A Concise Introduction to
Color Management and ICC profiles
[Note that there are many other, perhaps more comprehensive and
expansive "introduction to Color Management" resources on the web.
Google is your friend...]
Color management is a means of dealing with the fact that color
capture and output devices such as Cameras, Scanners, Displays and
Printers etc., all have different color capabilities and different
native ways of communicating color. In the modern world each device
is typically just part of a chain of devices and applications that
deal with color, so it is essential that there be some means for
each of these devices to communicate with each other about what they
mean by color.
Successful color management allows colors to be captured,
interchanged and reproduced by different devices in a consistent
manner, and in such a way as to minimize the impact of any technical
limitation each device has in relation to color. It must also deal
with the interaction of human vision and devices, allowing for such
fundamental vision characteristics as white point adaptation and
other phenomena. It should also allow the human end purposes to
influence the choice between tradeoffs in dealing with
practical device limitations.
The key means of implementing color management is to have a way of
relating what we see, to the numbers that each device uses to
represent color.
The human eye is known to have 3 type of receptors responsible for
color vision, the long, medium and short wavelength receptors.
Because there are 3 receptors, human color perception is a 3
dimensional phenomena, and therefore at least 3 channels are
necessary when communicating color information. Any device capable
of sensing or reproducing color must therefore have at least 3
channels, and any numerical representation of a full range of colors
must have at least 3 components and hence may be interpreted as a
point in a 3 dimensional space. Such a representation is referred to
as a Color Space.
Typically color capture and output devices expose their native color
spaces in their hardware interfaces. The native color space is
usually related to the particular technology they employ to capture
or reproduce color. Devices that emit light often choose Red Green and Blue (RGB) wavelengths, as these are particularly
efficient at independently stimulating the human eye's receptors,
and for capture devices R,G & B are roughly similar to the type
of spectral sensitivity of our eyes receptors. Devices that work by
taking a white background or illumination and filtering out (or subtracting) colors tend to use
Cyan, Magenta, and Yellow (CMY) filters or colorants to
manipulate the color, often augmented by a Black channel (CMYK). This is because a Cyan
filters out Red wavelengths, Magenta filters out Green wavelengths,
and Yellow filters out Blue wavelengths, allowing these colorants to
independently control how much RGB is emitted. Because it's
impossible to make filters that perfectly block C, M or Y
wavelengths without overlapping each other, C+M+Y filters together
tend to let some light through, making for an imperfect black.
Augmenting with an additional Black filter allows improving Black,
but the extra channel greatly complicates the choice of values to
create any particular color.
Many color devices have mechanisms for changing the way they respond
to or reproduce color, and such features are called Adjustments, or Calibration. Such features can
be very useful in adapting the device for use in a particular
situation, or for matching different instances of the device, or for
keeping its behavior constant in the face of component or
environmental changes. Sometimes there may be internal
transformations going on in the device so that it presents a more or
less expected type of color space in its hardware interface. [ Some
sophisticated devices have built in means of emulating the behavior
of other devices, but we won't go into such details here, as this is
really just a specialized implementation of color management. ]
To be able to communicate the way we see color, a common "language"
is needed, and the scientific basis for such a language was laid
down by the International Commission on Illumination (CIE) in 1931
with the establishment of the CIE 1931 XYZ color space. This provides a means of predicting
what light spectra will match in color for a Standard Observer, who
represents the typical response of the Human eye under given viewing
conditions. Such a color space is said to be Device Independent since it is
not related to a particular technological capture or reproduction
device. There are also closely related color-spaces which are direct
transformations of the XYZ space, such as the L* a* b* space which is a more
perceptually uniform device independent colorspace.
As mentioned above, the key to managing color is to be able to
relate different color spaces so that they can be compares and
transformed between. The most practical approach to doing this is to
relate all color spaces back to one common colorspace, and the CIE
XYZ colorspace is the logical choice for this. A description of the
relationship between a devices native color space and an XYZ based
colorspace is commonly referred to as a Color Profile. As a practical issue when dealing
with computers, it's important to have a common and widely
understood means to communicate such profiles, and the ICC profile format standardized
by the International Color Consortium is today's most widely
supported color profile format.
The ICC profile format refers to it's common color space as the Profile Connection Space (PCS), which is closely based on
the CIE XYZ space. ICC profile are based on a Tagged format, so they
are very flexible, and may contain a variety of ways to represent
profile information, and may also contain a lot of other optional
information.
There are several fundamental types of ICC profiles. Device and Named profiles represent color anchor points. Device Link and Abstract profiles represent journeys between anchor
points.
Device
These primarily provide a translation between
device space and PCS. They also typically provide a translation in
the reverse direction, from PCS to device space. They provide an
"color anchor" with which we are able to navigate our way around
device color. The mechanisms they use to do this are discussed in
more detail below.
Device Link
A Device Link profile provides a transformation
from one Device space to another. It is typically the result of
linking two device profiles, ie. Device A -> PCS -> Device B,
resulting in a direct Device A -> Device B transformation.
Abstract
An abstract profile contains a transformation
define in PCS space, and typically represents some sort of color
adjustment in a device independent manner.
Named
A Named profile is analogous to a device Profile,
but contains a list of named colors, and the equivalent PCS and
possibly Device values.
Most of the time when people talk about "ICC profiles" they mean Device Profiles. Profiles rely
on a set of mathematical models to define the translation from one
colorspace to another. The models represent a general framework,
while a specific profile will define the scope of the model as well
as it's specific parameters, resulting an a concrete translation.
Profiles are typically used by CMMs
(Color Management Modules), which are a piece of software (and
possibly hardware) that knows how to read and interpret an ICC
profile, and perform the translation it contains.
Often the function of a CMM will be to take two device profiles, one
representing the starting point and the other representing the
destination, and create a transformation between the two and
applying it to image pixel values.
Two basic models can be used in ICC profiles, a Matrix/shaper model and a cLUT (Color Lookup Table) model.
Models often contain several optional processing elements that are
applied one after the other in order to provide an overall
transformation.
The Matrix/Shaper model consists of a set of per channel lookup
curves followed by a 3x3 matrix. The curves may be defined as a
single power value, or as a one dimensional lookup table which
encodes a discretely represented curve (Lut). The matrix step can
only transform between 3 dimensional to 3 dimensional color spaces.
The cLUT model consists of an optional 3x3 matrix, a set of per
channel one dimensional LUTs, an N dimensional lookup table (cLUT)
and a set of per channel one dimensional LUTs. It can transform from
any dimension input to any dimension output.
All Lookup Tables are interpolated, so while they are defined by a
specific set of point values, in-between values are filled in using
(typically linear) interpolation.
For a one dimensional Lookup table, the number of points needed to
define it is equal to its resolution.
For an n-dimensional cLUT, the number of points needed to define it
is equal to it's resolution taken to the power of the number of
input channels. Because of this, the number of entries climbs rapidly with resolution, and
typical limited resolution tables are used to constrain profile file
size and processing time. cLUT's permit detailed, independent
control over the the transformation throughout the colorspace.
Limitations of CIE XYZ
Although CIE XYZ colorspace forms an excellent basis for connecting
what we can measure with what we see in regard to color, it has its
limitations. The primary limitation is that the visual match between
two colors with the same XYZ values assumes identical viewing
conditions. Our eyes are marvelously adaptable, automatically
adjusting to different viewing conditions so that we are able to
extract the maximum amount of useful visual information. There are
many practical situations in which the viewing conditions are not
identical - e.g. when evaluating an image against our memory of an
image seen in a different location, or in viewing images side by
side under mixed viewing conditions. One of the primary things that
can change is our adaptation to the white point of what we are
looking at. This can be accounted for in XYZ space by applying a
chromatic adaptation, which mimics the adaptation of the eye. The
ICC profile format PCS space by default adapts the XYZ values to a
common white point (D50), to facilitate ease of matching colors
amongst devices with different white points. Other viewing condition
effects (ie. image luminance level, viewing surround luminance and
flare) can be modeled using (for example) using CIECAM02 to modify
XYZ values.
Another limitation relates to spectral assumptions. CIE XYZ uses a
standard observer to convert spectral light values into XYZ values,
but in practice every observer may have slightly different spectral
sensitivities due to biological differences, including aging.
(People with color deficient vision may have radically different
spectral sensitivities.) Our eyes also have a fourth receptor
responsible for low light level vision, and in the eye's periphery
or at very low light levels it too comes to play a role in the color
we perceive, and is the source of a difference in the eye's spectral
sensitivity under these conditions.
Another spectral effect is in the practice of separating the color
of reflective prints from the light source used to view them, by
characterizing a prints color by it's reflectance. This is very
convenient, since a print will probably be taken into many different
lighting situations, but if the color is reduced to XYZ reflectance
the effect of the detailed interaction between the spectra of the
light source and print will lead to inaccuracies.