Color Imaging for Multimedia

GAURAV SHARMA, MEMBER, IEEE, MICHAEL J. VRHEL, MEMBER, IEEE, AND

H. JOEL TRUSSELL, FELLOW, IEEE

To a significant degree, multimedia applications derive their film, from digital still cameras, or even from camcorders.
effectiveness from the use of color graphics, images, and video. The images that are produced for the customer may be
However, the requirements for accurate color reproduction and viewed in a printed catalog or on the computer monitor
for the preservation of this information across display and print
devices that have very different characteristics and may be or downloaded and printed on the user’s local printer. To
geographically apart are often not clearly understood. This paper avoid unnecessary grief to the customer, the final image in
describes the basics of color science, color input and output all cases should be a faithful representation of the original
devices, color management, and calibration that help in defining product. Other examples could be given for remote medical
and meeting these requirements.
diagnosis, educational presentations, industrial design of
Keywords— Color, color imaging, colorimetry, display, multi- products, and, of course, entertainment.
media, printing, scanning, survey, tutorial.
Since color is such a large part of multimedia presen-
tation, it is important that both the designers and users
I. INTRODUCTION of multimedia systems understand how to describe and
In the last few years, computer-based multimedia systems present color accurately. The presentation of accurate color
have grown from their humble beginnings into systems requires a joint effort on the part of both the designer
that truly allow the integration of information from many or manufacturer and the consumer. The consumer should
different types of sources (media), including text, numeric be aware of the need for the calibration of both input
data, audio, graphics, still images, and video. In keeping and output devices. The designers of imaging devices for
with this trend, it is anticipated that multimedia systems will multimedia systems should make their devices easy to
continue to grow rapidly in the areas of information pre- calibrate and easy to maintain calibration. At the same time,
sentation, education, and entertainment, with applications designers of multimedia applications should make products
including teaching, electronic publishing, telemedicine, and that effectively use the calibration of the various devices.
retailing. In the initial euphoria surrounding these develop- There is a range of needs of the users that must be
ments, however, several issues have not received adequate spanned by the producers of multimedia products. The
attention. In particular, for the presentation of image data, level of color accuracy required for an application must
one question is still left over from the early days. How be identified in order to specify what hardware can be used
do you ensure that two images presented on two different and what information is to be given or supplied by the
devices can be accurately compared? user. In the case of accurate color presentation, it is often
The problem is much more complex now than in early important to know the intent of the user of the image. In
systems. Images that are to be compared in present multi- certain applications, such as mail-order catalogs, absolute
media applications may come from entirely different input colorimetric accuracy is desired so that chosen items may
sources and may be reproduced on entirely different output match or coordinate with existing articles. In other applica-
devices. Consider the publisher of a merchandise catalog tions, such as presentation of art on a computer, it is often
who is trying to sell products over the World Wide Web. necessary to sacrifice absolute colorimetric accuracy in
The images of the products may come from photographic favor of an image that appears visually similar to that in the
museum, under the much different viewing conditions of
Manuscript received August 15, 1997; revised December 1, 1997. The
Guest Editor coordinating the review of this paper and approving it for the monitor. To achieve either of these effects, the original
publication was A. M. Tekalp. image must be recorded correctly so that the data can
G. Sharma is with the Digital Imaging Technology Center, Xerox produce a good display. Information about the rendering
Corporation, Webster, NY 14580 USA (e-mail: sharma@wrc.xerox.com).
M. J. Vrhel is with Color Savvy Systems, Ltd., Springboro OH 45066 intent (for instance, colorimetric versus perceptual) for
USA (e-mail: mvrhel@colorsavvy.com). displaying the image must also be tagged to the image file
H. J. Trussell is with the Electrical and Computer Engineering De- for interpretation and use in a display device.
partment, North Carolina State University, Raleigh, NC 27695-7911 USA
(e-mail: hjt@eos.ncsu.edu). As can be seen from this brief introduction, there are
Publisher Item Identifier S 0018-9219(98)03520-8. several steps to producing good color images for a mul-

0018–9219/98$10.00  1998 IEEE

1088 PROCEEDINGS OF THE IEEE, VOL. 86, NO. 6, JUNE 1998

timedia application. This paper presents the fundamentals to produce a color match for a given stimulus (equiva-
that both the user and the designer should understand in lently, identical cone responses under the same viewing
order to get the most from the system. Section II describes conditions) by using only three light sources, known as
the basics of the color visual system, color terminology, primaries (see [1] and [2] for references and a description
and color relationships. The third section has three major using notation and terminology identical to that used here).
subtopics that are closely interrelated. Image character- The sensitivity functions, however, have been subsequently
istics and reproduction methods are discussed together, verified through physiological examinations. From (1), it
since many images used in multimedia presentations are can be seen that the cone responses depend only on the
themselves reproductions of other images. Image-recording orthogonal projection of the spectrum onto the column
methods are dependent on the characteristics of the image. space of , which is called the human visual subspace
Section IV presents the basics of color management and (HVSS) [3]–[6].
calibration, which are necessary to communicate color While the cone sensitivities themselves cannot be
information among the various devices. Section V details readily determined, alternate bases that span the HVSS
an illustrative example that uses the principles described, can be determined through color-matching experiments that
and Section VI gives a summary of the paper. At the outset determine the strengths of primaries required to match
itself, the authors would like to clarify that to keep the paper monochromatic color spectra over the entire visible range
brief, they have attempted to select a few representative [1], [2], [7]. These bases are known as color-matching
citations instead of attempting exhaustive coverage and duly functions (CMF’s) and depend on the primaries used in the
assigning credit. color-matching experiments. The 3 matrix formed by a
set of CMF’s is within a (nonsingular) linear transformation
II. COLOR VISION AND COLORIMETRY of . Out of the infinite set of CMF’s that can be used
The perception of color involves complex interactions for color specification, the International Commission on
between several physical, neural, and cognitive phenomena, Illumination (CIE) has defined two sets of CMF’s for use
which must be understood in order to comprehend color in standard colorimetry: 1) the CIE red-green-blue (RGB)
vision completely. While research is still ongoing in the CMF’s and 2) the CIE XYZ CMF’s [8]. The CIE RGB
integration of all these aspects of color, significant success CMF’s correspond to monochromatic primaries at 700 nm
has been achieved in understanding the physical and (to a (red), 546.1 nm (green), and 435.8 nm (blue). The CIE XYZ
lesser extent) neural phenomena involved in color sensa- CMF’s do not correspond to any physical set of primaries
tion. The goal of this section is to present an overview of but were selected by the CIE so as to have several properties
this current understanding, with particular emphasis on the that were considered desirable when the standards were
aspects that are of interest in color-imaging applications. defined [9], [10, p. 531]. In this paper, the CIE XYZ CMF’s
will be used. Using these CMF’s, the color of the irradiant
spectrum can be specified by its CIE XYZ tristimulus
A. Trichromacy and the Visual Subspace
value, which is given by , where
The sensation of color is produced when light in the vis- is a matrix whose columns are the CIE XYZ CMF’s.
ible region of the electromagnetic spectrum (approximately Since the results of color-matching experiments vary in
corresponding to the wavelength region from 400 to 700 accordance with the angular size of the viewing field, the
nm) excites retinal receptors in the eye. The human eye CIE has defined two sets of XYZ CMF’s: 2 CMF’s for
has three color-sensing receptors, called S, M, and L cones, use with uniform visual fields of angular subtense from 1
which are abbreviated forms of short-, medium-, and long- to 4 and 10 CMF’s for use with uniform visual fields of
wavelength sensitive cones, respectively. Under a fixed set larger angular subtense. Only the 2 CMF’s will be used
of viewing conditions, the response of these cones can throughout this paper because imaging applications involve
be accurately modeled by a linear system defined by the complex visual fields where the color-homogeneous areas
spectral sensitivities of the cones. For the purposes of this have small angular subtense. The (2 ) CIE XYZ CMF’s
paper, it will be convenient to assume that (color) spectra are shown in Fig. 1. The perceived lightness/brightness of
are represented by -vectors consisting of their samples at a color is often assumed to be a monotonic function of the
a discrete set of wavelength values that span tristimulus value, which is known as the luminance; and
the visible wavelength range. Thus, the cone responses pro- the corresponding CMF, , is often referred to as the
duced by light (stimulus) with spectral power distribution photopic luminosity function.
are given by As mentioned earlier, the 3 matrices and
(1) are within a (nonsingular) linear transformation of each
other. Therefore, the HVSS can alternately be defined as
where is the vector of cone responses the column space of . Also note that for colorimetry in
and is the 3 matrix of cone spectral specific applications, alternate standard CMF’s have also
sensitivities. been defined, an example being the National Television
Equation (1) was first inferred from experimental obser- Systems Committee (NTSC) [11]/Society for Motion and
vations of the linearity of color matching and the phe- Television Pictures (SMPTE) [12], [13] CMF’s developed
nomenon of trichromacy, which states that it is possible for use in TV. The same color can therefore be specified

SHARMA et al.: COLOR IMAGING FOR MULTIMEDIA 1089

Fig. 1. CIE 2 XYZ color-matching functions.

in numerous color spaces by using tristimulus values com- ment geometries used in common color instrumentation. An
puted with these different CMF’s. The term “color space” is additional limitation of (2) is its inability to account for
also applied to other coordinate systems used for specifying fluorescence.
color, which are typically transformations of the tristimulus Note that the linear model of (1) and (2) (or any equiva-
values. lent model) is limited in several respects. While this model
usually explains the color matching of two spatially uniform
B. Colorimetry for Reflective Objects patches placed over a constant background, it does not take
into account several adaptation effects in the eye that are
The discussion in the last section was based on the
inherently nonlinear [14, ch. 9]. As will be discussed in
assumption that is the spectral irradiance of the light
Section II-D1, the simple linear model does not completely
incident on the eye. When a reflective object with spectral
explain even the matching of uniform patches of color under
reflectance is viewed under an illuminant with spectrum
different viewing conditions, let alone the color-matching
given by the vector , the resulting spectral radiance at
characteristics of complex scenes. An additional limitation
the eye is the product of the illuminant spectrum and the
of the linear model, which will also be addressed in
reflectance at each wavelength. Therefore, the CIE XYZ
Section II-D1, is that Euclidean distances in the tristimulus
tristimulus value defining the color is given by
spaces defined by the model are in poor agreement with
(2) perceived color differences.

where is the matrix of CIE XYZ CMF’s (as before),

is the diagonal illuminant matrix with entries from C. Illuminants and Color Temperature
along the diagonal, and . Color measurement The tristimulus color specification of a reflective object
for transmissive objects can be similarly defined in terms is dependent on the viewing illuminant. The CIE has
of their spectral transmittance. In analogy with the HVSS, defined several standard illuminants for use in colorimetry
the column space of is defined as the human visual of reflective objects. The relative irradiance spectra of a
illuminant subspace [6]. Note that (2) is a simplified model number of these standard illuminants is shown in Fig. 2.
of the interaction between the illuminant and the surface To represent different phases of daylight, a continuum of
that does not account for several geometry/surface effects. daylight illuminants has been defined by the CIE [8], which
In particular, the reflected light from a surface typically has are uniquely specified in terms of their correlated color
a specular component (similar to reflection by a mirror) temperature (CCT). Since the temperature of a black-body
in addition to the diffuse component represented in (2). radiator describes its complete spectral power distribution
Throughout this paper, it will be assumed that the spec- and thereby its color, it is commonly referred to as the color
ular component is excluded from the color-measurement temperature of the black body. For an arbitrary illuminant,
process. This assumption is normally valid for the measure- the CCT is defined as the color temperature of the black-

1090 PROCEEDINGS OF THE IEEE, VOL. 86, NO. 6, JUNE 1998

Fig. 2. CIE standard illuminants.

body radiator that is visually closest to the illuminant (in outputs of the cones go through a process of nonlinear
color) [7]. The D65 and D50 illuminant spectra shown transformation and mixing in the neural pathways lead-
in Fig. 2 are two daylight illuminants commonly used ing from the eye to the brain. It is believed that this
in colorimetry, which have CCT’s of 6500 and 5000 K, process transforms the cone responses into an achromatic
respectively. The CIE illuminant “A” represents a black- channel carrying only information of lightness/brightness
body radiator at a temperature of 2856 K and closely and two opponent chromatic channels: one carrying a red-
approximates the spectra of incandescent lamps. Sources green chromatic-difference signal and the other carrying
with lower CCT tend to be more red, while those with a yellow-blue chromatic-difference signal. Research has
higher temperatures are bluer. Illuminants with similar CCT indicated that the achromatic channel carries information
are assumed to be similar with regard to their color render- at a high spatial resolution and the chromatic channels
ing of illuminated objects. This, however, is true only for use significantly reduced resolution. This framework for
illuminants whose spectra closely resemble that of a black- color vision is broadly accepted and is successful in qual-
body radiator, and other spectra that have identical CCT itatively explaining a wide variety of observations from
can have very different distributions and color-rendering psychophysical experiments. However, the exact nature of
properties [15]. An example of the problem with the use the transformations involved is still the subject of active
of CCT for specifying the color-rendering properties of research. The interested reader is referred to [10] and [14]
an illuminant is shown in Fig. 3, where two synthesized for recent descriptions and historical references. Several
illuminants are shown along with a reflectance spectrum semiempirical models that conform to the above framework
measured from a cyan print sample. Though the illuminants have been developed for use in color science and imaging.
have the same luminance and an identical CCT of 5000
1) Uniform Color Spaces and Color Appearance: One lim-
K, the color difference for the reflectance sample under the
itation of the CIE XYZ space (and other tristimulus spaces)
two illuminants is rather large, corresponding to 44.4
is that it is perceptually nonuniform, i.e., equal differences
units. (For the definition of see Section II-D1.)
in CIE XYZ tristimulus values (as measured by the Eu-
clidean distance) do not correspond to equal perceived
D. Psychophysical Aspects of Color differences in colors. Thus, the CIE XYZ space is primarily
The simple model of (1) describes only a small part of the useful for determining color matches and not the closeness
complex color-perception process. In particular, the cone of colors that are not in exact match. Since a measure of
responses cannot be directly related to the common color the perceived difference between colors is very useful in
attributes of hue, saturation, and lightness/brightness.1 The many applications, considerable research has focused on
transforming tristimulus values into alternate color spaces
1 The readers are referred to [7, pp. 487] and [16]–[18] for definitions
that are perceptually uniform. The CIE has defined two
of hue, saturation, lightness, brightness, and other color-appearance ter-
minology. Common notions of these terms, however, will suffice for the uniform color spaces (UCS’s) for use in colorimetry, of
purposes of this paper. which the more commonly used is the CIELAB color space

SHARMA et al.: COLOR IMAGING FOR MULTIMEDIA 1091

Fig. 3. CCT counterexample with two illuminants, with CCT = 5000 K and a spectral reflectance.

[8]. The CIELAB color space is based on a color vision data to be manipulated along these perceptual axes. In
model [19] and transforms XYZ tristimulus values into an addition, this transformation has been exploited in defin-
achromatic lightness value and two chromatic values ing alternate (non-Euclidean) color-difference formulas in
and using the transformations CIELAB space that attempt to remedy the deviations from
uniformity in CIELAB. Most of these formulas decompose
the total CIELAB color difference into lightness, hue, and
chroma differences and introduce weighting factors for
these differences that are determined by the lightness, hue,
and chroma. The weighting factors are determined so as
to obtain better agreement of the modified color-difference
(3) formula with experimental data. The most prominent and
successful among these difference formulas is the CMC(l:c)
where are the tristimuli of the “white stimulus” distance function [21]. Recently [22], the CIE issued a new
and recommendation for the computation of color differences in
CIELAB space that incorporates several of the robust and
(4) attractive features of the CMC(l:c) distance function. Color
differences computed in accordance with this recommenda-
The chromatic channel is a red-green opponent channel tion are referred to as color differences.
in that positive values for indicate redder colors and The human eye can sense color over a very large dynamic
negative values indicate greener colors, with the magnitude range of light intensities. One of the remarkable aspects of
indicating the strength of the redness/greenness. Similarly, color vision is that the color of reflective objects is largely
is a yellow-blue opponent color channel. Note that the unchanged over this vast range of light intensities and over
lightness value depends on only the luminance , and several different viewing illuminants. The term color con-
this is a result of a deliberate decision made in choosing stancy is used to describe this phenomenon of object colors’
the CIE XYZ CMF’s. remaining invariant under changes in viewing illumination.
The Euclidean distance between two color stimuli in The adaptation of the color-sensing mechanism in the eye
CIELAB space is denoted by (delta - ), and a in response to these changes is called chromatic adaptation.
value of around 2.3 corresponds to a just noticeable Obviously, this chromatic adaptation is not included in the
difference [20]. The radial distance ( ) and simple linear model of (1), which works only for a fixed set
angular position ( ), in the plane, serve of viewing conditions. Most models of chromatic adaptation
as correlates of chroma and hue, respectively. Thus, a assume that it is achieved through independent adaptive
transformation of CIELAB into cylindrical coordinates pro- gain control on each of the three cones. This approach
vides approximate lightness, chroma, and hue attributes. was first proposed by Von Kries [23], who hypothesized
This is extremely useful in applications that require color that the cone responses are scaled linearly so that the

1092 PROCEEDINGS OF THE IEEE, VOL. 86, NO. 6, JUNE 1998

white stimulus produces the same response under different filters for the , , and channels. Similar approaches
viewing conditions. Mathematically, this is achieved by based on slightly different transforms for luminance and
applying a diagonal scaling matrix to the cone responses chrominance channels have also been proposed in [26],
and is referred to as a Von Kries transformation [7]. Note [28], and [29]. While other models that are considerably
that the white normalization of the XYZ tristimuli in the more sophisticated have been developed for luminance
transformation to CIELAB in (3) is motivated by the information [30], [31], their extensions to color are yet to
same reason, though the transformation is not really a Von be developed.
Kries transform because the XYZ tristimuli are not cone
responses. As will be discussed in Section IV-E, a Von III. COLOR-IMAGE REPRODUCTION
Kries transform is often used in imaging applications to AND RECORDING DEVICES
account for the eye’s adaptation to changes in viewing In the physical world, color images exist as spatially
conditions. varying spectral radiance or reflectance distributions. To
Several researchers are investigating models for color be processed digitally, these images need to be sampled
vision that incorporate adaptation and other psychophysical both spatially and spectrally. The issues involved in spatial
effects and provide descriptors of color appearance that sampling and reconstruction of images have been discussed
are independent of the viewing conditions (as opposed at length in signal-processing literature and will not be
to simple CIE XYZ colorimetry). These color-appearance repeated here. The aspects of spectral sampling and color
models hold great promise in imaging applications that recording for images will be discussed here. Common
require images to be rendered on different devices (such color-reproduction devices, spectral characteristics of natu-
as a CRT monitor and print) with different viewing con- ral images and images from these reproduction devices, and
ditions. To obtain “identical appearing images,” the input methods for recording these images are described briefly.
viewing conditions and the colorimetric specification of the
image under these viewing conditions are used in a color- A. Image-Reproduction Systems
appearance model to obtain a color-appearance description There are several systems that are commonly used for
of the image. The color-appearance description, along with the reproduction of color images. These systems can be
the output viewing conditions, can then be used in the broadly classified into three types: additive, subtractive,
inverse of the appearance model to obtain the colorimetry and hybrid. Additive color systems produce color through
(under the output viewing conditions) for an image that the combination of differently colored lights, known as
appears identical to the input image. The reader is referred primaries. The qualifier additive is used to signify the fact
to [24] for a survey and comparison of several of the current that the final spectrum is the sum (or average) of the spectra
color-appearance models. of the individual lights. Examples of additive color systems
2) Spatial Aspects: The CIELAB color space attempts to include color cathode-ray tube (CRT) displays and projec-
account for the nonlinearities in the color-sensing process tion video systems. Color in subtractive systems is produced
for the comparison of relatively large uniform color regions through a process of removing (subtracting) unwanted
and does not account for any spatial interactions in the spectral components from “white” light. Typically, such
eye. Since the eye acts as a spatial low-pass filter, which systems produce color on transparent or reflective media,
can average over high spatial-frequency differences in which are illuminated by white light for viewing. Dye
images, a point-by-point computation of color differences in sublimation printers, color photographic prints, and color
CIELAB color space is not appropriate for the comparison slides are representatives of the subtractive process. Hybrid
of images. Though the eye exhibits low-pass characteristics systems use a combination of additive and subtractive
for both luminance2 and chrominance spatial information, processes to produce color. The main use of a hybrid system
the bandwidth for the chrominance channels is much lower is in color halftone printing, which is commonly used for
than that of the luminance channels. This fact has been ex- lithographic printing and in most desktop color printers.
ploited in numerous applications, including color television Any practical output system is capable of producing only
(TV), where the transmission bandwidth allocated to the a limited range of colors. The range of producible colors
chrominance channels is significantly lower than that for on a device is referred to as its gamut. Typical additive
the luminance channel. systems use red, green, and blue primaries because they
Several researchers have modeled the spatial properties of offer the largest possible gamut, and for the same reason,
the eye as a linear filter and studied the frequency response subtractive systems use cyan (C), magenta (M), and yellow
of the eye for luminance and chrominance spatial patterns (Y) colorants. In addition, subtractive systems often use a
[25], [26]. However, a complete model for human vision fourth black (K) colorant in addition to CMY to achieve
that is perceptually uniform and incorporates the effects greater contrast, better reproduction of achromatic (gray)
of the spatial interactions is yet to be developed. Recent colors, and lower cost through the replacement of the more
attempts in this direction include the S-CIELAB model [27] expensive CMY colorants with the K.
that builds on CIELAB by incorporating spatial low-pass The gamut of a device can be represented by a three-
2 Strictly speaking, the luminance response of the eye is believed to be
dimensional (3-D) solid in any color space such as CIE
bandpass, but for all practical purposes, the low-frequency attenuation can XYZ/CIELAB. Since two-dimensional (2-D) representa-
be ignored. tions are more convenient for display, it is common to

SHARMA et al.: COLOR IMAGING FOR MULTIMEDIA 1093

Fig. 4. CIE xy chromaticity diagram showing locations of colors in chromaticity space.

utilize CIE chromaticity diagrams [7] for this purpose, ability to manipulate these views interactively is extremely
where the and chromaticity coordinates are useful in understanding the capabilities and limitations of
and , respectively. A graphical the different color-output devices. Three different views
representation of the CIE diagram showing the relative comparing the gamuts of a CRT monitor and the gamut of a
locations of the different colors is displayed in Fig. 4. [32] dye-sublimation contone color printer are shown in Fig. 5,
On the CIE chromaticity diagram, the gamut of an addi- and identical views for an inkjet halftone color printer are
tive system appears as a convex polygon with the primaries shown in Fig. 6. In both cases, the wire frames represent
representing the vertices. For the usual case of three red, the gamut of a CRT monitor and the solids represent the
green, and blue primaries, the gamut is a triangle on the CIE gamuts of the printer. These views demonstrate that the
chromaticity diagram. Since most subtractive and hybrid gamuts of these three output devices are fairly different with
systems are nonlinear, their gamuts have irregular shape and several colors that can be produced on one device and not
are not characterized by such elegant geometric constructs. on another. Overall, the gamut of the monitor is the largest,
The two-dimensional representation of the gamut on the followed by the gamut of the continuous-tone printer and
CIE diagram presents only an incomplete (and difficult then by the inkjet halftone printer, which has a rather small
to interpret) picture because it does not represent the full 3- gamut in comparison to the other devices. As will be seen
D data. With the increase in computing speeds and advances in Section IV-D, this mismatch in the gamut between the
in computer graphics, visualization techniques are now devices poses significant challenges in cross-media color
used to render 3-D views of the gamuts [33], [34]. The reproduction.

1094 PROCEEDINGS OF THE IEEE, VOL. 86, NO. 6, JUNE 1998

 (a)

(b)

(c)
Fig. 5. Comparison of a CRT monitor gamut (shown as a wire frame) and a continuous-tone
dye-sublimation printer gamut (shown as a solid) in CIELAB space. (a) Top view along the L3
axis. (b) and (c) Perspective projections.

1) CRT Displays: The most widely used display device at each phosphor site is controlled by electrons striking it
for television and computer monitors is the color CRT. The from an electron gun. Images are displayed by spatially
face of a CRT is coated with a mosaic of red, green, and modulating the RGB electron guns in accordance with the
blue light emitting phosphors. The intensity of light emitted image signal. Since the mosaic is extremely fine, the eye

SHARMA et al.: COLOR IMAGING FOR MULTIMEDIA 1095

 (a)

(b)

(c)
Fig. 6. Comparison of a CRT monitor gamut (shown as a wire frame) and an inkjet printer gamut
(shown as a solid) in CIELAB space. (a) Top view along the L3 axis. (b) and (c) Perspective
projections.

does not see the color of the individual phosphor sites but The light emitted by each phosphor has a constant
only a spatial average of the light from the individual red, spectral composition independent of the driving voltages for
green, and blue phosphors in a small region. The CRT is the electron guns, except for an intensity scale factor. The
thus an additive color system. scale factor is a nonlinear function of the driving voltage. A

1096 PROCEEDINGS OF THE IEEE, VOL. 86, NO. 6, JUNE 1998

typical model for the radiant spectrum emitted by a CRT is Ideally, if one had CMY colorants that absorbed light
in nonoverlapping regions of the spectrum and no black
colorant, the above system of equations could be simplified
(5) and linearized. Actual physical colorants, however, have
significant “unwanted absorptions” resulting in significant
where is the emission spectrum of the red CRT overlap. In addition, due to nonlinear interactions between
phosphor, is the voltage applied to its associated electron the colorants in the printing process, (6) is often useful only
gun, is the maximum voltage for the red gun, and as a qualitative model of the printing process. The same
is the monitor-gamma for the red gun. The variables model (without the black colorant) is also used to represent
corresponding to the green and blue electron guns are both prints and slides from color photography.
defined similarly. The monitor-gamma is normally around Contone printers are rather sophisticated and expensive
2.2 for most color monitors (for all three channels). Since devices. Halftone color printers offer a significantly less
the above parametric model is only approximate, several expensive alternative with somewhat lower image quality.
modifications of it have been used by researchers [35]–[38]. These printers produce color images by placing halftone
The CRT phosphors define a set of additive primaries. dots composed of the CMYK colorants on paper. The
If the CMF’s corresponding to these primaries are used in colorants combine subtractively over the regions over which
color specification, they can be directly used to drive the they overlap, resulting in a mosaic of differently colored
electron guns, provided the signals are precorrected for the regions on paper. Due to the low-pass nature of the eye’s
power-law nonlinearity mentioned above. Typically, this spatial response, the effective spectrum seen by the eye is
is done in TV transmission. Since the operation involves the average of the spectra over a small angular subtense.
exponentiation to the power of , it is known as gamma Different colors are produced by varying the relative areas
correction. A similar operation can also be done for digital- of the CMYK halftone dots, and the concentrations of the
image data to be displayed on a computer CRT display. The colorants within a dot are not varied. Since the overlapping
quantization of gamma-corrected RGB tristimuli results in of colorant dots is subtractive and the averaging over the
wider quantization intervals at higher amplitudes where the spatial mosaic is additive, halftoning is a hybrid process.
sensitivity of the eye is also lower. Therefore, just like A model for halftone printers was first suggested by
speech companding, gamma correction of color tristimuli Neugebauer [42], who observed that the overlap of CMY
prior to quantization in a digital system (or transmission in colorant halftone dots produces eight primary colors, which
a limited bandwidth system) reduces the perceptibility of are now known as Neugebauer primaries. These include the
errors and contours in comparison to a scheme in which no paper (no ink), each of the three colorants (cyan, magenta,
gamma correction is used [39], [40]. and yellow), combinations of two colorants (red, green-
2) Color Printers: Unlike CRT’s, which emit their blue), and combinations of all three (black). Neugebauer
own light, color printers produce images on paper (or proposed that the tristimulus values of the print can be
transparency) for viewing under an external source of expressed as the weighted average of the tristimuli of the
light. Color printers can be classified into two types: 1) Neugebauer primaries with the weights equal to their frac-
continuous-tone (contone) printers and 2) halftone printers. tional area coverages. The model can readily be generalized
Both contone and halftone printers produce color images from the three-colorant case to the case of four or any
by covering the paper with cyan, magenta, and yellow arbitrary number of colorants. Due to the penetration and
colorants that absorb light (largely) in the red, green, scattering of light in paper, known as the Yule–Nielsen
and blue spectral regions, respectively, and (commonly) effect (or alternately as the optical dot gain) [43], [44],
a black colorant that absorbs light uniformly over the the simple Neugebauer model does not perform well in
visible spectrum. practice. As a result, several empirical modifications have
Contone printers are subtractive color devices. They use been suggested for the model [45]. Recently, considerable
different concentrations of the cyan, magenta, and yellow success has been demonstrated in using a spectral version
colorants to control the absorption in the red, green, and of the Neugebauer model with empirical corrections for the
blue regions and thereby produce different colors. A simple Yule–Nielsen effect [46]–[48], in which the reflectance of
model of the subtractive process represents the spectrum of a CMYK halftone print is modeled as
reflected light from a contone print as [1], [41]
(7)
(6)

where is the viewing illuminant spectrum, is where and are, respectively, the spectral re-
the paper reflectance, are the concentrations of flectance and the fractional area of the th primary (there
the colorants (normalized with respect to the maximum), are 16 primaries for a four-colorant (CMYK) system) and
and are their optical densities corresponding to is the empirical Yule–Nielsen correction factor.
the maximum concentrations. The digital values used to Normally, the fractional areas are themselves related
drive the printer determine the concentrations and to the fractional areas and covered by the
thereby produce different colors. cyan, magenta, yellow, and black colorants, respectively.

SHARMA et al.: COLOR IMAGING FOR MULTIMEDIA 1097

The relations determining the from and C. Computer-Generated Imagery
depend on the method used to generate the halftone images. The last section focused on natural image scenes and
Equations for these relations were first derived by Demichel hard-copy prints of these images. In the areas of multimedia
[49] by assuming independent random coverage for the applications, the computer is an alternate, rapidly growing
colorants. Rotated halftone screens [50] that approximate source of color images. Computer-generated scenes and
this random independent coverage assumption are often animations are now common in video games and movies
used for printing because they are robust to commonly oc- [63], [64]. These applications strive for realistic images and
curring registration (alignment) errors between the different therefore have to model the interactions between the lights
colorants’ images (separations). Alternate halftone screens and the objects in the scene using both geometric optics and
have different characteristics and consequently different colorimetry to produce proper color, shading, and texture
relationships between the fractional colorant coverages and for the surfaces involved [65, ch. 14]. For these applications
the primary areas [51], [52]. to be successful in multimedia applications, it is necessary
that the color information be recorded in a form that permits
B. Image Characteristics accurate rendering of the image on multiple output devices.
This is particularly relevant in recording on movie film,
The spectral radiance or reflectance of an image carries which has entirely different characteristics from the CRT
the most complete color information. Most color spectra [66].
are fairly smooth, and a 10-nm wavelength sampling is Color can also be used to effectively encode information
normally sufficient for representing these spectra [53]. for presentation and visualization. Bar graphs and pie charts
The spectra of fluorescent gas-discharge lamps have sharp used in presentations are simple examples of the use of
spectral peaks and therefore require a higher sampling rate colors to distinguish different regions. The capabilities of
or alternate model-based approaches [54]–[56]. A 10-nm color encoding are also exploited more fully in applications
sampling of spectra in the 400–700-nm region provides that use color to represent quantitative differences in the
samples3 for each spectrum. Thus, color spectra data meaningfully. Examples of such applications include
lie in a 31-dimensional vector space. Color spectra are visualization of multispectral/stereo data using pseudo-
usually smooth, and therefore, they do not really exhibit color images and use of color to portray 3-D relationships
31 degrees of freedom. A principal-components analysis of in multidimensional data bases. The proper use of color in
common reflectance spectra reveals that a large fraction of these and other multimedia applications greatly enhances
the energy in these spectra is contained in a few principal their effectiveness [65, p. 601], [67]. Usually, the color
components [57]. Several researchers have investigated this images are designed for viewing on a CRT monitor and tend
problem and have estimated that color-reflectance spectra to exploit the full capabilities of the monitor. When printing
can be represented by using anywhere between four and 12 these images, it is necessary to preserve their information
principal components depending on the accuracy desired and effectiveness. This is often a significant challenge
[58]–[61]. because the printer gamuts are significantly smaller than
The spectral-reflectance characteristics discussed in the monitor gamuts. Some of these problems and attempted
last paragraph covered almost the entire range of naturally solutions will be mentioned in Section IV-D.
occurring reflective objects.4 Since most color-reproduction
systems exploit trichromacy and have only three addi-
tive/subtractive primaries, their spectra often have even D. Image-Recording Systems
fewer degrees of freedom. Thus, the spectra from a CRT Though the spectral radiance/reflectance carries the most
display lie in a 3-D space defined by the spectra of the complete color information for an image, it is usually both
phosphor primaries. The same is true of all additive systems expensive and unnecessary to record this complete informa-
that utilize three primaries. For subtractive and hybrid tion. Instead, typical color-recording devices employ only
systems, though, there are often only three primaries; they three channels and exploit trichromacy of the eye or the
are not confined to a 3-D space due to the nonlinearities in characteristics of the original image medium to obtain color
the subtractive process. For hybrid systems, the Neugebauer information.
primaries may be viewed as additive primaries, but typi- Photography was the first widely available system for
cally, the spectra of CMYK prints have fewer significant recording color images. Color photographs and slides
principal components than the 16 primaries in (7). are still one of the primary input sources for digital-
imaging systems. Color photography film contains three
light-sensitive emulsion layers that are exposed by the
3 The 400–700-nm interval is chosen for illustration, and the exact
red, green, and blue spectral components of the incoming
interval is not central to the argument here. Typical color-measurement
instruments report spectra over this or a larger wavelength region. light. Through a process of development and printing,
4 The color spectra of some naturally occurring objects arise from color images are produced by using cyan, magenta, and
interference phenomena, which yield nonsmooth spectra that are not yellow colored dyes in accordance with the subtractive
represented well by the principal components approach. Examples of principles discussed earlier. Due to inherent nonlinearities
such spectra are colors produced due to multiple film interference in
certain minerals and iridescent colors on some bird feathers and in shells and limitations in the chemical processes involved, exact
containing calcium carbonate [62, pp. 261, 267]. colorimetric reproduction is difficult in photography.

1098 PROCEEDINGS OF THE IEEE, VOL. 86, NO. 6, JUNE 1998

In addition, due to the adaptation of the eye, exact image by scanning the image over a raster of scan lines. In
colorimetric matches are also not desirable. Therefore, recordings for television, the raster lines are interlaced [73].
color photography normally attempts preferred color While this approach works fine for displaying video, usually
reproduction instead of accurate colorimetry [68, p. 191]. restoration schemes are required to obtain still images for
Since the original scene is rarely available for side-by-side printing/display [74].
comparison, this usually produces color images that are It is necessary to recognize that three channels are not
acceptable. sufficient for rendition of a natural reflectance scene under
In digital color cameras, the film is replaced by an ar- multiple viewing illuminants, and therefore additional chan-
ray of light-sensing charge-coupled-device (CCD) elements nels should be used if modification of viewing conditions or
placed in 2-D arrays. The analog output from the light- scene parameters is anticipated. Recently, researchers have
sensing elements is quantized to obtain a digital image. developed hyperspectral cameras that capture many image
Different schemes may be used to achieve the spatial- channels [75]. Currently, these devices are experimental
sampling and color-filtering operations concurrently. One and have been used in color-constancy research [76] and
arrangement uses three CCD arrays with red, green, and in simulation of digital cameras [77].
blue color filters, respectively. In such an arrangement, 1) Requirements for Accurate Color Recording: The color-
precise mechanical and optical alignment is necessary to recording process in digital cameras, scanners, and video
maintain spatial correspondence between the images from cameras can be represented by very similar mathematical
the different color channels. For economy and in order to models. In a manner analogous to (1), the output from
avoid the problems of registering multiple images, another these devices in response to a radiant spectrum can be
arrangement, commonly referred to as a color filter array represented as , where is the vector of
(CFA), uses a color filter mosaic, which is overlaid on measurements from the device, is the 3 matrix of
the CCD array during the semiconductor processing steps. spectral sensitivities of the (usually three) device channels,
Since the green region of the spectrum is perceptually more and is the measurement noise [1].
significant, such mosaics are laid out so as to have green, From this model, it can be readily inferred that in the
red, and blue recording pixels in the ratio 2 : 1 : 1 or 3 : 1 : 1 absence of noise, exact CIE XYZ tristimulus values can
[69], [70]. Image-restoration techniques are then used to be obtained from the data recorded by the device if there
reconstruct the full images for each of the channels [71], exists a transformation that transforms the sensor response
[72]. matrix into the matrix of CIE XYZ color-matching
For recording color images that are available as prints, functions, [41]. For devices using three channels, this
slides, negatives, or transparencies, color scanners are used. reduces to the requirement that be a nonsingular linear
These devices operate on principles identical to those of the transformation of . This fact has been known for some
digital cameras except that they include an illuminant for time and is referred to as the Luther–Ives condition [78],
use in the scanning process. The most common design uses [79]. Recent reiterations of this result can be found in [80]
an array of three linear CCD sensors with red, green, and and [81]. A device that satisfies (generalizations of) the
blue color filters. The linear array covers one dimension Luther–Ives condition will be said to be colorimetric.
of the scanned image and is scanned across in the other Due to practical design constraints, the Luther–Ives con-
direction to obtain a complete image record. Since color dition is rarely satisfied in practical devices. Besides, in
slides and negatives have a very high spatial resolution and the presence of noise, the condition does not necessarily
dynamic range in comparison with color prints, scanners imply optimality [82], and devices satisfying this condition
for scanning these media are designed with high spatial can provide colorimetric information only under a single
sampling rates and large dynamic range. Also, these de- viewing illuminant. Consequently, several researchers have
vices are often designed not for direct colorimetry (see proposed measures for evaluating the sensitivities of a
Section III-D1) but for determining the densities of the color-recording device. The first such measure was pro-
cyan, magenta, and yellow colorants in the original. Thus, posed by Neugebauer [83]. For a single filter, Neugebauer’s
they employ color filters with narrow transmittance bands quality factor measures the degree to which it is a linear
centered around the absorptance peaks of the colorants. A transformation of the CMF’s. Vora and Trussell [6], [84]
change in the colorants can therefore produce significant extended the measure to an arbitrary number of device
changes in the device’s response, resulting in poor color channels and viewing illuminants. A more comprehensive
recording unless the device is recalibrated for the new figure of merit that accounts for both the device noise
colorants. This is a significant problem since the colorants and the nonlinearities in the color-perception process was
from different manufacturers vary significantly. proposed in [85]. This work also gave a unified framework
Cameras for capturing video images record a number for all these measures and work by other researchers in
of image frames per second. These are then displayed in designing color filters that did not explicitly use a quality
succession on the display device to reconstruct the tempo- measure [86], [87].
rally varying scene. As far as color recording goes, these It may be emphasized here that for a rendition of the
devices also typically employ three red, green, and blue recorded reflectance scene under multiple viewing illumi-
channels and are not very different from digital cameras. nants, more than three channels are usually necessary. A
Older designs use analog electronics that obtained the 2-D simple generalization of the Luther–Ives condition would

SHARMA et al.: COLOR IMAGING FOR MULTIMEDIA 1099

require 3 channels for different viewing illuminants. are opponent chrominance channels similar to and ,
In practice, however, between four and seven optimally respectively.
designed spectral channels provide sufficient accuracy for
common viewing illuminants [86], [88]. F. Device Color Spaces
Note that in recording color images digitally, both the
spectrum and the spatial dimensions need to be “sam- The measurements from color-recording devices and the
pled.” The different quality measures mentioned above control values for color-output devices are the color repre-
consider only the spectral sampling aspect of the above sentations of recorded images or the images to be repro-
problem. These are therefore suitable for evaluating the duced, respectively. Hence, it is common to say that these
color-recording fidelity for large patches of uniform color values represent colors in the device’s color space. Thus,
and do not represent the complete performance for the typically, there are RGB color spaces for scanners, cam-
image-recording system. This is particularly true in CFA eras, and other input devices, and there are CMY/CMYK
camera devices, where the spatial and spectral sampling color spaces for color printers. Unlike the CIE and other
are performed jointly. color spaces mentioned in Section II-A, most of these
color spaces are not standard and cannot be directly used
E. Quantization and Coding for the meaningful archival/communication of image data.
However, if these device spaces are related to the standard
Color images recorded with the different input devices
color spaces in a clear, unambiguous way, these can also be
described in the last section need to be quantized for digital
potentially used for the exchange of color information. An
processing. Both scalar- and vector-quantization techniques
advantage of such a scheme is that no transformations of the
can be used in the quantization of color data. For simplicity,
image data are required for display if the data are specified
most color devices do independent quantization of the
in the native color space of the device. Since a large
RGB channels with 8–12 bits per channel, with either
fraction of the images on the World Wide Web are primarily
uniform quantizers or companded quantizers that perform a
targeted for display on CRT monitors that have very similar
gamma correction before the quantization. As mentioned
characteristics, a new standard color space, sRGB, has
earlier, the gamma correction significantly reduces the
been proposed for use based on these characteristics [96].
perceptibility of quantization errors, particularly in the
The sRGB color space is basically a gamma-corrected
8-bit devices. For computer color displays based on a
tristimulus space that uses the CRT phosphors as primaries
frame-buffer architecture [89], often only 8, 12, or 16
for determining the CMF’s. In addition, the sRGB standard
bits of video memory are allocated to each pixel, thereby
includes provisions for specifying the viewing conditions
allowing simultaneous display of only 2 , 2 , or 2
(white-point chromaticities, image surround, flare, etc.).
colors, respectively. Vector-quantization techniques have
therefore been extensively used for displaying images on
these devices. A brief survey of some of these methods can IV. COLOR MANAGEMENT AND CALIBRATION
be found in [1]. For proper color reproduction, the input and output
With the proliferation of digital color imagery, the prob- devices involved need to be calibrated. Historically, the
lem of coding color images for transmission and storage systems used for color reproduction were calibrated in
has gained increased importance. It was recognized early on a closed-loop configuration. As shown in Fig. 7, in a
that the highly correlated RGB spaces were not suitable for closed-loop configuration, the complete system is calibrated
independent coding [90]. Consequently, most of the meth- from input through output. Thus, for color photography,
ods transform the data into a luminance channel, and two the film sensitivities, dye absorptances, and developmental
chrominance channels that are then coded independently. A interactions were appropriately chosen so as to result in
luminance chrominance space also allows coding schemes acceptable reproduction. In offset printing, the scanner was
to exploit the properties of human vision by allocating used to generate CMYK “separations” that were suitable for
significantly less bits to the high-frequency chrominance generating halftone prints. With the increased use of digital
components, which are perceptually less significant. computers and the evolution of desktop printing, it became
The most prevalent compression scheme at present is the obvious that such an approach has severe limitations. In
Joint Photographic Experts Group standard for still images particular, as the number of devices increases, calibrations
[91] and the Moving Picture Experts Group standard for for each input-output device pair are difficult to construct
video data [92]. These standards are both based on the and maintain. In addition, since the calibrated data in a
discrete cosine transform [93]. While these standards do closed-loop calibration scheme is specific to one output
not explicitly specify the color spaces to be used, in current device, it is not suitable for archival or exchange with
implementations, it is common to use the YCrCb space [94], devices outside the system.
with the Cr and Cb components subsampled by a factor With the growth of networking and increased exchange of
of two along both spatial dimensions [95]. The YCrCb color-image data between geographically divided systems,
color space is a luminance chrominance color space based it was recognized that several of these problems can be
on gamma-corrected RGB that has been proposed for use solved by calibrating each device to a standard device-
as a standard in high-definition TV. The Y component is independent (DVI) color space, which can then be used
a luminance channel similar to , and the Cr and Cb for the exchange of data between different devices and for

1100 PROCEEDINGS OF THE IEEE, VOL. 86, NO. 6, JUNE 1998

 tems and algorithms that use the calibrations effectively to
achieve desired results. These components are the subject
of the remainder of this section.

A. Color-Measurement Instrumentation
A number of devices are available for the measurement
of color. The correct instrument for a particular applica-
tion depends on several factors, including the cost of the
instrument, the accuracy desired, the variety of viewing
conditions for which a calibration is desired, and the
number and type of devices that need to be calibrated.
1) Spectroradiometers: A spectroradiometer is a device
that measures the power of optical radiation as a function
of wavelength. The measured spectrum provides the most
complete description of color from which color descriptors
in different color spaces can readily be computed. Most
spectroradiometers contain an optical grating and a linear
CCD array. Typical devices used for colorimetric work
report data from about 380 to 780 nm at increments ranging
Fig. 7. Closed-loop system calibration. from 2 to 10 nm. For most color work (including that using
fluorescent lamps as sources), 2-nm sampling is sufficient
[53]. Spectroradiometers can be used to measure both self-
luminous and reflective objects. For the measurement of
reflectance spectra, a spectrally smooth (preferably white)
light source and a spectrally referenced sample (preferably
white) are required. A comparison measurement between
the known sample and the sample under question is made
under identical conditions, allowing the determination of
the unknown sample’s reflectance spectrum, from which
the color descriptor under any viewing illuminant can be
obtained. Normally, the exact spectrum of the illuminant
used with the spectroradiometer in the reflectance measure-
ment is immaterial, provided it has sufficient power over the
spectral range of interest. However, for the measurement of
fluorescent materials, the power in the ultraviolet regions is
also important, and the illuminant must closely approximate
the desired standard illuminant under which colors are to
be computed [97].
While spectroradiometers can measure both luminous and
nonluminous objects, they are expensive, larger and less
portable than the other instruments, and generally more
difficult to operate due to their multiple operating modes.
In particular, it can be difficult to set up a spectroradiome-
ter for measuring reflectance samples under controlled
conditions. There are automated systems for performing
measurements on multiple samples using an - stage and
Fig. 8. Device-independent color calibration.
a single light source.
2) Spectrophotometers: A spectrophotometer is a device
archival. As shown in Fig. 8, in these systems, the data
for measuring the spectral reflectance of an object. Unlike
from an input device are converted to a device-independent a spectroradiometer, a spectrophotometer does not measure
color space and then transformed into the device space of self-luminous objects. Therefore, spectrophotometers are
the target device for reproduction. useful for the calibration of printers and scanners but not of
To enable proper management of color, several compo- CRT displays. Spectrophotometers have their own internal
nents are required. Color-measurement instrumentation is light source, which illuminates the sample under measure-
necessary for calibrating input and output systems. Standard ment. There are many geometries for the sensor/illuminant
formats for the storage and communication of these device combination, and each may be optimal for different applica-
calibrations are required so that different applications can tions. For publishing applications, the sensor and illuminant
make use of the calibrations. Also necessary are sys- in the device are often set up for a 45 /0 condition,

SHARMA et al.: COLOR IMAGING FOR MULTIMEDIA 1101

 find use primarily in the calibration of CRT’s when the
chromaticity of the CRT phosphors is known.

B. Calibration and Profiles

Calibration of a color-imaging device relates its in-
put/output to DVI color values. For an input device, calibra-
tion provides a mapping from device-measurement values
(e.g., scanner RGB) to DVI color descriptors (e.g., CIE
XYZ, CIELAB, etc.). For an output device, the calibration
process yields a mapping from DVI color descriptors to
device control values (e.g., CMYK, monitor RGB), which
produce those color descriptors.
Fig. 9. 45 /0 measurement geometry. 1) Input-Device Calibration: To calibrate a scanner, the
first step is to select a collection of color patches that span
the gamut of interest. Ideally, these colors should not be
as shown in Fig. 9. There are devices that illuminate the metameric for the scanner or the eye (under the illuminant
sample diffusely and measure off axis, allowing an option for which the calibration is being produced). Metamerism
to include or exclude the specular component from the is defined as the property in which different spectra map
sample. Some devices contain an optical grating; others to the same values under a set of sensitivity functions.
contain a filter wheel; and yet others use spectrally different These patches are measured using a color-measurement
light sources to illuminate the sample. Most devices report instrument, such as a spectrophotometer or a colorimeter,
data from about 380 to 730 nm in 10–20-nm increments. which will provide the DVI color values , where
Since reflectances of natural objects are smooth functions denotes the number of patches. Any of the (DVI) color
of wavelength and have low dimensionality, as discussed spaces such as CIE XYZ, CIELAB, etc. can be used for this
in Section III-B, these sampling rates are almost always purpose. The use of CIELAB is common since this space
sufficient [53]. includes information on the viewing illuminant in the white
Spectrophotometers measure reflectance as the ratio of point. The patches are also measured with the scanner to
two (uncalibrated) spectroradiometric measurements [1], obtain the scanner measurements .
[97] using the same principle that was described in the last To determine the CIE values for an arbitrary measured
section. They therefore come with a standard sample reflec- patch, the collected data are used to construct an interpo-
tor, which is used to periodically recalibrate the instrument. lating function, which maps from the space of scanner-
Modern spectrophotometers are handheld, self-contained, measurement values to the chosen DVI color space. This
and easy to operate. Some include automated stages for function, , can then be used to relate any of the
measuring multiple samples on a page. A disadvantage of scanner RGB values to colorimetric XYZ values or LAB
most spectrophotometers is their cost, which, while less values, i.e., . Normally, a parametric form
than that of a spectroradiometer, can often be prohibitive. is chosen for , and the parameters are determined
The devices are also fragile, especially those with optical through regression. Several different schemes have been
gratings or filter wheels. used for this purpose, ranging from straightforward linear
3) Colorimeters and Photometers: As suggested by its and polynomial regression to neural networks [98, ch.
name, a colorimeter measures color tristimuli and reports 11]. Usually, is complicated and computationally
these as color values in CIE XYZ, CIELAB, or related expensive. For this reason, usually is used to produce
color spaces. Some colorimeters have an internal light a finely sampled lookup table (LUT) from which at
source for the measurement of color of reflective objects, arbitrary points is obtained by using simple interpolation
while others measure only self-luminous or externally schemes [98, ch. 4].
illuminated objects. For some devices, tristimulus values for Calibration for digital cameras and video cameras is
the sample under a few different illuminants are available. usually done in a similar fashion by using a target of
Most colorimeters are small handheld devices with no patches with known reflectances or color values. Since the
moving parts, and they achieve their spectral separation by eye is very sensitive to deviations from the neutral (achro-
way of color filters or with spectrally different light sources. matic) colors, sometimes an additional one-dimensional
Colorimeters are less expensive than spectrophotometers (1-D) transform is included on each of the RGB channels so
and spectroradiometers, but they do not provide the detailed that the line corresponds to neutral colors. This
spectral information that allows the calibration of a printer procedure is commonly referred to as gray/white balancing.
for an arbitrary viewing illuminant. Those that measure 2) Output-Device Calibration: For calibrating an output
self-luminous sources are used in the calibration of CRT’s. device, a transformation from DVI color values to the space
Photometers are single-channel devices that provide a of device control values is required. This requires a two-
measurement of the luminance of a self-luminous or step procedure. In the first step, the printer characterization,
externally illuminated object. They are inexpensive and which determines the forward transformation from printer

1102 PROCEEDINGS OF THE IEEE, VOL. 86, NO. 6, JUNE 1998

control values to DVI color values, is determined. Then, this Color Consortium profile format specifies a wide variety
forward transform is used in the next step to determine the of input and output device profiles suitable for efficiently
inverse mapping from DVI color values to device control representing the color-calibration information. The use of
values. a standardized format allows the profiles to be used by
Since CRT monitors are represented well by the para- different applications from different manufacturers.
metric models described in Section III-A1, the forward
characterization of these devices is usually done by deter- C. Color-Management Systems (CMS’s)
mining the model parameters from a few measurements. It
A CMS is responsible for interpreting the device profiles
can also be readily seen that due to their additive nature, the
and performing the appropriate transformations to/from the
inverse transformation from CIE XYZ (or other tristimulus)
DVI space. The goal of a CMS is to provide predictable
values to the CRT control voltages can be computed by
and consistent color without requiring specialized skills
means of a simple matrix transformation followed by a 1-D
from the user. Thus CMS’s tie together device profiles
transform for the gamma correction. This scheme is used
with color-matching modules that use these device profiles
in all cases, except those requiring the highest accuracy,
to transform device-dependent image data to DVI color
for which LUT-based schemes may be used and additional
spaces or to the device color spaces of target output devices
corrections may be made for the surface reflection (flare)
on which the images are to be displayed. In addition, the
from the monitor screen. In addition, it may be necessary
CMS provides the user with flexibility to choose different
to correct for the significant spatial nonuniformity over the
rendering intents for different images. Thus, for instance, in
CRT screen [1], [99].
reproducing a company logo, a perfect colorimetric match
For the forward characterization of printers, commonly
is usually desired; in producing bar graphs and pie charts
an empirical scheme similar to that described for scanner
for presentation, it is desirable that the colors be highly
calibration is used. By selecting a set of printer control
saturated for maximal impact; and in reproducing pictorial
values covering the range of allowable control
images, it is desirable that the closest perceptual match be
values, measuring the corresponding DVI color values
obtained (which will depend on viewing conditions). Color-
, and using some interpolation scheme, the forward
management functions can be performed at several different
mapping from control values to DVI color values
phases of the imaging process: in the devices (e.g., Adobe
is determined. For halftone printers, alternately, the Neuge-
Postscript level 2 for printers), in device drivers (e.g., Canon
bauer models mentioned in Section III-A2 have also been
Colorgear), in applications (e.g., Adobe Photoshop), or in
used. Since even the models are nonlinear and not readily
the operating system (e.g., Apple’s ColorSync). The notion
invertible, for the inverse mapping , invariably an
of embedding color management in the operating system
interpolating function is used to get a finely sampled LUT.
has the potential of making the process transparent to
Due to the four degrees of freedom in the control values
the end user and is therefore gaining acceptance rapidly,
of CMYK printers, there exist multiple control values that
with several vendors of operating systems for desktop
result in the same printed color. Since this poses a challenge
and workstation computers incorporating CMS’s into their
in obtaining a smooth inverse mapping, often the inverse
products.
determines the amounts of three virtual CMY colorants,
Ideally, with color management, one could accurately
from which the CMYK control values are obtained by
transfer color information from one medium (e.g., CRT)
some functional relation. This process is viewed as the
to another (e.g., print). Unfortunately, this is an extremely
incorporation of black ink and removal of underlying CMY
difficult task due to two reasons: 1) there are significant dif-
inks and is commonly referred to as undercolor removal
ferences in gamuts of different devices (this was mentioned
(UCR). Typically, the UCR mapping from the virtual CMY
and demonstrated in Section III-A) and 2) the difference in
values to actual CMYK control values is designed so as
typical viewing conditions for different media implies that
to render achromatic colors better, reduce total colorant
a simple colorimetric match does not give an appearance
amounts (for faster drying/better adhesion to paper), and
match. There is therefore significant interest in gamut-
(in some cases) reduce the use of the expensive CMY
mapping algorithms that map the colors in an image to
colorants.
suitable colors that can be reproduced on the target device.
3) Device Profiles: To make the calibration transforma-
Methods that model the adaptations of the eye and allow the
tions available to different applications that wish to use
computation of appearance-matched images under different
them, the calibration transformation for each device is
viewing conditions are also an active area of research.
stored in a device profile. In the early days of color
management, different manufacturers used their own pro-
prietary formats for the storage of these profiles, which D. Gamut Mapping
were therefore useful only for applications from the same Gamut mapping is the process of mapping the displayable
manufacturer. To realize the full benefits from the DVI colors from one media to those of another media. As
calibration of devices, the desktop-publishing industry is defined earlier, the ideal goal for color matching depends
increasingly moving toward open systems. A standard on the type of the image and the intent of the user. The
format for the storage of device profiles has been de- problem is probably most acute for pictorial images for
fined and is being widely adopted [100]. This International which ideally one would like to map the colors so as

SHARMA et al.: COLOR IMAGING FOR MULTIMEDIA 1103

to obtain the best possible appearance match between the in a loss of fidelity and introduce visually noticeable
images on the different media [101], [102]. The strategy errors due to the accumulation of errors associated with
for gamut mapping could be either image dependent or im- finite precision arithmetic. Some CMS’s can cache input
age independent. Since image-dependent methods can use and output transformations, concatenate these, and perform
different strategies for different images, they often produce these in a single operation when the final image is desired.
better results. However, they are also significantly slower This not only saves time but also improves accuracy.
than image-independent techniques because they require a
fresh computation for each image, and are therefore seldom E. Appearance Matching
used in automatic gamut-mapping schemes. Several color
applications, however, indicate out-of-gamut colors in an Section II-D briefly discussed the adaptation in the eye in
image via a key color or a flashing highlight (on a CRT). response to changes in viewing conditions. Since different
color-reproduction media have different viewing condi-
This allows the user to perform a transformation such
tions, in reproducing images across these different media,
that these colors are mapped to his satisfaction. Such an
an appearance match instead of a pure colorimetric match
approach, however, requires considerable operator skill in
is usually desired. The simplest instance of appearance
order to obtain satisfactory results.
matching is the white-point matching method based on
The simplest technique of gamut mapping (particularly
the Von Kries transformation, as mentioned in Section II-
for CRT monitors) is clipping in the space of device
D1. This transformation converts tristimuli into a space of
control values. Since the space of control values is not
cone responses and applies a diagonal correction matrix
a UCS, clipping of control values does not yield the
that equates the white points under the two viewing con-
closest printable color (as perceived by an observer). To
ditions [7, p. 432]. White-point matching is often used in
remedy this problem, a simple extension would be to
color-imaging applications, and there is support for it in
map out-of-gamut colors to the nearest in-gamut color
several color-management applications and packages, such
in a UCS. While this approach offers significantly better
as Postscript.
results than the device space clipping [103], it can often
In addition to white-point adaptation, there are several
result in unacceptable hue shifts that are perceptually very
well-characterized psychophysical effects that change with
objectionable. An additional limitation of this and the
changes in viewing conditions. In particular, it is well
clipping approach is that smoothly varying regions beyond
documented that the apparent contrast (perceived intensity
the device gamut can potentially be mapped to a single
gradient) of an image decreases in a dark surround in
color, creating undesirable abrupt edges in the previously
comparison to a bright surround [68, pp. 49–57]. Often, this
smooth regions and causing loss of significant information
change in contrast is modeled by relating the luminance to
(such as shape from shading).
the perceived lightness as a power-law relation, similar to
Another approach used for gamut mapping is a gamut-
that for CIELAB in (3), with the exponent’s increasing as
compression algorithm, which compresses all the colors
the surround gets brighter [68, pp. 56–57], [107]. Thus,
in the image in a manner that reduces the colorimetric
“gamma correction” has also been used extensively to
dynamic range in the image while ensuring that the colors
compensate for these effects, in addition to correction for
can be reproduced. For example, one could move all the
monitor nonlinearity.
colors in the image toward one point, such as a midgray,
There has been considerable research in defining color-
until all the colors in the image are within the device
appearance models that account for chromatic adaptation,
gamut. Unlike the clipping approach, this method will
influence of surround, and other psychophysical phenomena
retain some of the variation in smoothly varying image
that affect the perception of images [24]. Comparative
regions that are beyond the device gamut. In addition to
results of several different color-appearance models in
UCS’s, notions of hue, chroma, and saturation/value are
imaging applications can be found in [24] and [108]. These
extremely useful in gamut-mapping research since these
models hold tremendous potential for use in cross-media
can be directly related to viewers’ objections to artifacts
color reproduction. The CIE is involved in an ongoing
produced by gamut-mapping algorithms and can be used
attempt to define a standard appearance model for use
in appropriately choosing the clipping and compression.
in imaging applications. An interim version of the model
In Section II-D1, it was pointed out that in addition to
is available [109], which is expected to be subsequently
being a UCS, the CIELAB space allows the computation
refined.
of correlates of lightness, hue, and chroma. As a result,
CIELAB has been used extensively in gamut-mapping
research. A recent comparison of a number of clipping- and V. ILLUSTRATIVE EXAMPLE
compression-based gamut-mapping schemes in CIELAB This paper has reviewed the background of color sci-
space can be found in [104], [105]. ence and color terminology, discussed the instrumentation,
The transformations to and from the perceptual spaces are and discussed how to communicate color among various
usually nonlinear, and speed is often an issue [106]. For this devices. Let us consider how to put all of these things
reason, some CMS’s may use a high-resolution LUT along together in an illustrative application. Rather than review in
with a linear interpolator to perform the transformation. detail published applications that demonstrate a few parts
Performing multiple transforms on the data can result of the color process, a fictitious but plausible application is

1104 PROCEEDINGS OF THE IEEE, VOL. 86, NO. 6, JUNE 1998

described that uses most of the concepts and tools discussed other objects may shift. The tristimulus values can be
in the previous sections. computed via (8). However, note that to view the image
In both animation and simulation, the problem of chang- correctly, the white point of the monitor should be changed
ing lighting conditions arises. In the case of animation, the to correspond to the new viewing conditions. To obtain
user specifies the lighting conditions and surface properties the most accurate color transformation, more than three
of the objects. The process that is addressed here is creating color bands of reflectance should be used to describe the
a new image from an original and reproducing it accurately reflectance of the pixel. If the color transformation is done
on a different medium. We will assume a case with a single from the three lab values, some error is inevitable. This
light source. error can be minimized by using a statistical description of
Since a single illumination is used, the radiant spectrum the spectra of the objects in the image [110].
of every object can be represented by the model used in The designer is happy with the appearance of the image
(2). It is not required that every object be represented by a on the monitor. Now it is desired to print the image on film
31-dimensional vector since, as mentioned previously, most for display. Again, the process of color communication is
reflective objects in the real world have many fewer degrees started. The film will be projected so that the white point
of freedom. To obtain accurate color under multiple illumi- of the display is fixed by the lamp and screen combination.
nants, however, more than three dimensions are required. The pixel values are now transformed to exposure values,
In any case, all of the vector equations and principles which control the density of the dyes on the film. Again,
discussed earlier hold, regardless of dimensionality. The gamut mapping must be done. Since this is the final display
change in color for each pixel can be computed directly for the design procedure, this is the gamut that should
from the vector that describes the pixel without auxiliary control the entire process.
computation, which requires geometrical considerations. In
this case, the new color is given by VI. SUMMARY
(8) To use color in multimedia applications effectively, one
needs to understand color from a perspective that goes
where is the vector representing the reflective spectra of beyond the simple RGB “color space” model often used
the pixel, represents the new illumination [see (2)], in computer applications. This paper aims at explaining
is the CMF, and is the vector of new tristimulus values. the issues involved in using calibrated and accurate color
Let us assume that an image or sequence of images has for multimedia applications. The basics of color science,
been created that correspond to midday illumination. It is management, and calibration are explained along with the
desired to render the scene as it would appear in the late challenges involved in reproducing images on a multiplicity
afternoon. The data that describe the image, not the objects of output devices. The use of the concepts is demonstrated
in the image, are the tristimulus values of each pixel. As through an illustrative example.
is common, let us use CIELAB. Since the CIELAB values
depend on a normalizing white point, this point must be
ACKNOWLEDGMENT
specified for a complete definition of the image. In this
case, the white point would correspond to the midday The authors wish to thank Dr. R. Balasubramanian for
(approximately D6500) spectra of the assumed illuminant. his helpful suggestions and pointers and J. Stinehour for
The image was designed using a computer and thus making available the data used in Figs. 5 and 6.
viewed on a CRT monitor. To use the monitor for eval-
uation, the monitor must be calibrated using the methods REFERENCES
described in Section IV-B2. The image-lab values, along [1] G. Sharma and H. J. Trussell, “Digital color imaging,” IEEE
with the white point, are converted to monitor RGB values Trans. Image Processing, vol. 6, pp. 901–932, July 1997.
using the calibration. The transformation from lab to RGB [2] H. J. Trussell, “Application of set theoretic models to color
systems,” Color Res. Appl., vol. 16, no. 1, pp. 31–41, Feb. 1991.
may also require gamut mapping. The designer who created [3] B. K. P. Horn, “Exact reproduction of color images,” Comput.
the original image should take into account the gamut of Vis., Graphics Image Process., vol. 26, pp. 135–167, 1984.
the devices that will be used. It is prudent to avoid out- [4] J. B. Cohen, “Color and color mixture: Scalar and vector
fundamentals,” Color Res. Appl., vol. 13, no. 1, pp. 5–39, Feb.
of-gamut colors as much as possible. The designer would 1988.
probably alter the colors, i.e., reflectances, of the objects to [5] H. J. Trussell and J. R. Sullivan, “A vector-space approach
obtain the best visual effect. to color imaging systems,” in Proc. SPIE Image Processing
The change of illuminant to late afternoon (approximately Algorithms and Techniques, K. S. Pennington, Ed., 1990, vol.
1244, pp. 264–271.
D2000) results in new lab values and a new white point. [6] P. L. Vora and H. J. Trussell, “Measure of goodness of a set
Recall that the white point is used as a compensation for the of color scanning filters,” J. Opt. Soc. Amer. A, vol. 10, no. 7,
adaptive ability of the eye. An example of this adaptation pp. 1499–1508, 1993.
[7] G. Wyszecki and W. S. Stiles, Color Science: Concepts and
is that white objects will appear white to the viewer under Methods, Quantitative Data and Formulae, 2nd ed. New York:
a wide range of illumination. Thus, despite the fact that Wiley, 1982.
the reflected spectrum from a white object in the afternoon [8] International Commission on Illumination, “Colorimetry,” Cen-
tral Bureau of the CIE, Vienna, Austria, CIE Pub. 15.2, 1986.
would have a much higher red component than at midday, [9] H. S. Fairman, M. H. Brill, and H. Hemmendinger, “How the
it will still appear white. However, the relative colors of CIE 1931 color-matching functions were derived from Wright-

SHARMA et al.: COLOR IMAGING FOR MULTIMEDIA 1105

 Guild data,” Color Res. Appl., vol. 22, no. 1, pp. 11–23, Feb. color monitor in terms of standard CIE coordinates,” Comput.
1997. Graph., vol. 17, pp. 315–321, July 1983.
[10] P. K. Kaiser and R. M. Boynton, Human Color Vision, 2nd ed. [36] R. S. Gentile, J. P. Allebach, and E. Walowit, “Quantization
Washington, DC: Optical Society of America, 1996. of color images based on uniform color spaces,” J. Imaging
[11] National Television Systems Committee, “NTSC signal speci- Technol., vol. 16, pp. 11–21, Feb. 1990.
fications,” Proc. IRE, vol. 42, pp. 17–19, Jan. 1954. [37] R. S. Berns, R. J. Motta, and M. E. Gorzynski, “CRT colorime-
[12] Society for Motion and Television Pictures, “Color temperature try, Part I: Theory and practice,” Color Res. Appl., vol. 18, no.
for color television studio monitors,” SMPTE Recommended 5, pp. 299–314, Oct. 1993.
Practice RP 37-1969, White Plains, NY, July 1969. [38] R. S. Berns, M. E. Gorzynski, and R. J. Motta, “CRT colorime-
[13] Society for Motion and Television Pictures, “Color monitor try, Part II: Metrology,” Color Res. Appl., vol. 18, no. 5, pp.
colorimetry,” SMPTE recommended practice RP 145–1987, 315–325, Oct. 1993.
White Plains, NY, June 1987. [39] F. Kretz, “Subjectively optimal quantization of pictures,” IEEE
[14] B. A. Wandell, Foundations of Vision. Sunderland, MA: Sin- Trans. Commun., vol. COM-23, pp. 1288–1292, Nov. 1975.
auer Associates, 1995. [40] C. A. Poynton, “‘Gamma’ and its disguises: The nonlinear
[15] C. S. McCamy, “Correlated color temperature as an explicit mappings of intensity in perception, CRT’s, film, and video,”
function of chromaticity coordinates,” Color Res. Appl., vol. SMPTE J., vol. 102, pp. 377–385, Dec. 1993.
17, no. 2, pp. 142–144, Apr. 1992. [41] H. J. Trussell, “DSP solutions run the gamut for color systems,”
[16] R. W. G. Hunt, “The specification of color appearance. I. IEEE Signal Processing Mag., vol. 10, pp. 8–23, Apr. 1993.
Concepts and terms,” Color Res. Appl., vol. 2, no. 2, pp. 55–68, [42] H. E. J. Neugebauer, “Die Theoretischen Grundlagen des Mehr-
Summer 1977. farbenbuchdrucks,” Zeitschrift für Wissenschaftliche Photogra-
[17] , “Color terminology,” Color Res. Appl., vol. 3, no. 2, pp. phie Photophysik und Photochemie, vol. 36, no. 4, pp. 73–89,
79–87, Summer 1978. Apr. 1937.
[18] , Measuring Color, 2nd ed. New York: Ellis Horwood, [43] J. A. C. Yule and W. J. Neilsen, “The penetration of light into
1991. paper and its effect on halftone reproduction,” in Proc. TAGA,
[19] E. Q. Adams, “A theory of color vision,” Psychol. Rev., vol. May 7–9, 1951, pp. 65–76.
36, pp. 56–76, 1923. [44] F. R. Clapper and J. A. C. Yule, “Reproduction of color with
[20] M. Mahy, L. Van Eyckden, and A. Oosterlinck, “Evaluation of halftone images,” in Proc. 7th Annual Tech. Meeting TAGA, May
uniform color spaces developed after the adoption of CIELAB 9, 1955, pp. 1–14.
and CIELUV,” Color Res. Appl., vol. 19, no. 2, pp. 105–121, [45] K. Sayangi, Ed., Proceedings of the SPIE: Neugebauer Memo-
Apr. 1994. rial Seminar on Color Reproduction, Bellingham, WA, Dec.
[21] F. J. J. Clarke, R. McDonald, and B. Rigg, “Modification to the 14–15, 1989, vol. 1184.
JPC79 color-difference formula,” J. Soc. Dyers Colorists, vol. [46] J. A. S. Viggiano, “Modeling the color of multi-colored
100, pp. 128–132, 1984. halftones,” Proc. TAGA, 1990, pp. 44–62.
[22] International Commission on Illumination, “Industrial color [47] R. Rolleston and R. Balasubramanian, “Accuracy of various
difference evaluation,” Central Bureau of the CIE, Vienna, CIE types of Neugebauer models,” in Proc. IS&T/SID Color Imaging
Pub. 116–1995, 1995. Conf.: Transforms and Portability of Color, Nov. 1993, pp.
[23] J. A. Von Kries, “Die Gesichtsempfindungen,” in Handbuch 32–37.
der Physiologie der Menschen, Tiel III, W. Nagel, Ed., vol. 3. [48] R. Balasubramanian, “Colorimetric modeling of binary color
Brunswick: Vieweg, 1905, pp. 109–282. printers,” in Proc. IEEE Int. Conf. Image Processing, Nov. 1995,
[24] M. D. Fairchild, Color Appearance Models. Reading, MA: pp. II-327–330.
Addison-Wesley, 1998. [49] E. Demichel, Procédé, vol. 26, pp. 17–21, 26–27, 1924.
[25] D. H. Kelly, “Spatial and temporal interactions in color vision,” [50] J. A. C Yule, Principles of Color Reproduction, Applied to
J. Imaging Tech., vol. 15, no. 2, pp. 82–89, Apr. 1989. Photomechanical Reproduction, Color Photography, and the
[26] E. Granger, “Uniform color space as a function of spatial Ink, Paper, and Other Related Industries. New York: Wiley,
frequency,” in Proc. IS&T 7th Int. Conf. Advances in Non- 1967.
Impact Printing Technology, 1991, pp. 309–322. [51] R. Balasubramanian, “A printer model for dot-on-dot halftone
[27] X. Zhang and B. A. Wandell. (1996). Spatial extension of screens,” in Proc. SPIE Color Hard Copy and Graphic Arts IV,
CIELAB for digital color image reproduction, Proceedings of J. Bares, Ed., 1995, vol. 2413, pp. 356–364.
the SID International Conference. [Online]. Available URL: [52] W. L. Rhodes and C. M. Hains, “The influence of halftone
http://white.stanford.edu:80/html/xmei/. orientation on color gamut and registration sensitivity,” in Proc.
[28] W. Frei and B. Baxter, “Rate-distortion coding simulation IS&T’s 46th Ann. Conf., May 1993, pp. 180–182.
for color images,” IEEE Trans. Commun., vol. COM-25, pp. [53] H. J. Trussell and M. S. Kulkarni, “Sampling and processing
1385–1392, Nov. 1977. of color signals,” IEEE Trans. Image Processing, vol. 5, pp.
[29] O. D. Faugeras, “Digital color image processing within the 677–681, Apr. 1996.
framework of a human visual model,” IEEE Trans. Acoust., [54] D. L. MacAdam, Color Measurement: Theme and Variations,
Speech, Signal Processing, vol. 27, pp. 380–393, Aug. 1979. 2nd ed. New York: Springer-Verlag, 1981.
[30] S. Daly, “The visual differences predictor: An algorithm for the [55] H. J. Trussell and M. Kulkarni, “Estimation of color under fluo-
assessment of image fidelity,” in Digital Images and Human rescent illuminants,” in Proc. IEEE Int. Conf. Image Processing,
Vision, A. B. Watson, Ed. Cambridge, MA: MIT Press, 1993, 1994, pp. III1006–1010.
pp. 179–206. [56] G. Sharma and H. J. Trussell, “Decomposition of fluorescent
[31] J. Lubin, “A visual discrimination model for imaging system de- illuminant spectra for accurate colorimetry,” in Proc. IEEE Int.
sign and evaluation,” in Vision Models for Target Detection and Conf. Image Processing, Nov. 1994, pp. II1002–1006.
Recognition, E. Peli, Ed. River Edge, NJ: World Scientific, [57] I. T. Jolliffe, Principal Components Analysis. Berlin, Ger-
1995, pp. 245–283. many: Springer-Verlag, 1986.
[32] G. Sharma, M. J. Vrhel, and H. J. Trussell. [58] J. Cohen, “Dependency of the spectral reflectance curves of the
(1998). Color figures for this paper. [Online]. Munsell color chips,” Psychronomic Sci., vol. 1, pp. 369–370,
Available URL: http://www4.ncsu.edu/gsharma/public 1964.
html/color multimedia/; http://www4.ncsu.edu/eos/users/h/hjt/ [59] L. T. Maloney, “Evaluation of linear models of surface re-
WWW/index figures.html. flectance reflectance with a small number of parameters,” J.
[33] R. Rolleston, “Visualization of colorimetric calibration,” in Opt. Soc. Amer. A, vol. 3, pp. 1673–1683, 1986.
Proc. SPIE Color Hard Copy and Graphic Arts II, J. Bares, [60] B. A. Wandell, “The synthesis and analysis of color images,”
Ed., 1993, vol. 1912, pp. 299–309. IEEE Trans. Pattern Anal. Machine Intell., vol. PAMI-9, pp.
[34] D. Kalra, “GamOpt: A tool for visualization and optimization 2–13, Jan. 1987.
of gamuts,” in Proc. SPIE Color Hard Copy and Graphic Arts [61] M. J. Vrhel, R. Gershon, and L. S. Iwan, “Measurement and
III, J. Bares, Ed., 1994, vol. 2171, pp. 299–309. analysis of object reflectance spectra,” Color Res. Appl., vol.
[35] W. B. Cowan, “An inexpensive scheme for calibration of a 19, no. 1, pp. 4–9, Feb. 1994.

1106 PROCEEDINGS OF THE IEEE, VOL. 86, NO. 6, JUNE 1998

[62] K. Nassau, The Physics and Chemistry of Color: The Fifteen “Optimization of sensor response functions for colorimetry of
Causes of Color. New York: Wiley, 1983. reflective and emissive objects,” IEEE Trans. Image Processing,
[63] B. Robertson, “Toy Story: A triumph of animation,” Comput. vol. 5, no. 3, pp. 507–517, Mar. 1996.
Graph. World, vol. 18, no. 8, pp. 28–38, Aug. 1995. [88] G. Sharma, H. J. Trussell, and M. J. Vrhel, “Optimal nonnega-
[64] D. Coco, “Breathing life into 3-d humans,” Comput. Graph. tive color scanning filters,” IEEE Trans. Image Processing, vol.
World, vol. 18, no. 8, pp. 28–38, Aug. 1995. 7, no. 1, pp. 129–133, Jan. 1998.
[65] J. D. Foley, A. van Dam, and J. F. Hughes, Computer Graphics: [89] D. Travis, Effective Color Displays: Theory and Practice. San
Principles and Practice, 2nd ed. Reading, MA: Addison- Diego, CA: Academic, 1991.
Wesley, 1993. [90] W. K. Pratt, “Spatial transform coding of color images,” IEEE
[66] M. Kulkarni and M. Grant, “Color management and film Trans. Commun., vol. COM–19, no. 12, pp. 980–992, Dec.
recorders,” Photographic Process., pp. 52–53, Jan. 1995. 1971.
[67] R. Vetter, C. Ward, and S. Shapiro, “Using color and text in [91] W. B. Pennebaker and J. L. Mitchell, JPEG Still Image Data
multimedia projections,” IEEE Multimedia Mag., vol. 2, pp. Compression Standard. New York: Van Nostrand, 1993.
46–54, Winter 1995. [92] J. L. Mitchell, C. Fogg, D. J. LeGall, and W. B. Pennebaker,
[68] R. W. G. Hunt, The Reproduction of Color in Photography, Eds., MPEG Digital Video Compression Standard. New York:
Printing, and Television, 4th ed. Tolworth, England: Fountain Chapman & Hall, 1997.
Press, 1987. [93] N. Ahmed and K. R. Rao, Orthogonal Transforms for Digital
[69] R. P. Khosla, “From photons to bits,” Phys. Today, vol. 45, no. Signal Processing. New York: Springer-Verlag, 1975.
12, pp. 42–49, Dec. 1992. [94] International Radio Consultative Committee, “Encoding param-
[70] K. A. Parulski, L. J. D’Luna, B. L. Benamati, and P. R. Shelley, eters of digital television for studios,” International Telecom-
“High-performance digital color video camera,” J. Electron. munications Union, Geneva, Switzerland, CCIR Rec. 601–2,
Imaging, vol. 1, no. 1, pp. 35–45, Jan. 1992. 1990.
[71] D. H. Brainard, “Bayesian method for reconstructing color [95] A. Netravali and B. Haskell, Eds., Digital Pictures, Repre-
images from trichromatic samples,” in IS&T’s 47th Ann. Conf., sentation, Compression, and Standards. New York: Plenum,
ICPS’94: Physics and Chemistry of Imaging Systems, May 1995.
15–20, 1994, vol. 2, pp. 375–380. [96] M. Anderson, R. Motta, S. Chandrasekar, and M. Stokes,
[72] D. R. Cok, “Reconstruction of CCD images using template “Proposal for a standard default color space for the Inter-
matching,” in IS&T’s 47th Ann. Conf., ICPS’94: Physics and net—sRGB,” in Proc. IS&T/SID 4th Color Imaging Conf.:
Chemistry of Imaging Systems, May 15–20, 1994, vol. 2, pp. Color Science, Systems and Applications, Scottsdale, AZ, Nov.
380–385. 19–22, 1996, pp. 238–246.
[73] K. B. Benson and J. C. Whitaker (Eds.), Television Engineering [97] J. C. Zwinkels, “Color-measuring instruments and their calibra-
Handbook: Featuring HDTV Systems. New York: McGraw- tion,” Displays, vol. 16, no. 4, pp. 163–171, 1996.
Hill, 1992. [98] H. R. Kang, Color Technology for Electronic Imaging Devices.
[74] A. J. Patti, M. I. Sezan, and A. M. Tekalp, “Robust methods Bellingham, WA: SPIE, 1997.
for high-quality stills from interlaced video in the presence of [99] J. N. Cook, P. A. Sample, and R. N. Weinreb, “Solution to
dominant motion,” IEEE Trans. Circuits Syst. Video Technol., spatial inhomogenity on video monitors,” Color Res. Appl., vol.
vol. 7, pp. 328–342, Apr. 1997.
18, no. 5, pp. 334–340, Oct. 1993.
[75] D. H. Brainard. (1997). Hyper-spectral image data. [Online].
[100] International Color Consortium. (Nov. 20, 1995). International
Available WWW: http://color.psych.ucsb.edu/hyperspectral/.
Color Consortium profile format, ver. 3.2. [Online]. Available
[76] D. H. Brainard, W. A. Brunt, and J. M. Speigle, “Color
ftp: ftp.fogra.org.
constancy in the nearly natural image, 1. Asymmetric matches,”
J. Opt. Soc. Amer. A, vol. 14, no. 9, pp. 2091–2110, Sept. 1997. [101] M. C. Stone, W. B. Cowan, and J. C. Beatty, “Color gamut
mapping and the printing of digital color images,” ACM Trans.
[77] P. L. Vora, M. L. Harville, J. E. Farrell, J. D. Tietz,
and D. H. Brainard, “Image capture: Synthesis of sensor Graph., vol. 7, no. 4, pp. 249–292, Oct. 1988.
responses from multispectral images,” in Proc. SPIE Color [102] J. Meyer and B. Barth, “Color gamut mapping for hard copy,”
Imaging: Device Independent Color, Color Hard Copy, and in SID Dig., 1989, pp. 86–89.
Graphic Arts II, 1997, vol. 3018, pp. 2–11. (Available URL: [103] R. S. Gentile, J. P. Allebach, and E. Walowit, “A comparison of
http://color.psych.ucsb.edu/hyperspectral/.) techniques for color gamut mismatch compensation,” in Proc.
[78] R. Luther, “Aus Dem Gebiet der Farbreizmetrik,” Z. Tech. SPIE Human Vision, Visual Processing, and Digital Display, B.
Phys., vol. 8, pp. 540–558, 1927. E. Rogowitz, Ed., 1989, vol. 1077, pp. 342–354.
[79] H. E. Ives, “The transformation of color-mixture equations from [104] E. D. Montag and M. D. Fairchild, “Simulated color gamut
one system to another,” J. Franklin Inst., vol. 16, pp. 673–701, mapping using simple rendered images,” in J. Bares, Ed., Proc.
1915. SPIE Color Imaging Device-Independent Color, Color Hard
[80] P. C. Hung, “Colorimetric calibration for scanners and media,” Copy, and Graphic Arts, 1996, vol. 2658, pp. 316–325.
in Proc. SPIE, 1991, vol. 1448, pp. 164–174. [105] , “Evaluation of gamut mapping techniques using simple
[81] J. J. Gordon and R. A. Holub, “On the use of linear transfor- rendered images and artificial gamut boundaries,” IEEE Trans.
mations for scanner calibration,” Color Res. Appl., vol. 18, no. Image Processing, vol. 6, pp. 977–989, July 1997.
3, pp. 218–219, June 1993. [106] C. Connolly and T. Fliess, “A study of efficiency and accuracy
[82] G. Sharma and H. J. Trussell, “Color scanner performance in the transformation from RGB to CIELAB color space,” IEEE
trade-offs,” in J. Bares, Ed., Proc. SPIE Color Imaging Device- Trans. Image Processing, vol. 6, pp. 1046–1048, July 1997.
Independent Color, Color Hard Copy, and Graphic Arts, 1996, [107] M. D. Fairchild, “Considering the surround in device inde-
vol. 2658, pp. 270–278. pendent color imaging,” Color Res. Appl., vol. 20, no. 6, pp.
[83] H. E. J. Neugebauer, “Quality factor for filters whose spectral 352–363, Dec. 1995.
transmittances are different from color mixture curves, and its [108] K. M. Braun and M. D. Fairchild, “Testing five color-
application to color photography,” J. Opt. Soc. Amer., vol. 46, appearance models for changes in viewing conditions,” Color
no. 10, pp. 821–824, Oct. 1956. Res. Appl., vol. 22, no. 3, pp. 165–173, June 1997.
[84] P. L. Vora and H. J. Trussell, “Mathematical methods for the [109] International Commission on Illumination, “The CIE 1997 in-
design of color scanning filters,” IEEE Trans. Image Processing, terim color appearance model (simple version), CIECAM97s,”
vol. 6, pp. 312–320, Feb. 1997. Rep. CIE Tech. Committee TC1-34, Aug. 1997.
[85] G. Sharma and H. J. Trussell, “Figures of merit for color [110] M. J. Vrhel and H. J. Trussell, “Color correction using principal
scanners,” IEEE Trans. Image Processing, vol. 6, no. 7, pp. components,” Color Res. Appl., vol. 17, no. 5, pp. 328–338,
990–1001, July 1997. Oct. 1992.
[86] M. J. Vrhel and H. J. Trussell, “Optimal color filters in the [111] D. L. MacAdam, Ed., Sources of Color Science. Cambridge,
presence of noise,” IEEE Trans. Image Processing, vol. 4, no. MA: MIT Press, 1970.
6, pp. 814–823, June 1995. [112] J. Bares, Ed., Proc. SPIE Color Imaging: Device-Independent
[87] M. Wolski, J. P. Allebach, C. A. Bouman, and E. Walowit, Color, Color Hard Copy, and Graphic Arts, 1996, vol. 2658.

SHARMA et al.: COLOR IMAGING FOR MULTIMEDIA 1107

 Gaurav Sharma (Member, IEEE) was born H. Joel Trussell (Fellow, IEEE) received the
in Dehradun, India, on October 12, 1968. He B.S. degree from the Georgia Institute of Tech-
received the B.E. degree in electronics and com- nology, Atlanta, in 1967, the M.S. degree from
munication engineering from the University of Florida State University, Tallahassee, in 1968,
Roorkee, India, in 1990, the M.E. degree in and the Ph.D. degree from the University of New
electrical communication engineering from the Mexico, Albuquerque, in 1976.
Indian Institute of Science, Bangalore, in 1992, He joined the Los Alamos Scientific Labo-
and the M.S. degree in applied mathematics ratory, Los Alamos, NM, in 1969, where he
and the Ph.D. degree in electrical engineering began working in image and signal processing
from North Carolina State University (NCSU), in 1971. During 1978–1979, he was a Visiting
Raleigh, in 1995 and 1996, respectively. Professor at Heriot-Watt University, Edinburgh,
During 1992–1996, he was a Research Assistant with the Center for Scotland, where he worked with both the university and industry on image-
Advanced Computing and Communications in the Electrical and Computer processing problems. In 1980, he joined the Electrical and Computer
Engineering Department at NCSU. Since 1996, he has been a Member Engineering Department at North Carolina State University, Raleigh.
of Research and Technical Staff at Xerox Corporation’s Digital Imaging During 1988–1989, he was a Visiting Scientist with the Eastman Kodak
Technology Center in Webster, NY. His current research interests include Company, Rochester, NY. During 1997–1998, he was a Visiting Scientist
color science and imaging, signal restoration, image halftoning, and error- with Color Savvy Systems, Ltd., Springboro, OH. His research has been
correction coding. in signal and image restoration. Most recently, he has concentrated on
Dr. Sharma is an associate member of Sigma Xi and a member of the accurate measurement and reproduction of color.
Society for Imaging Science and Technology, the International Society for Dr. Trussell is a past Associate Editor of IEEE TRANSACTIONS ON
Optical Engineering, the IEEE Signal Processing Society, and Phi Kappa ACOUSTICS, SPEECH, AND SIGNAL PROCESSING and is currently an Associate
Phi. Editor for IEEE SIGNAL PROCESSING LETTERS. He was a member and past
Chairman of the Image and Multidimensional Digital Signal Processing
Committee of the IEEE Signal Processing Society. He was a member
of the Board of Governors of the Signal Processing Society. He was a
corecipient of the IEEE Acoustics, Speech, and Signal Processing Society
Michael J. Vrhel (Member, IEEE) was born in
Senior Paper Award in 1986 (with M. R. Civanlar) and the IEEE Signal
St. Joseph, MI, in 1964. He received the B.S.
Processing Society Paper Award in 1993 (with P. L. Combettes).
degree in electrical engineering from Michigan
Technological University, Houghton, in 1987
and the M.S. and Ph.D. degrees in electrical en-
gineering from North Carolina State University,
Raleigh, in 1989 and 1993, respectively.
From 1993 to 1996, he was a National Re-
search Council Associate at the National Insti-
tutes of Health, Biomedical Engineering and In-
strumentation Program. Currently, he is the Se-
nior Scientist at Color Savvy Systems, Ltd., Springboro, OH. His research
interests include color reproduction, signal restoration/reconstruction, and
wavelets.
Dr. Vrhel received a Kodak Fellowship from Eastman Kodak Company,
Rochester, NY, from 1989 to 1993.

1108 PROCEEDINGS OF THE IEEE, VOL. 86, NO. 6, JUNE 1998