
fig. 3(a)).

4.1.2 Printer CMY color space

The CMY color model stands for Cyan, Magenta and Yellow which are the com-
plements of Red, Green and Blue respectively. This system is used for printing.
CMY colors are called ”subtractive primaries”, white is at (0.0, 0.0, 0.0) and
black is at (1.0, 1.0, 1.0). If you start with white and subtract no colors, you get
white. If you start with white and subtract all colors equally, you get black (see
fig. 3(b)).

4.2 CIE XYZ and xyY color spaces

The CIE color standard is based on imaginary primary colors XYZ i.e. which
don’t exist physically. They are purely theoretical and independent of device-
dependent color gamut such as RGB or CMY . These virtual primary colors
have, however, been selected so that all colors which can be perceived by the

6
 (a) RGB color space (b) CMY color space

Figure 3: Visualization of RGB and CMY color spaces

human eye lie within this color space.

The XYZ system is based on the response curves of the three color receptors
of the eye’s. Since these differ slightly from one person to another person, CIE
has defined a ”standard observer” whose spectral response corresponds more or
less to the average response of the population. This objectifies the colorimetric
determination of colors.
XYZ (fig. 4(a)) is a 3D linear color space, and it is quite awkward to work
in it directly. It is common to project this space to the X + Y + Z = 1 plane.
The result is a 2D space known as the CIE chromaticity diagram (see fig. 2).
The coordinates in this space are usually called x and y and they are derived
from XYZ using the following equations:
X Y Z
x= y= z= (1)
X +Y +Z X +Y +Z X +Y +Z
As the z component bears no additional information, it is often omitted.
Note that since xy space is just a projection of the 3D XYZ space, each point
in xy corresponds to many points in the original space. The missing information
is luminance Y. Color is usually described by xyY coordinates, where x and y
determine the chromaticity and Y the lightness component of color (fig. 4(b)).

4.2.1 RGB to CIE XYZ Conversion

There are different mathematical models to transform RGB device dependent
color to XYZ tristimulus values. Conversion from RGB to XYZ can take the
form of a simple matrix transformation (equ. 2) or a more complex transfor-
mation depending of the hardware used (e.g. to acquire or to display color
information).

7
 In this section we will define how to compute a linear transformation model.
This model may by a correct approximation for CCD sensor (RGB to XYZ
transform) and CRT display (XYZ to RGB transform). But do not forget, this
is only an approximate model.
   
R X
 G  = A.  Y  (2)
B Z
We can use for example this transformation:
     
R 3.06322 −1.39333 −0.475802 X
 G  =  −0.969243 1.87597 0.0415551  .  Y 
B 0.0678713 −0.228834 1.06925 Z

and the inverse transform simply uses the inverse matrix.

   −1  
X 3.06322 −1.39333 −0.475802 R
 Y  =  −0.969243 1.87597 0.0415551  .  G 
Z 0.0678713 −0.228834 1.06925 B

This is all very useful, but the interesting question is ”Where do these num-
bers come from?”. Figuring out the numbers to put in the matrix is the hard
part. The numbers depend on the color system of the output device we are
using. The important parts of a color system are the x and y chromaticity
coordinates and the luminance component of the primaries (xyY ). However, if
we don’t know the Y values, which is often the case, then we have a problem.
However, we can solve this problem if we know the chromaticity coordinates of
the white point. In the previous example we have used a the color system which
has the following specifications:

Coordinate xRed yRed xGreen yGreen xBlue yBlue

Value 0.64 0.33 0.29 0.60 0.15 0.06

Coordinate xW hite yW hite YW hite

Value 0.3127 0.3291 1

These terms will be abbreviated to xr , yr , xg , yg , xb , yb , xw , yw and Yw . We

know already that the relation 1 links the tristimulus values to the chromaticity
coordinates.
We can transform these relations:
x z
X= Y Z= Y
y y
The first step is to use these relationships to determine the luminance Y
values. So we can calculate the tristimulus values as follows

8
  Y
Xr = Yyrr xr Xg = ygg xg Xb = Yybb xb

 Yr = Yr Yg = Yg Yb = Yb
Y
Zr = Yyrr zr Zg = ygg zg Zb = Yybb zb

For the tristimulus values of the white point:

xw zw
Xw = yw Yw = Yw Zw = yw

We now make the assumption that the sum of full intensity values of red
green and blue will be white. Using this assumption we can write this relation-
ship:

 Xw = Xr + Xg + Xb
Yw = Yr + Yg + Yb

Zw = Zr + Zg + Zb
We can then substitute the previous equations to the current one and then
rewrite this latter as a matrix relationship:
 xw   xr xg xb   
yw Y w yr yg yb Yr
 Y w  =  1.0 1.0 1.0  .  Yg 
zw zr zg zb
yw Y w yr yg yb Yb

This matrix can be re-written as follows:

   xr xg xb −1  xw 
Yr yr yg yb yw Y w
 Yg  =  1.0 1.0 1.0  .  Yw 
zr zg zb zw
Yb yr yg yb yw Y w

We now have the luminance values Yr , Yg , Yb and we can substitute these

values into the previous equations to find Xr , Xg , Xb , Zr , Zg , and Zb . The final
step is to define the relationship between tristimulus values and RGB values
as follows. The RGB matrix R should be the result of a multiplication of the
conversion matrix C by the tristimulus matrix T :
    
1 1 0 0 ? ? ? Xw Xr Xg Xb
 1 0 1 0  =  ? ? ?   Yw Yr Yg Yb 
1 0 0 1 ? ? ? Zw Zr Zg Zb
Then the conversion matrix can be calculated as follows
C = R ∗ T T ∗ (T ∗ T T )−1
If we follow this procedure using the values given in the previous example
then we arrive at the following solution:
 
3.06322 −1.39333 −0.475802
C =  −0.969243 1.87597 0.0415551 
0.0678713 −0.228834 1.06925

9
 (a) XYZ color space (b) xyY color space

Figure 4: Visualization of XYZ and xyY color spaces

4.2.2 Chromaticity coordinates of phosphors

Name xr yr xg yg xb yb White point

Short-Persistence 0.61 0.35 0.29 0.59 0.15 0.063 N/A
Long-Persistence 0.62 0.33 0.21 0.685 0.15 0.063 N/A
NTSC 0.67 0.33 0.21 0.71 0.14 0.08 Illuminant C
EBU 0.64 0.33 0.30 0.60 0.15 0.06 Illuminant D65
Dell 0.625 0.340 0.275 0.605 0.150 0.065 9300 K
SMPTE 0.630 0.340 0.310 0.595 0.155 0.070 Illuminant D65
HB LEDs 0.700 0.300 0.170 0.700 0.130 0.075 xw =.31 yw =.32

4.2.3 Standard white points

Name xw yw
Illuminant A 0.44757 0.40745
Illuminant B 0.34842 0.35161
Illuminant C 0.31006 0.31616
Illuminant D65 0.3127 0.3291
Direct Sunlight 0.3362 0.3502
Light from overcast sky 0.3134 0.3275
Illuminant E 1/3 1/3

4.3 A better model for RGB to CIE XYZ conversion

ColorSpace enables user to apply a more accurate model for RGB to CIE XYZ
conversion. This model (equ. 3) includes an offset suitable to calibrate CCD or
CMOS sensors.

10
      
X R Xof f set
 Y  = A.  G  +  Yof f set  (3)
Z B Zof f set

4.4 CIE L∗ a∗ b∗ and CIE L∗ u∗ v ∗ color spaces

There are based directly on CIE XYZ (1931) and are another attempt to
linearize the perceptibility of unit vector color differences. there are non-linear,
and the conversions are still reversible. Coloring information is referred to the
color of the white point of the system. The non-linear relationships for CIE
L∗ a∗ b∗ (see equ. 4 and fig. 5(a)) are not the same as for CIE L∗ u∗ v ∗ (see equ. 7
and fig. 5(b)), both are intended to mimic the logarithmic response of the eye.
 ³ ´ 13

 L ∗
= 116 Y
− 16 if YY0 > 0.008856

 Y

 ³0 ´
 L∗ = 903.3 Y if YY0 ≤ 0.008856
h Y0 i (4)

 a∗
= 500 f ( X
) − f ( Y
)



 h X0 Y0 i
 b∗ = 200 f ( Y ) − f ( Z )
Y0 Z0

with ½ 1
f (U ) = U3 if U > 0.008856
(5)
f (U ) = 7.787U + 16/116 if U ≤ 0.008856
and
4X 9Y
U (X, Y, Z) = et V (X, Y, Z) = (6)
X + 15Y + 3Z X + 15Y + 3Z

 ³ ´ 13

 L∗ = 116 Y
− 16 if Y
> 0.008856

 Y Y0

 ³0 ´
 L∗ = 903.3 YY0 if Y
≤ 0.008856
h i Y0
(7)

 u∗ = 13L∗ U (X, Y, Z) − U (X0 , Y0 , Z0 )



 h i
 v∗ = 13L∗ V (X, Y, Z) − V (X0 , Y0 , Z0

4.5 Color spaces used in video standards

Y U V and Y IQ are standard color spaces used for analogue television transmis-
sion. Y U V is used in European TVs (see fig. 6(a)) and Y IQ in North American
TVs (NTSC) (see fig. 6(b)). Y is linked to the component of luminance, and
U, V and I, Q are linked to the components of chrominance. Y comes from the
standard CIE XYZ.

11
(a) L∗ a∗ b∗ color space (b) L∗ u∗ v ∗ color space

Figure 5: Visualization of L∗ a∗ b∗ and L∗ u∗ v ∗ color spaces

(a) Y U V color space (b) Y IQ color space

Figure 6: Visualization of Y U V and Y IQ color spaces

12
 RGB to Y U V transformation

 Y = 0.299 × R + 0.587 × G + 0.114 × B
U = −0.147 × R − 0.289 × G + 0.436 × B

V = 0.615 × R − 0.515 × G − 0.100 × B
RGB to Y IQ transformation

 Y = 0.299 × R + 0.587 × G + 0.114 × B
I = 0.596 × R − 0.274 × G − 0.322 × B

Q = 0.212 × R − 0.523 × G + 0.311 × B
With these formulae the Y range is [0; 1], but U, V, I, and Q can be as well
negative as positive.

Y CbCr (see fig. 4.5) is a color space similar to Y U V and Y IQ. The trans-
formation formulae for this color space depend on the recommendation used.
We use the recommendation Rec 601-1 which gives the value 0.2989 for red, the
value 0.5866 for green and the value 0.1145 for blue.

Figure 7: Y CbCr color space

RGB to Y CbCr transformation


 Y = 0.2989 × R + 0.5866 × G + 0.1145 × B
Cb = −0.1688 × R − 0.3312 × G + 0.5000 × B

Cr = 0.5000 × R − 0.4184 × G − 0.0816 × B

4.6 Linear transformations of RGB

4.6.1 I1 I2 I3 color space
Ohta [3] introduced, after a colorimetric analysis of 8 images, this color space.
This color space is a linear tranformation of RGB (see fig. 4.6.1).

13
 Figure 8: I1 I2 I3 color space

RGB to I1 I2 I3 transformation
 1
 I1 = 3 (R + G + B)
1
I2 = 2 (R − B)
 1
I3 = 4 (2G − R − B)

4.6.2 LSLM color space

This color space is a linear transformation of RGB based on the opponent signals
of the cones: black–white, red–green, and yellow–blue (see fig. 4.6.2).

RGB to LSLM transformation


 L = 0.209(R − 0.5) + 0.715(G − 0.5) + 0.076(B − 0.5)
S = 0.209(R − 0.5) + 0.715(G − 0.5) − 0.924(B − 0.5)

LM = 3.148(R − 0.5) − 2.799(G − 0.5) − 0.349(B − 0.5)

4.7 HSV and HSI color spaces

The representation of the colors in the RGB and CMY color spaces are designed
for specific devices. But for a human observer, they have not accurate defini-
tions. For user interfaces a more intuitive color space is preferred. Such color
spaces can be:

• HSI ; Hue, Saturation and Intensity, which can be thought of as a RGB

cube tipped up onto one corner (see fig. 10(b) and equ. 8).

14
 Figure 9: LSLM color space

(a) HSV color space (b) HSI color space

Figure 10: Visualization of HSV and HSI

15
 RGB to HSI transformation
 β
 H = arctan(
p α)
S = α2 + β 2 (8)

I = (R + G + B)/3

with
½
α=R √
− 12 (G + B)
3
β = 2 (G − B)

• There is different way to compute the HSV (Hue, Saturation and Value)
color space [4]. We use the following algorithm 4.7.

RGB to HSV transformation

1 if (R > G) then M ax = R; M in = G; position = 0;
2 else M ax = G; M in = R; position = 1;
3 fi
4 if (M ax < B) then M ax = B; position = 2 fi
5 if (M in > B) then M in = B; fi
6 V = M ax;
7 if (M ax 6= 0) then S = M ax−MM ax
in
;
8 else S = 0;
9 fi
10 if (S 6= 0) then
11 if (position = 0)
G−B
12 then H = 1 + M ax−M in ;
13 else if (position = 1);
B−R
14 then H = 3 + M ax−M in ;
R−G
15 else H = 5 + M ax−M in ;
16 fi
17 fi

The polar representation of HSV (see fig. 11(a)) and HSI (see fig. 11(b))
color spaces leads a new visualization model of these color spaces (suitable for
color selection).

4.8 LHC and LHS color spaces

The L∗ a∗ b∗ (and L∗ u∗ v ∗ ) has the same problem as RGB, they are not very
interesting for user interface. That’s why you will prefer the LHC equa. 9 (and
LHS equa. 10), a color space based on L∗ a∗ b∗ (and LHS). LHC stand for
Luminosity , Chroma and Hue.

16
 (a) HSV color space (polar representa- (b) HSI color space (polar representa-
tion) tion)

Figure 11: Visualization of HSV and HSI color spaces (polar representation)

L∗ a∗ b∗ to LHC transformation
 ∗

 L=L √



 C = a∗2 + b∗2



 H = 0 whether a∗ = 0


 H = (arctan(b∗ /a∗ ) + k.π/2)/(2π)
whether a 6= 0 (add π/2 to H if H < 0) (9)



 and k = 0 if a∗ >= 0 and b∗ >= 0



 or k = 1 if a∗ > 0 and b∗ < 0



 or k = 2 if a∗ < 0 and b∗ < 0

or k = 3 if a∗ < 0 and b∗ > 0

L∗ u∗ v ∗ to LHS transformation


 L = L∗p



 S = 13 (u∗ − u∗w )2 + (u∗ − u∗w )2



 H = 0 whether u∗ = 0


 H = (arctan(v ∗ /u∗ ) + k.π/2)/(2π)
whether u 6= 0 (add π/2 to H if H < 0) (10)



 and k = 0 if u∗ >= 0 and v ∗ >= 0



 or k = 1 if u∗ > 0 and v ∗ < 0



 or k = 2 if u∗ < 0 and v ∗ < 0

or k = 3 if u∗ < 0 and v ∗ > 0

In order to have a correct visualization (with a good dynamic) of LHS and

LHC color spaces we used the following color transformations:

17
 L∗ a∗ b∗ to LHC transformation used in ColorSpace
 ∗

 L=L √

C = a∗2 + b∗2
(11)

 H = 0 whether a∗ = 0
 b∗
H = 180
π (π + arctan( a∗ )

L∗ u∗ v ∗ to LHS transformation used in ColorSpace



 L = L∗ p

S = 1.3 (u∗ − u∗w )2 + (v ∗ − vw
∗ )2
∗ (12)
 H = 0 whether u = 0

 v∗
H = 180
π (π + arctan( u∗ )

(a) LHC color space (b) LHS color space

Figure 12: Visualization of LHC and LHS color spaces

4.9 Spectral (λSY ) color space

λSY is a color space representation based on brightness, dominant wavelength
and saturation attributes. λSY color coordinates are defined from xyY color
coordinates.

Let us consider the xy chromaticy diagram given by Figure 13(b). Then, any
real color X that lies within the region enclosed by the spectrum locus line and
upper the lines BW and W R can be considered to be a mixture of illuminant
W and spectrum light of its dominant wavelength λd which is determined by
extending the line W X until it intersects the spectrum locus [9].

Any color Y that lies on the opposite side of the illuminant point and below
the lines BW and W R can be described both by a dominant wavelength λd and

18
 λd
λd
Green Green

xs max
x

xc λa−d
x
Spectrum locus Spectrum locus
White White

Red y Red
ys
max
Blue Blue

(a) (b)

(c) (d)

Figure 13: λSY color space transformation.

19
by its complementary wavelength λcd which is determined by extending the line
Y W until it intersects the line BR (i.e. the purple line).
The saturation S is determined in the xy chromaticity diagram, either by
the relative distance of the sample point and the corresponding spectrum point
from the illuminant point, either by the relative distance of the sample point
and the corresponding purple point from the illuminant point.

5 Decorrelated hybrid color spaces

The basic idea of hybrid color spaces is to combine either adaptively, either
interactively, different color components from different color spaces to: (a) in-
crease the effectiveness of color components to discriminate color data, and (b)
reduce rate of correlation between color components [2].
It is established that we can all the more reduce, from K to 3, the number
of color dimensions that: (a) most of color spaces are linked the ones to the
others, either by linear transformations or by non-linear transformations, and
(b) all color spaces are defined by a 3 dimensional system.

(a) (b) (c) (d) (e)

(f) (g) (h) (i) (j)

Figure 14: (a) RGB Color image, made of 6 regions (Brown, Orange, Yellow,
Pink, Green and Dark Green), projected on different color components. (b),
(c), (d) R, G, B projections. Among the three R, G, B color components, at
most 3 regions can be identified with the component G. (e), (f), (g) Y, Cb , Cr
projections. Among the three Y, Cb , Cr color components, at most 3 regions can
be identified with the component Y . (h), (i), (j) H, S, V projections. Among
the three H, S, V color components, at most 3 regions can be identified with
the component H. In combining G, Y, H color components, all regions can be
identified.

Considering that there is a high redundancy between colors components it

is, in a general way, quite difficult to define criteria of analysis to compute au-

20
tomatically the most relevant color components corresponding to a selected set
of color components. That is the reason why, in order to build a hybrid color
space, based on K 0 color components, from K selected color components, such
as K 0 << K (see Figure 14), we propose the following method: (1) select K
color components, by using a specific interface which enables the user to weight
each selected color components, and build the corresponding image of dimension
K, (2) compute the covariance matrix (of size K × K) of K color components
selected, (3) compute the eigenvectors and the eigen values of this matrix, (4)
reduce to K 0 the number of color components in computing the K 0 most signif-
icant eigen values of the covariance matrix from a principal component analysis
(PCA).
Next, the three first principal components computed (i.e. the decorrelated
hybrid colors components) are used to compute the 3D representation which
best characterizes the image studied (see Figure 15).

(a) Original Image (b) 3D representation of the hybrid colors

Figure 15: Decorrelated hybrid color space visualization

Figures 16 shows some 3D visualization of decorrelated hybrid color spaces

in different configurations in 16(b) 1 , 16(c) 1 and 16(d) 2
1 Using Components R,G and B of RGB, X and Y of XYZ, L, M and S of LM S, cos of

λSY P olar, without weight

1 Using Components L∗ and a∗ of L∗ a∗ b∗ , H and C of LHC, H and S of HSI, without

weight
2 Using Components Y of xyY , L∗ of L∗ a∗ b∗ ,L of LHC, I of HSI, without weight

21
 (a) Parrot image (b)

(c) (d)

Figure 16: Decorrelated hybrid color spaces examples on parrot image

22
6 Decorrelated hybrid color spaces applied to
image database
6.1 Decorrelated hybrid color spaces: an extension
We have introduced the hybrid construction scheme in the precedent paragraph,
based on one initial image. This strategy can be easily applied to a list of im-
ages, considering the set of images as one unique image.
We will use the following notations, and we will suppose (for simplicity of for-
mulæ only) all color spaces are normalized.
• S the set of n images and Sl the l-th image.
• K the set of selected color spaces components and Ki the i-th component.
• Kil (x, y) the corresponding value of pixel (x, y) of component Ki of image
Sl .
• Size (Sl ) the size in pixel of the image Sl .

Let introduce the Sum and Cross matrix, defined by:

X
Sumli = Kil (x, y)
xy∈Sl

X
Crosslij = Kil (x, y) ∗ Kjl (x, y)
xy∈Sl

We note, that for one image Sl , the covariance is then defined by:

l
Crosslij Sumli Sumlj
Covij = − ×
Size (Sl ) Size (Sl ) Size (Sl )

Then, to expand the formula to n images:

P l
P l
P l
1≤l≤n Crossij 1≤l≤n Sumi 1≤l≤n Sumj
Covij = P −P ×P
1≤l≤n Size (Sl ) 1≤l≤n Size (Sl ) 1≤l≤n Size (Sl )

At this point, the computation of the hybrid color space follows the previ-
ously cited steps:
4. principal component analysis;
5. selection of the 3 most significant axis.
We have developed via ICobra and ColorSpace applications a web interface
system3 [14] to manage these hybrid color spaces. The process is divided in two
steps:
3 Available at: http://www.icobra.info/hybrid.php

23
 • Off-line computation. This part intends to compute the main portion of
calculus required by the hybrid color space computation.

• Online interface. This part intends, via the interface, to select color spaces
components and images, to complete the calculus, and, then, to show the
selected image in the computed hybrid space.

Before describes rapidly these two parts, we can note: all color spaces are nor-
malized during the computation (there is a scale rapport from 1 to 200 between
some spaces); the transfer values (primaries and white settings) used are, at this
moment, still approximation.

6.2 Off-line computation

Figure 17: Off-line scheme

The figure 17 illustrates the off-line computation. For each image, we will
compute and store:
• for each couple of color spaces components, the Cross value corresponding.
It results a m × m matrix, where t is the number of possible color spaces
components, presently 72 (3 × 24) components.

• for each color spaces components, the Sum value corresponding. It results
a vector of size m.
• the size of image.

6.3 Online interface

As shown in 17 the application may be split into several sections:

• World Wide Web Interface, as figure 19 illustrates. It permits:

– to browse the different image databases;
– to select by clicks a list of images. An empty list means that the
decorrelated hybrid color space is computed on self-image, as usual;

24
Figure 18: Overall online scheme

Figure 19: WWW interface

25
 – to select a list of color spaces components;
– to choose the mode of visualization: 2D image, 3D, or 3D histogram;
– to launch ColorSpace with the selected parameters;
• A computational part: the final covariance matrix is computed, using
pre-calculated data;
• A CSI file generation: a csi file (Color Space Interface) is generated, in-
cluding all settings and information in order to compute the PCA and
displaying the selected visualization;

• ColorSpace launching: ColorSpace is launched by the browser (applica-

tion/csi mime-type bind) with the csi file as parameter. The software
computes the PCA, and renders the selected visualization.

Figure 20 illustrates this tool within some screenshots using different images
and color spaces components.

7 References and lectures

[1] P. Colantoni and A. Trémeau, “3d visualization of color data to analyze
color images,” in The PICS Conference, (Rochester, USA), May 2003.
[2] N. Vandenbrouke, L. Macaire, and J. G. Postaire, “Color pixels classifi-
cation in a hybrid color space,” in Proceedings of IEEE, pp. 176–180,
1998.
[3] Y. I. Ohta, T. Kanade, and T. Sakai, “Color information for region seg-
mentation,” Computer Graphics and Image Processing, vol. 13, pp. 222–
241, 1980.
[4] D. F. Rogers, Procedural Elements for Computer Graphics. Mc Graw Hill,
1985.
[5] G. A. Agoston, Color Theory and its Application in Art and Design. Op-
tical Sciences, Springer-Verlag, 1987.
[6] CIE, “Parametric effects in colour-difference evaluation,” Tech. Rep. 101,
Bureau Central de la CIE, 1993.
[7] CIE, “Technical collection 1993,” tech. rep., Bureau Central de la CIE,
1993.
[8] CIE, “Industrial colour-difference evaluation,” Tech. Rep. 116, Bureau
Central de la CIE, 1995.
[9] D. L. MacAdam, Color Measurement, theme and variation. Optical Sci-
ences, Springer-Verlag, second revised ed., 1985.
[10] G. Wyszecki and W. S. Stiles, Color Science: Concepts and Methods,
Quantitative Data and Formulae. John Wiley & sons, second ed., 1982.

26
(a) Original image (b) Selected com- (c) Decorrelated hybrid color
ponents and images space representation
list

(d) (e) (f)

(g) (h) (i)

Figure 20: Decorrelated hybrid color spaces: some examples

27
[11] H. R. Kang, Color Technology for Electronic Imaging Devices. SPIE
Optical Engineering Press, 1997.
[12] K. N. Plataniotis and A. N. Venetsanopoulos, Color Image Processing
and Application. Springer, 2000.
[13] R. Hall, Illumination and Color in Computer Generated Imagery.
Springer-Verlag, 1988.
[14] J. D. Rugna, P. Colantoni, and N. Boukala, “Hybrid color spaces applied
to image database,” vol. 5304, pp. 254–264, Electronic Imaging, SPIE,
2004.

28