Page 1

305

JOURNAL OF IMAGING SCIENCE AND TECHNOLOGY® • Volume 45, Number 4, July/August 2001

storage, retrieval, and display (both in soft and printed

forms) with an eye on cost, time, and who will do the

work is extremely challenging.1–3 Because the scope of

this publication is limited to direct digital capture of

paintings and will not address digitizing photographic

collections or photographic reproductions of paintings,

conservation issues are especially important. The pro-

cess of digitizing must not damage the painting, which

could occur all too easily by excessive handling and ir-

radiation by high-intensity light sources. Blackwell4 and

Lossau and Liebetruth5 have considered these issues in

practice. Clearly, it is desirable to digitize the art once

and in a manner that facilitates a variety of derivatives

created for a variety of applications, be it web-based,

printed publication, scientific analysis, or any number

of scholarly endeavors by art historians.

Thus the purpose of this publication is to provide suf-

ficient background to aid in setting specifications for

color image-capture systems. Having presented this

background, it is possible to propose a methodology for

digital image capture that minimizes the inherent limi-

tations of many digital systems in use today. Finally,

considerations will be given to the future. What are the

optimal characteristics of a digital imaging system de-

signed to record paintings?

A Review of Digital Imaging

The Human-Visual System

The first step in understanding digital imaging is to

understand about the human visual system. Four con-

cepts will be defined: spectral sensitivity, opponent en-

coding, spatial resolution, and nonlinear response. This

description is very simplified. For greater detail, but

still at a qualitative level, see Berns.6 For a thorough

understanding of the human visual system, see Wandell7

and Kaiser and Boynton.8

Incident light interacts with visual receptors, rods and

cones. Following a chemical reaction, light energy is con-

Introduction

Digital image databases have become ubiquitous with

museums, galleries, archives, and libraries. (For simplic-

ity, “museum” will be used to represent the many reposi-

tories.) Connect to their websites and a wealth of visual

information is available. Museums are sharing their im-

ages to increase access such as the Art Museum Image

Consortium, AMICO, a subscription program for educa-

tional institutions (see www.amico.org). The quality of

the digital images can be quite varied. Leaving aside the

issue of the long-term storage of digital information,

clearly a critical issue but beyond the scope of this publi-

cation, there are many reasons that quality might be poor.

Some are philosophical in origin where there is a delib-

erate decision not to create a high-quality accurate rep-

resentation, rather, a database of low-quality “thumbnail”

images. Some reasons are technical in origin; and some

are caused by constraints imposed by the sheer magni-

tude of creating and maintaining a digital archive of many

thousands, and in some cases, hundreds of thousands of

images. Finally, images available to the public through

the internet are usually of low resolution as well as de-

signed for CRT display. These are “derivative” images

created from the “master” image. A digital master is a

result of direct digital capture of the work of art.

The technical problems are manifest almost immedi-

ately. Trying to define specifications for image capture,

The Science of Digitizing Paintings for Color-Accurate Image Archives:

A Review

Roy S. Bernsv*

National Gallery of Art, Washington, DC

A review of the human visual system, the CIE L*, a*, b* color space and its use in evaluating color image quality, and digital

image capture is presented, the goal of which is to provide background information for imaging professionals involved in creating

digital image databases for museums, galleries, archives, and libraries. Following this review, an analysis was performed to

determine the effects of bit depth, dynamic range, gamma correction, and color correction on the ability to estimate colorimetric

data from R, G, B digital images with a minimum of error. The proper use of gray scale and color targets was also considered.

Recommendations are presented for the direct digital image capture of paintings. Finally, a brief look into the future using

spectral imaging techniques is presented.

Journal of Imaging Science and Technology 45: 305–325 (2001)

Original manuscript received June 5, 2000

v IS&T Member

* Permanent address: Munsell Color Science Laboratory, Chester F.

Carlson Center for Imaging Science, Rochester Institute of Technology,

Rochester, New York; berns@cis.rit.edu

Supplemental Materials—Web Figures 1 through 3 can be found in

color on the IS&T website (www.imaging.org) for a period of no less

than 2 years from the date of publication.

©2001, IS&T—The Society for Imaging Science and Technology

Feature Article

306 Journal of Imaging Science and Technology®

Berns

verted to a neural signal. There are three classes of cones

(our color receptors), L, M, and S, named for their domi-

nant sensitivity to long, medium, and short wavelengths

of light. Each has a unique spectral sensitivity, shown

in Fig. 1. Spectral sensitivity defines a detector’s sensi-

tivity as a function of wavelength. Because there are

only three types of cones and their spectral sensitivi-

ties are broad, many different objects can produce the

same cone responses, leading to the identical color. This

is known as metamerism (pronounced “me-tam’-er-ism”)

and explains why color reproduction is possible in which

small numbers of colorants can reproduce our world,

composed of thousands of colorants. A metameric match

between foliage, illuminated by natural daylight, and a

CRT color reproduction of the foliage is shown in Fig. 2.

Despite large differences in their spectral properties,

these two stimuli match in color.

The cones combine spatially and form opponent sig-

nals, white/black, red/green, and yellow/blue. During the

late 1800’s, Hering considered these six colors as elemen-

tal, shown in Fig. 3. Lines connecting the various el-

emental colors indicate possible continuous perceptions.

There are not lines connecting yellow and blue and red

and green. It is not possible to have a color that is si-

multaneously reddish and greenish, for example.

Because there are different numbers of each cone type,

when they combine to form opponent signals, there are

different spatial resolutions among the opponent signals.

Spatial resolution defines the resolving power of an im-

aging system and leads to the ability to discern fine de-

tail. The black/white signal has the highest spatial

resolution, followed by the red/green signal. The yellow/

blue signal has quite low spatial resolution. These dif-

ferences in spatial resolution have been exploited in im-

age compression such as JPEG.10 The concept of “visually

lossless compression” originates with this property of the

eye. Spatial resolution is reduced in the chromatic chan-

nels in a manner that results in images that look identi-

cal to their uncompressed counterparts when viewed at

typical distances, hence the term “visually lossless.” How-

ever, information is still being discarded.

The neural processing from cone-receptor signals

through signals interpreted by the brain and result-

ing in color names such as yellow, brown, and gray is

exceedingly complex. Every year, vision scientists fill

in more pieces of the puzzle. From a practical perspec-

tive, it is very useful to have a simple mathematical

model that enables color perceptions to be estimated

from light imaged onto our visual system. A model was

derived by the International Commission on Illumina-

tion (CIE) in 1976, known as CIE L*, a*, b* (pronounced

“el-star” and so on) or its official abbreviation, CIELAB

(pronounced “see-lab”).11 The coordinates, L*, a*, and

b* represent the perceptions of lightness, redness/

greenness, and yellowness/blueness, respectively.

Figure 1. The spectral sensitivities of the human-visual

system’s cone9 (normalized to equal area).

Figure 2. Spectral properties of a metameric pair formed by

foliage, illuminated by natural daylight (dashed line), and a

CRT display (solid line).6

Figure 3. Six elemental colors postulated by Hering along with

lines indicating possible perceptions, e.g., yellowish red (W =

white; Y = yellow; R = red; K = black; B = blue; G = green) .

The Science of Digitizing Paintings for Color-Accurate Image Archives: A Review

Vol. 45, No. 4, July/August 2001 307

CIELAB is considered a color space in which positions

indicate an object’s color. This type of color space is

diagrammed in Fig. 4. The coordinates are calculated

from the knowledge of an object’s spectral reflectance

or transmittance properties, a light source interacting

with the object, and the spectral sensitivities of an

observer. For a number of reasons, the L, M, and S

spectral sensitivities, shown in Fig. 1, are not used;

rather, a set of curves mathematically related to spec-

tral sensitivities, known as color-matching functions,6,11

are used instead.

The complex signal processing of the visual system

results in a nonlinear (curved) relationship between

light imaged onto the eye and color perceptions, shown

in Fig. 5 for lightness. This means that for dark colors,

small changes in an object’s reflectance or transmittance

lead to large changes in lightness. For light colors, the

opposite occurs: Large changes in an object’s reflectance

or transmittance lead to small changes in lightness. This

curvature is known as a compressive function. This com-

pression also occurs for the chromatic channels.

In summary, the human visual system has three cone

types, each with unique spectral sensitivities that over-

lap greatly. They combine spatially forming opponent

signals: white/black, red/green, and yellow/blue. Each

channel has a different spatial resolution. The relation-

ship between incident light and color perceptions is non-

linear. A simple model, CIELAB, can be used to calculate

color perceptions from measurements of the object, its

illumination, and the observer.

The Digital Camera

Digital cameras consist of three main components: in-

put optics, sensor, and signal processing. The optical

system is typical of conventional photography in which

a scene is imaged on to a two-dimensional plane, much

like the human visual system.12 Light interacts with a

sensor, often a charged-couple device (CCD) that con-

verts light energy to an electrical signal.13 The sensor

may be one-dimensional—a single row of sensors—or

two-dimensional—a grid of sensors. In one-dimensional

systems, the row of sensors is scanned across the image

plane. These are “scanbacks” and are similar to flat-

bed scanners. Because scanning occurs, the objects be-

ing digitized are stationary. Through signal processing,

the electrical signal is converted to a digital signal.

For color, the one-dimensional detector array is

trebled, each array having either a red, green, or blue

filter in front of the detector, often referred to as “tri-

linear arrays.” In some systems, a color filter wheel is

positioned between the optical system and a mono-

chrome, i.e., grayscale, scanning array. The array scans

across the image plane multiple times, once for each

filter-wheel position. For two-dimensional arrays, de-

tectors are filtered in a mosaic pattern. Similar to the

human visual system, there is not the same number of

red, green, and blue sensors. Image processing is used

to interpolate missing data for each color plane, some-

times called “demosaicing.” The design of the mosaic and

accompanying image processing is optimized to mini-

mize artifacts caused by sampling, known as “aliasing.”

It is also possible to have three two-dimensional mono-

chrome arrays, each filtered with either a red, green, or

blue filter or a single two-dimensional array and filter

wheel. These systems do not require the complex inter-

polation procedures.

For color accuracy, the most important characteristic

of the digital camera is its spectral sensitivities. Ide-

ally, they should closely resemble the human visual

system’s spectral sensitivities shown in Fig. 1. Strictly,

a camera’s spectral sensitivities should be a linear trans-

formation of the human visual system’s spectral sensi-

tivities.14 That is, through a linear transformation, the

L, M, and S sensitivities are well estimated. Thus, CIE

color-matching functions, which do not resemble L, M,

and S sensitivities, meet this criterion. This is also

known as the Luther or Luther-Ives condition. For many

scanbacks, this critical characteristic is seldom achieved,

shown in Fig. 6. It is seen that these sensitivities have

very little overlap and the position of the red sensitiv-

ity is shifted considerably to longer wavelengths. It is

Figure 4. Conceptualized CIELAB color space.6

Figure 5. Nonlinear response between incident light and light-

ness. The incident light can have units of radiance, irradiance,

luminance, illuminance, or simply power per unit area. (CIE

L* is used as a representation of lightness.)

308 Journal of Imaging Science and Technology®

Berns

important to note that these sensors were designed for

flat-bed scanners. These devices are optimized for im-

aging photographic materials. These sensitivities are

“densitometric,” rather than “colorimetric.” They are

designed to record the amounts, i.e., densities, of cyan,

magenta, and yellow dyes in color photographic materi-

als.15 Also shown in Fig. 6 is the spectral density, calcu-

lated by Dλらむだ = -Log10(Rλらむだ,gray/Rλらむだ,paper), of color-paper

photographic dyes reproducing a medium gray color. The

blue and green peak sensitivities of the scanback are

approximately coincident with the yellow and magenta

dye peaks. The red sensitivity is not coincident to the

cyan dye, but shifted to shorter wavelengths. This is a

result of the particular IR cut-off filter used in this cam-

era system, selected to improve its colorimetric perfor-

mance when used as a scanback. When the sensor is

used in a scanner, a different IR cut-off filter would be

used. If a photographic collection were digitized using

this type of scanback, it is possible to transform the R,

G, and B data to highly accurate L*, a*, b* estimates,

though the computations are complex and for some pho-

tographic materials, the spectral sensitivities would

need to be changed.16–18 When digitizing paintings, den-

sitometric spectral sensitivities will result in large er-

rors, even with the addition of color management,

demonstrated in a latter section. Unfortunately, these

types of digital cameras are the most common types used

to create digital archives.4

Figure 7 is a plot of the spectral sensitivities of typi-

cal digital cameras employing color CCD two-dimen-

sional arrays. There is considerably more overlap than

densitometric scanbacks. However, the red sensitivity

is still shifted towards longer wavelengths. It must be

noted that color accuracy is only one of a number of cri-

teria when designing a digital camera. Low-light sensi-

tivity, image noise, resolution, read-out speed,

manufacturing costs, and so on all must be considered.

Many of these design criteria are mutually exclusive;

the final design is always a compromise.

Figure 8 is a plot of the spectral sensitivities of a mono-

chrome scanner with a filter wheel. The key design cri-

terion was color accuracy. The spectral sensitivities of

this camera system are much closer to the human vi-

Figure 6. Spectral sensitivities of a typical scanback (solid

lines), normalized to equal area. These sensitivities include

the CCD spectral sensitivity, filter transmittances, and infra-

red radiation blocking filter. The human visual system’s spec-

tral sensitivities, shown in Fig. 1, are also plotted (dashed lines)

as is the spectral density of typical photographic dyes com-

bined to reproduce a medium-gray color (dashed-dotted line).

Figure 7. Spectral sensitivities of a typical color CCD two-

dimensional array (solid lines), normalized to equal area. These

sensitivities include the CCD spectral sensitivity, filter trans-

mittances, and infrared radiation blocking filter. The human

visual system’s spectral sensitivities, shown in Fig. 1, are also

plotted (dashed lines).

Figure 8. Spectral sensitivities of a monochrome scanback with

three optimized color filters (solid lines), normalized to equal

area. These sensitivities include the CCD spectral sensitivity,

filter transmittances, and infrared radiation blocking filter.

The human visual system’s spectral sensitivities, shown in Fig.

1, are also plotted (dashed lines).

The Science of Digitizing Paintings for Color-Accurate Image Archives: A Review

Vol. 45, No. 4, July/August 2001 309

sual system than the other sensitivities shown in Figs.

6 and 7. The red sensitivity has its peak response coin-

cident with the eye’s L sensitivity.

Finally, it is worthwhile showing the spectral sensitiv-

ity of a typical color transparency film, shown in Fig. 9.

Because of the inherent design of film in which sensitivi-

ties are stacked, limited choices for photosensitive mate-

rials, and a system that must simultaneously consider

image capture and display, the spectral sensitivities do

not closely resemble the human visual system. Further-

more, color-photographic materials are designed to make

pleasing reproductions, not necessarily accurate repro-

ductions. For all of these reasons, photographic repro-

ductions are not color accurate although with experience,

it is possible through filtration using color-compensat-

ing filters to achieve global color matches. This is done

routinely in museum photographic departments.

CCD detectors have a linear response to incident light,

shown in Fig. 10. This is a dramatic difference to the

nonlinearity shown in Fig. 5. Depending on how the

“raw” detector signal is processed and the digital prop-

erties of the stored image, this difference can be incon-

sequential or result in large visual artifacts. This will

be considered in detail later in this article.

The raw detector signals, in the form of an analog sig-

nal, i.e., millivolts, are amplified and digitized using an

analog-to-digital converter (ADC). The range of digital

values depends on the number of bits in the ADC. Most

commonly, 8, 12, or 16 bit ADC’s are used. This means

that there are 28 (256), 212 (4096) or 216 (65,536) number

of gray levels for each color channel (also called “color

levels” as well as “bit depth”). The greater the number

of levels, the less the amount of error caused by the con-

version from analog to digital signals. Visual artifacts

caused by an insufficient number of gray levels are de-

scribed and shown below in the next section.

Thus the camera has an inherent spatial and color

resolution. The spatial resolution is determined in large

part by the number of detectors. For example, high-reso-

lution scanning systems commonly have one-dimen-

sional arrays with 6000 or 8000 detector elements that

scan in 8000 or more steps. Thus the camera would have

spatial resolution of 6000 × 8000 or 8000 × 8000 pixels

(“picture elements”). This does not mean that a digital

camera with more pixels is always better than one with

less. Having many pixels combined with poor optics may

have poorer image quality than fewer pixels with excel-

lent optics. Ultimately, spatial resolution is determined

for the entire camera system including the optics, sen-

sor, and image processing, e.g., spatial interpolation, by

imaging targets designed to quantify resolution and

mathematically analyzing the digital images of these

targets.2,13,19 Color resolution is determined by the num-

ber of bits in the ADC. For high-quality digital cameras,

the ADC’s have a minimum of 12 bits.

The Digital Image

A digital image is simply a two-dimensional array of

numbers in which each number relates loosely to the

amount of light reflected by an object. The image reso-

lution defines the dimensions of the array. The concept

of relating numbers to light reflection is easier to un-

derstand when a black and white, i.e., “monochrome” or

“grayscale”, image is considered. Essentially, dark ar-

eas have small numbers while light areas have large

numbers. The most common bit depth for images is 8

bits per channel. Because digital values begin at 0, the

numbers range from 0 to 255 (28 – 1). A white is near

255 while a black is near 0.

There is a tendency to treat digital values as if they

were perceptions: a perfect white is 255, a perfect black

is 0, and a medium gray, e.g., Kodak Gray Card, is 128.

For color, all three channels have the identical digital

values. Although it is always true that a larger number

corresponds to a lighter gray, the relationship between

digital values and perception is rarely one to one. Re-

Figure 9. Spectral sensitivities of typical transparency film

(solid lines), in this case Eastman Kodak Ektachrome 64T.15

The human visual system’s spectral sensitivities, shown in Fig.

1, are also plotted (dashed lines).

Figure 10. Relationship between incident light and raw de-

tector signals for CCD arrays.

310 Journal of Imaging Science and Technology®

Berns

lating digital values to perception requires system char-

acterization using standard test targets.

This difficulty in relating numbers to perceptions be-

comes further exacerbated when an image is output and

viewed. Simply taking a digital image and displaying it

on different computer platforms and printing it out on

different printers reveals the large range of renditions

that result from the identical image file. Again, this

problem is minimized through system characterization

using standard test targets and measurement equip-

ment. The origins of this problem begin with the inher-

ent nonlinearity of many imaging devices. Conventional

photography, broadcast television, computer-controlled

CRT and LCD displays, and printing all have a nonlin-

ear relationship between their input, e.g., exposure,

voltage, digital counts, dot area, and their output, mea-

sured as reflectance, transmittance, or light energy.

A specific nonlinearity worth exploring is the

nonlinearity of CRT displays, shown in Fig. 11. By coin-

cidence, this nonlinearity is similar to the human vi-

sual system, except inverted. This is an expansive

function that is inherent to all vacuum tubes. During

the early 20th century, a simple mathematical equation

was derived to relate input and output, shown as Eq.

1.20,21

The normalized output was predicted by

exponentiating the normalized input. The specific ex-

ponent is notated by the Greek symbol gamma, γがんま. The

use of γがんま to notate an exponent stems from photographic

science.22 However, a film’s gamma has a different physi-

cal meaning than a display’s gamma. Digital systems

use gamma as shown in Eq. 1. This is a simplification.

For very small digital counts, the exponent is replaced

with a slope term, for example, sRGB.23 Furthermore,

this equation will only accurately model computer-con-

trolled CRT displays if the monitor has its black level

set perfectly: at 0 digital code value, there is not any

emitted light while at 1 digital code value, light is emit-

ted. Because this display set up is rarely achieved, more

complex equations have been derived.24,25

output

maximum output

input

maximum input











 =













γがんま

(1)

Because CRT displays have inherent nonlinearity, raw

detector signals from digital cameras will lead to poor

image quality unless the image is first “gamma cor-

rected,” that is, the inherent nonlinearity is appropri-

ately compensated. The raw signals are exponentiated

by the inverse of the display’s gamma, shown in Eq. 2.

output maximum output

raw input

maximum raw input

= (

)













1γがんま

(2)

The net effect of gamma correcting the raw image and

the display’s inherent nonlinearity is an image that ap-

pears correct. That is, the tonal properties appear rea-

sonable. Complete gamma correction is also a

simplification. Quite often, gamma correction is not fully

applied in order to compensate for dim (e.g., broadcast

television) or dark, e.g., projected slides, viewing condi-

tions reducing the perceived contrast in images. See

Hunt26 for greater details. This is shown in Fig. 12. Im-

posing an exponential function to a digital image is a

common and useful procedure.

The vast majority of color images are 24-bit images, 8

bits for each of the red, green, and blue channels. Yet,

for high-quality digital cameras, their analog-to-digital

converters are 12 or 16 bits per channel. In order to have

an 8-bit per channel image, the 4096 or 65,536 poten-

tial levels from the camera must be transformed to 256

levels. This is a reduction of information, and in simi-

lar fashion to spatial compression, this should be done

in a manner that minimizes visual artifacts. The most

common artifacts are banding in which smooth grada-

tions become banded, or blocking in which details are

lost, shown in Fig. 13. These are known as “quantiza-

tion errors.” When a continuous signal, i.e., analog sig-

nal, becomes discrete, i.e., digital, it is quantized into a

number of specific levels. The specific number of levels

is determined by the bit depth. For all imaging applica-

tions, 216 number of levels is a sufficient number of lev-

els such that quantization errors are essentially

eliminated. However, when converting to 8 bits per chan-

nel, quantization errors can be quite noticeable depend-

ing on the method of bit reduction. A numerical analysis

is considered in a later section, Optimal Encoding—

Lightness.

A useful method to analyze quantization is by evalu-

ating image histograms. The histograms for the Fig. 13

images are shown in Fig. 14. The height of a peak rep-

resents the number of pixels with a particular digital

value. The left-hand image has a histogram in which

each level between the minimum and maximum digital

counts has a number of pixels. Conversely, the right-

hand image is missing data, resulting in only a few lev-

els. The visual banding is quantified via an image

histogram.

Assessing Color Image Quality

Obviously, any digital-capture system is capable of

transforming an object into a digital image. The impor-

tant question is whether the digital image has archival

value as a digital representation. Can the digital val-

ues be used to estimate the color of the object? Even

though any “picture tells a thousand words,” digital

image-archives should do more; they should facilitate

Figure 11. CRT nonlinearity. This nonlinearity is inherent to

all vacuum tubes.20,21

The Science of Digitizing Paintings for Color-Accurate Image Archives: A Review

Vol. 45, No. 4, July/August 2001 311

Figure 12. Photograph and gray scale imaged linearly and displayed with (left) and without (right) “gamma correction.” (Image

courtesy of C. McCabe).

Figure 13. The effects of quantization on image quality is shown in the right-hand figure. Notice that smooth gradations become

banded and that fine detail in the woman’s hair are lost.

312 Journal of Imaging Science and Technology®

Berns

scientific documentation and study of works of art. The

goal is to have the capability to estimate quantitative

information about the object, not create a visually

equivalent representation for a defined output device

such as a display or printed document. Imaging test

targets along with the work of art and comparing stan-

dard values of the test target with its estimates is most

easily accomplished for analyzing the ability to estimate

the color of cultural heritage. The standard values of

interest are those that predict what observers see.

CIELAB provides these required values.

Two types of test targets should be used. The first is a

gray scale, such as the Kodak Gray Scale Q-13 or ISO

14524 chart.27 Gray-scale targets characterize the rela-

tionship between digital values and lightness (by plot-

ting the target’s average digital counts and measured

values), dynamic range (by evaluating whether each step

has a unique average digital count), and whether the

image is gray-balanced (by evaluating whether each

channel has matched digital counts). Although it is com-

mon practice to include the Kodak Separation Guide

along with a gray scale, this target cannot be used to

effectively evaluate color errors or for color management.

It was designed to validate the correct correspondence

between four-color film separations and screens. This

practice is unnecessary in modern graphic reproduction.

Ideally, gray scale and color targets should span the

lightness and color range of imaged objects and have

similar spectral, i.e., made using similar colorants, and

surface, i.e., gloss and texture, properties. For modern

photographic and printed materials, IT8 standard tar-

gets are available.28,29 For paintings, targets are largely

nonexistent. A target that is often used in broadcast

television in order to evaluate the color accuracy of

broadcast television cameras is the GretagMacbeth

ColorChecker Color Rendition Chart (usually called the

ColorChecker chart),30 shown in Web Fig. 1 and Ref. 6.

It is produced using painted papers. Because it has a

number of pigments, it is a convenient target to evalu-

ate the accuracy of digitized paintings. Both targets can

be cut up to make smaller targets so that resolution is

not significantly reduced when including these targets

along with the object undergoing digitization. The

GretagMacbeth ColorChecker Color Rendition Chart is

also available as a smaller target. A more comprehen-

sive target of greater than 100 colored patches has also

been recently released, the ColorChecker DC.

Both targets are first measured using a spectropho-

tometer, usually having bidirectional geometry, a geom-

etry that minimizes specular reflection from con-

tributing to the measured values.1 From the spectral

measurements, CIELAB values are calculated for a stan-

dard illuminant and observer, the specific standards

dependent on the intent of the archive. The choice of

illuminant and observer is complex. Should the object’s

color be defined by its appearance in an exhibition gal-

lery, its appearance during conservation, or its appear-

ance as envisioned by the artist? Exhibition galleries

can have a range of illumination from bluish natural

daylight (7500 K) to yellowish incandescent (2200 K).

7500 K and 2200 K refers to correlated color tempera-

ture, the temperature of a blackbody radiator which

generates a white light matching the color of the source

of interest. Sources and displays are often defined in

terms of their correlated color temperatures. Conserva-

tors tend to use a combination of natural and artificial

daylight, probably averaging to 6500 K. Depending on

the artist, there can be one or more illumination choices,

for example a plein air versus studio painter. Quite of-

ten, the image archive is defined by how the archive

will be used rather than the above considerations. If

the end product is a printed publication, CIE illuminant

D50 and the 1931 standard observer are used.31 If the

end product is web based, CIE illuminant D65 and the

1931 standard observer are used.23 These output-ori-

ented images should be considered derivative images,

created using principles of color management.6,15,32 For

the analyses in this article, CIE illuminant D65 and the

1964 standard observer were used because this corre-

sponds to viewing objects in a natural-daylight lit stu-

dio with north-facing windows or typical conservation

laboratories. In the author’s experience, this combina-

tion leads to the best correlation between numerical and

visual color quality because CIELAB is most visually

uniform for D65 and metrics such as CIE94 are opti-

mized for visual data subtending a 10° field of view, rep-

resented by the 1964 standard observer. Also, for

metameric matching, the 1964 standard observer bet-

ter correlates with visual observations. This combina-

tion of illuminant and observer is in contradiction with

a number of imaging standards (e.g., Refs. 23, 31, and

32) that require the use of the 1931 standard observer

and often illuminant D50.

Assuming that the digital-camera system is gray-bal-

anced, the gray scale is used to ascertain whether the

image is linear or nonlinear with respect to light input.

Because CIELAB L* is nonlinearly related to incident

light as shown in Fig. 5, luminance factor, instead is used.

Luminance factor, also known as CIE tristimulus value

Y, is linearly related to incident light and is a standard

of light measurement.11 It is also used in the calculation

of L*. Spectrophotometers designed for color measure-

ment provide both CIE tristimulus values, X, Y, and Z,

and CIELAB coordinates. The average digital values for

each patch from the two images of the Kodak gray scale

shown in Fig. 12 are plotted against luminance factor in

Fig. 15. As expected, the raw digital counts are linearly

related to incident light. The gamma-corrected digital

values are nonlinearly related to incident light. This

nonlinearity can compensate for typical display

Figure 14. Histograms of the images shown in Fig. 13.

The Science of Digitizing Paintings for Color-Accurate Image Archives: A Review

Vol. 45, No. 4, July/August 2001 313

nonlinearities, such as that shown in Fig. 11. This re-

sults in images that have reasonable tone reproduction,

as demonstrated in Fig. 12. Because of various types of

imaging noise, the curves are not perfectly smooth.

A scanback-type digital camera with spectral sensitivi-

ties similar to those shown in Fig. 6 was used to digitize

the Kodak Gray Scale and ColorChecker chart. The raw

signals were transformed to L*, a*, and b* coordinates,

the general methodology described below, Optimal En-

coding – Color. Both numerical and graphical analyses

are performed in order to evaluate the camera’s accu-

racy in estimating colorimetric data measured using a

spectrophotometer. Graphically, “vector plots” are made,

shown in Fig. 16. The uppermost portion of Fig. 16 is an

a*b* projection. The end of the tail represents the mea-

sured coordinates while the head represents the esti-

mated coordinates. Vectors pointing either towards or

away from the origin (the center of the dashed lines at

a* = 0, b* = 0) indicate chroma error. Chroma is defined

as the degree of departure of a color from a gray of the

same lightness.6 Vectors pointing away from the origin

indicate the estimate is over-predicting chroma. Vectors

pointing towards the origin indicate the estimate is un-

der-predicting chroma. Vectors pointing in directions

other than towards or away from the origin indicate hue

errors. The length of the vector indicates the magnitude

of the chromatic error. Ideally, the errors should be suffi-

ciently small such that only the arrowhead is shown with-

out a tail. There are some systematic trends: yellows have

large chroma errors, blues have large hue errors, and

reds and greens have relatively smaller errors. The bot-

tom graph of Fig. 16 shows an L*C*ab projection where

C*ab represents chroma ( a

b

*

*

2

2

+

). Vectors that are point-

ing upwards indicate that the estimation is too light;

vectors pointing downward indicate that the estimation

is too dark. Vectors parallel with the C*ab axis indicate

accurate lightness estimation. There are two systematic

trends. First, many of the chromatic samples’ lightnesses

have been underestimated because the vectors are point-

ing downwards. Second, the lighter samples of the

ColorChecker’s gray scale have been over estimated (vec-

tors pointing upwards) and the darker neutrals have been

under estimated (vectors pointing downwards). Ordi-

narily, errors of this type for neutral samples indicate

calibration errors in which there is a mismatch between

actual and assumed photometric properties of the image

capture system, that is, a mismatch in gamma. In this

case, these estimation errors were caused by spatial non-

uniformities in illumination. Non-uniform illumination

of the Kodak gray scale was interpreted as a slight non-

linear photometric response. This was compensated for

during color management. Because the spatial non-uni-

formities varied across the image plane, this compensa-

tion resulted in systematic errors in different image

locations.

These differences are quantified numerically by cal-

culating differences in color positions. CIELAB is a rect-

angular color space with L*, a*, and b* axes (shown in

Fig. 4). Thus ∆L*, ∆a*, and ∆b* values are calculated

where the Greek symbol ∆ (“delta”) represents differ-

ence. ∆L* is calculated by subtracting the measured

value from the estimated value, i.e., ∆L* = L*estimated –

L*measured, and in similar fashion for the other coordi-

nates. A positive ∆ value indicates that the estimated

value has more of the particular quantity than the mea-

sured value. It is also useful to describe differences in

Figure 15. Luminance factor plotted against digital counts

for the Kodak gray scales from Fig. 12.

Figure 16. CIELAB vector plots in which the arrowhead de-

fines the coordinates of the estimated values while the end of

the arrow tail defines the coordinates of the measured values.

314 Journal of Imaging Science and Technology®

Berns

lightness, chroma (∆C*ab), and hue (∆H*ab), rather than

lightness, redness/greenness, and yellowness/blueness,

particularly for chromatic samples. The equations are

given in Refs. 6 and 11. All of these values have been

tabulated in Table I. With experience, the numerical

information can be quite useful. In many cases, the

graphical information is more intuitive.

An obvious question concerns the magnitude of error.

Are these errors large or small? Are they visible? Are

they objectionable? Simply stated, these errors are large

and objectionable, particularly since the goal is to esti-

mate colorimetry from an image archive. One problem

is the systematic nature of some of the errors shown in

Fig. 16. The second problem is the size of these errors.

When all three dimensions are considered simulta-

neously, the length of the vector can be used as an error

metric. In CIE terminology, this is called a total color

difference and is notated by ∆E*ab. However, CIELAB

as a color-difference space is poorly correlated with vi-

sual judgments of color difference.6 The last column in

Table I lists a weighted color difference, ∆E*94, also re-

ferred to as CIE94. This equation was designed for in-

dustries manufacturing colored products.33

The

weighting was optimized to improve correlation with

visual tolerances. A CIE94 color difference of unity is

slightly above a visual threshold. Thus, if pairs of col-

ored patches were prepared with these differences in

their CIELAB coordinates, the differences would all be

obvious, simulated in Web Fig. 2.

It is important to point out that CIE94 was designed

to correlate with comparisons of the quality of colored

patches, not colored images. As a rule of thumb, two

pictorial images viewed side by side with systematic

errors that result in an average ∆E*ab of 2.2 or less are

indistinguishable from one another.34,35 If the errors are

around 5∆E*ab on average, the accuracy is acceptable

for high-quality color reproduction. However, the 2.2 and

5∆E*ab rule of thumb applies to imaging systems where

errors from both image input and output occur. Because

this analysis is only considering image input, the er-

rors should be even smaller. In the author’s opinion, a

well-designed image input device should result in aver-

age errors of less than 2∆E*ab for the ColorChecker. Pres-

ently, there has been insufficient evaluation of CIE94’s

effectiveness for predicting the color quality of pictorial

images and accordingly, recommended figures of merit

cannot be given. Also, this equation will soon be updated

by the CIE.

As a final analysis, a color-difference histogram should

be evaluated, shown in Fig. 17. Quite often, differences

are not normally distributed and average and maximum

errors can be misleading.

Despite color management, these errors are large al-

though not unexpected. It is not surprising given the

large discrepancy between the camera and human spec-

tral sensitivities, plotted in Fig. 6. Although the cam-

era does a fine job in recording red, green, and blue light,

the human visual system does not “record” light in the

same manner. No matter how complex the color man-

agement system, there will always be estimation errors.

Optimal Encoding – Lightness

From the review of digital imaging, minimizing quanti-

zation errors is critical in order to maximize an archive’s

scientific value. This is achieved by the optimal encod-

ing of the lightness properties of the work of art. The

first step towards optimization is by review of the proper

digital capture practices for two-dimensional works of

art and documents, listed in Table II.

TABLE I. Colorimetric Errors Comparing Measured and Estimated Values for a GretagMacbeth ColorChecker

Color Rendition Chart Imaged with a Scanback-Type Digital Camera Followed by Color Management

Sample

∆L*

∆a*

∆b*

∆C*ab

∆H*ab

∆E*ab

∆E*94

Dark Skin

-6.0

-0.3

3.6

2.5

-2.6

7.0

6.5

Light Skin

-2.6

4.1

-0.4

2.6

3.2

4.9

3.8

Blue Sky

-4.2

0.5

-3.9

3.7

-1.3

5.8

4.7

Foliage

-6.8

-1.4

14.9

14.5

4.0

16.5

10.3

Blue Flower

-4.9

-1.0

-4.7

4.3

2.1

6.8

5.5

Bluish Green

-5.8

5.2

-4.6

-5.6

-4.2

9.1

6.9

Orange

-7.0

-0.8

4.5

3.4

-3.1

8.4

7.2

Purplish Blue

-4.1

-6.6

-7.5

6.9

7.1

10.7

6.5

Moderate Red

-3.8

4.8

-4.3

3.7

5.3

7.5

5.1

Purple

-6.4

-8.8

-0.3

-5.4

7.0

10.9

8.2

Yellow Green

-6.5

-2.5

18.0

18.0

2.7

19.3

8.2

Orange Yellow

-8.2

-7.1

12.6

10.2

-10.2

16.7

10.0

Blue

-8.1

-11.7

-12.5

10.4

13.6

18.9

11.6

Green

-6.5

6.4

7.0

1.6

9.3

11.5

8.5

Red

-8.7

6.0

-4.2

4.0

6.1

11.3

9.4

Yellow

-4.6

-4.8

28.5

28.2

-6.1

29.3

8.3

Magenta

-3.3

-1.1

-1.8

-0.3

2.1

3.9

3.5

Cyan

-6.3

12.1

-12.7

1.3

-17.5

18.6

12.9

White

6.1

-3.9

1.1

2.4

-3.2

7.3

7.1

Neutral 8

1.5

0.6

-1.7

0.2

1.8

2.3

2.3

Neutral 6.5

-3.3

0.5

-1.2

1.1

0.7

3.5

3.5

Neutral 5

-5.7

0.3

-0.5

-0.2

-0.5

5.8

5.8

Neutral 3.5

-6.6

0.7

0.4

-0.5

-0.6

6.7

6.7

Black

-8.8

0.6

1.1

-0.8

-0.9

8.8

8.8

Average

-5.0

-0.4

1.3

4.4

0.6

10.5

7.1

Maximum

29.3

12.9

The Science of Digitizing Paintings for Color-Accurate Image Archives: A Review

Vol. 45, No. 4, July/August 2001 315

Many of these practices are consistent with proper

conventional photographic practices. It is worth stress-

ing the importance of having sufficient illumination and

the maximum aperture to keep exposure (scan) times

short. With an increase in exposure time, noise accu-

mulates.13 This is observed most easily by evaluating

the blue channel in dark image areas. Noise appears as

a random speckled pattern, shown in Fig. 18. The blue

channel is used because CCD detectors have their poor-

est sensitivity to the blue region of the visible spectrum,

and therefore are most affected by noise accumulation,

particularly when incandescent illumination is used.

From a conservation perspective, minimizing exposure

times is also desirable, often a problem with scanbacks.4

Techniques to evaluate image noise in digital capture

are described in Refs. 2, 13, 36, and 37.

In conventional photography, many films are “very for-

giving” if the image is either under- or over-exposed.

Because most films are designed for photographing

scenes with very large dynamic ranges, there is more

exposure latitude when capturing paintings. CCD de-

tectors have less exposure latitude; an incorrect expo-

sure is “less forgiving.” If the exposure time is too long,

the CCD array saturates causing blooming. A specular

highlight smears in the direction of the one-dimensional

array. Also, light areas are clipped, mapped to the same

maximum value, resulting in a loss of highlight infor-

mation. If the exposure time is too short, the raw data

do not encompass the full range of possible digital val-

ues. This results in quantization error.

A technique to evaluate the effects of quantization on

color quality is to define an imaging system, assume

that there is 1 bit of uncertainty (accomplished by add-

ing 1 bit to the signal), and evaluate the effect of this

uncertainty on colorimetric error.38 For this analysis,

only ∆L* errors were considered based on analyzing the

Kodak gray scale. If the raw camera data were 16 bit

and the digital image encompassed the full possible

dynamic range (0 – 65,535), there would not be any sig-

nificant error as shown in Table III. When the quanti-

zation is reduced to 12 bits, errors appear, largely for

the darker samples. The 10-bit (0 – 1023) level of error

could be observable in smoothly varying colors. At 8 bits

(0 – 255), there are large errors, particularly for dark

colors. If exposure times are too short, this is equiva-

lent to reducing the bit depth. Thus the 10 and 8 bit

data shown in Table III correspond to the effects of un-

derexposure using a digital camera. Digital uncertainty

from linear signals has a greater visual effect for dark

colors. The systematic trend, in which errors increase

with decreasing lightness, is explained by the human

visual system’s nonlinear lightness response, shown in

Fig. 5. An error of one bit for a dark color results in a

larger visual difference in lightness than a one-bit error

Figure 17. CIE94 histogram of the data given in Table I.

TABLE II. List of Proper Digital Capture Practices

Object and image planes parallel

Correct camera aperture for appropriate depth of field

Spatially uniform illumination across the object

Sufficient amount of illumination

Amount of ultraviolet and infrared radiation minimized (or eliminated by

filtering)

Flare (stray light) minimized

Specular reflections minimized

Digital cropping minimized

Characterization targets included along with object

Exposure time appropriate to maximize dynamic range of raw digital

data

Exposure times or amplification for each channel appropriate to yield

gray balance

Figure 18. Incorrect exposure time or an insufficient amount

of light results in excessive image noise, shown for the blue

channel.

316 Journal of Imaging Science and Technology®

Berns

for a light color. This is seen in the 8 bit data. The er-

rors decrease with increasing lightness of the gray scale.

This would result in large banding, especially in shad-

ows and dark colors. Shadow detail would also be lost.

Clearly, it is critical to set the exposure time properly.

Based on this simple analysis, a camera’s analog to

digital converter should have a minimum of 12 bits. This

is in agreement with Olson39 who found that quantiza-

tion errors should result in a maximum ∆L* of 0.25 or

less. His analysis accounted for both the human visual

system’s color and spatial properties. Hill40 performed

an extensive computational analysis for theoretical and

actual color gamuts and also concluded that linear en-

coding required 11 or 12 bits.

As Table III shows, 8-bit quantization results in large

visual errors for linear signals. However, when 12 or 16

bit data are transformed to 8 bit per channel data, there

isn’t a restriction that the transformation be linear. It

was noted above that by coincidence, the nonlinearities

of CRT displays are nearly the exact inverse of the hu-

man visual system’s nonlinearity (seen by comparing

Figs. 5 and 11). As a consequence, gamma correcting

signals as part of the bit-reduction transformation has

the “free” benefit of reducing visual errors. This is shown

in Table IV in which 12 bit signals encompassing the

full dynamic range (0 – 4095) are gamma corrected [i.e.,

Eq. 2] before being mapped to 8 bits. Mapping linear

12-bit raw signals nonlinearly onto 8-bit signals is

clearly advantageous. The level of improvement is

equivalent to increasing bit depth.

The relationship between incident light, gamma-cor-

rected signals, and normalized lightness, i.e., CIE L*/

100, are plotted in Fig. 19. All these curves have simi-

lar shape; their differences are not significant in terms

of how the encoding affects quantization errors. As a

comparison, a Kodak gray scale was photographed us-

ing 4” × 5” positive film in a museum photographic de-

partment, then digitized using a drum scanner with a

linear response. The relationship between incident light

and raw 12-bit signals is plotted in Fig. 20. This

nonlinearity is typical of conventional photography,

though quite different than what is usually shown as

the nonlinear response of film. This is for two reasons:

linear signals are plotted rather than logarithmic sig-

nals and when photographing paintings using controlled

TABLE III. The Effects of Quantization Levels on Lightness

Errors, ∆L*

Kodak gray scale L*

∆L* for

∆L* for

∆L* for

∆L* for

16 bits

12 bits

10 bits

8 bits

19

11.70

0.01

0.15

0.47

2.22

18

11.86

0.01

0.15

0.50

2.22

17

14.27

0.01

0.13

0.50

1.93

B

16.54

0.01

0.11

0.50

1.72

15

19.43

0.01

0.09

0.49

1.43

14

21.53

0.01

0.08

0.48

1.33

13

25.46

0.00

0.07

0.47

1.05

12

27.63

0.00

0.06

0.46

0.95

11

31.41

0.00

0.05

0.45

0.82

10

35.09

0.00

0.04

0.44

0.70

9

38.62

0.00

0.04

0.43

0.62

8

43.56

0.00

0.03

0.42

0.52

M

48.08

0.00

0.03

0.41

0.45

6

53.35

0.00

0.02

0.40

0.38

5

59.17

0.00

0.02

0.39

0.32

4

65.66

0.00

0.02

0.39

0.28

3

72.53

0.00

0.01

0.01

0.24

2

79.37

0.00

0.01

0.35

0.20

1

87.88

0.00

0.01

0.41

0.17

A

96.63

0.00

0.01

0.01

0.14

Average ∆L*

0.00

0.06

0.40

0.88

Maximum ∆L*

0.01

0.15

0.50

2.22

TABLE IV. The Effects of Gamma on Lightness Errors, ∆I*.

(See text for an explanation of the computations.)

Kodak gray scale

L*

∆L* for

∆L* for ∆L* for ∆L* for ∆L* for scanned

γがんま = 3

γがんま =2.5

γがんま = 2

γがんま = 1

photography

19

11.70 0.00

0.00

0.01

2.22

1.43

18

11.86 0.03

0.04

0.08

2.22

1.43

17

14.27 0.17

0.20

0.27

1.93

1.31

B

16.54 0.23

0.26

0.34

1.72

1.13

15

19.43 0.29

0.31

0.38

1.43

1.01

14

21.53 0.31

0.33

0.39

1.33

0.92

13

25.46 0.35

0.36

0.40

1.05

0.76

12

27.63 0.36

0.37

0.41

0.95

0.69

11

31.41 0.38

0.38

0.40

0.82

0.59

10

35.09 0.39

0.39

0.40

0.70

0.49

9

38.62 0.40

0.39

0.39

0.62

0.43

8

43.56 0.41

0.39

0.38

0.52

0.34

M

48.08 0.42

0.39

0.37

0.45

0.27

6

53.35 0.42

0.39

0.36

0.38

0.21

5

59.17 0.43

0.38

0.35

0.32

0.17

4

65.66 0.43

0.38

0.34

0.28

0.18

3

72.53 0.43

0.38

0.32

0.24

0.24

2

79.37 0.43

0.37

0.31

0.20

0.32

1

87.88 0.43

0.37

0.30

0.17

0.40

A

96.63 0.44

0.36

0.29

0.14

0.46

Average ∆L*

0.34

0.32

0.32

0.88

0.64

Maximum ∆L*

0.44

0.39

0.41

2.22

1.43

Figure 19. Relationship between light, expressed as luminance

factor, and lightness (solid line) or gamma-corrected signals (γがんま

= 3: dashed line; γがんま = 2.5: dotted line; γがんま = 2: dot-dashed line).

The Science of Digitizing Paintings for Color-Accurate Image Archives: A Review

Vol. 45, No. 4, July/August 2001 317

illumination, the full dynamic range of conventional film

is not utilized. These data were used to perform a quan-

tization error analysis in which the photographic

nonlinearity was used rather than a gamma function in

transforming 12-bit to 8-bit per channel data, also shown

in Table IV. The photographic nonlinearity has much

larger error then any of the gamma-corrected signals.

That is, the digitized photographic transparency re-

sulted in greater quantization errors than direct digi-

tal capture.

There are several different gamma functions used in

the imaging industry. These can be either de facto or

legitimate national or international standards. Choices

include 1.8, the system gamma of Apple computer sys-

tems; 2.2, the inherent nonlinearity of CRT displays

when modeled by Eq. 1 and recently incorporated into

the sRGB display encoding;23 3.0, the nonlinearity in-

corporated into CIELAB;11 2.3, an empirical fit to CIE

L* used in the RLAB color-appearance model.41 The CIE

L* equation, L* = 116(Y/Yn)1/3 – 16 where Y is luminance

factor of a stimulus and Yn is the luminance factor of

the white object-color stimulus, has an offset term. If

the offset is removed and the following equation is fit:

L*/100 = (Y/Yn)1/γがんま, a gamma of 2.3 results when the

independent data are uniformly sampled by L* (i.e.,

Y/Yn = 0.00, 0.01, 0.03, 0.06, 0.11 … ) and 2.4 when the

independent data are uniformly sampled by Y/Yn (i.e.,

Y/Yn = 0.00, 0.10, 0.20, 0.30, 0.40 … ). When an offset is

included, the sRGB exponent is 2.4. The analyses shown

in Tables III and IV were repeated. For γがんま = 1.8 the aver-

age error was 0.35 ∆L*; for the other exponents, the

average was 0.34. These differences were not statisti-

cally significant, therefore, the particular exponent is

not critical.

The above analyses lead to the following recommen-

dations. First, the digital range of the raw data should

be maximized, but without clipping. For 12-bit data, the

maximum digital values should not exceed about 4000,

likely specular highlights. For sample A of the Kodak

Gray Scale, the average digital value should be around

3900. For 16-bit data, the maximum digital values

should not exceed about 64,000 and sample A of the

Kodak Gray Scale should have an average digital value

of about 63,000. When writing images in standardized

formats such as TIFF, the raw data should be directly

saved as 16-bit TIFF. For 12-bit data, the digital values

are premultiplied by 24, i.e., they are bit shifted so that

the images can be viewed in Photoshop™ without ad-

justment. This does not affect quantization errors. If

this premultiplification (bit shifting) is not performed

the displayed image will be very dark. For 8-bit TIFF,

the raw data must be rescaled and gamma corrected.

The maximum digital value should not exceed about 250.

The National Archives and Records Administration

specifies digital values of 247 and 8 for patches A and

19, respectively of the Kodak Gray Scale.42 For sample

A of the Kodak Gray Scale, the average digital value

should be around 245. The minimum values for all bit

depths should approach 0. However, for a variety of com-

putational and image noise reasons, the values for

sample 19 of the Kodak Gray scale will be about 5, 80,

or 1000 for 8-, 12-, and 16-bit data, respectively. The

key is ensuring that the available range of digital val-

ues is maximized during image acquisition but without

clipping. That is, it is critical to achieve proper expo-

sure. If software is available to display raw data histo-

grams, this can provide the needed information. Proper

exposure should not be verified by looking at the digi-

tized image on a monitor. Because the display is 8 bits

per channel, camera software will have user controlled

adjustments to impose a transfer function, sometimes

called a “process curve”, to convert from raw to display

data. The same curve will be used to generate the 8 bit

TIFF file. Although the image may look reasonable,

quantization error is still being introduced.

Optimal Encoding – Color

Dealing with color is significantly more complex than

optimally encoding lightness. As described above, most

digital cameras in use today have inappropriate spec-

tral sensitivities. If the sensor was optimized for photo-

graphic materials, shown in Fig. 6, or for consumer

applications, shown in Fig. 7, they are not accurate color

measurement instruments. For scientific imaging, the

camera should be thought of as an imaging colorimeter

or spectrometer. Several systems have been designed

as imaging colorimeters: the IBM Pro 3000 system,43–45

the VASARI system,46,47 and the MARC system.48 Sys-

tems designed as imaging spectrometers are still at the

research and development stage, summarized in Refs.

49, 50 and 51.

An experiment was performed to evaluate to colori-

metric potential of four imaging systems, three digital

cameras and scanned conventional photography. Cam-

era A has spectral sensitivities closely related to the

human visual system, similar to those shown in Fig. 8.

It consists of a monochrome sensor and three color fil-

ters. Camera B has spectral sensitivities similar to those

shown in Fig. 7. It consists of a two-dimensional color

CCD array. Camera C has spectral sensitivities similar

to those shown in Fig. 6. It is a scanback employing a

flat-bed scanner-type trilinear color-filter array. An 8”

× 10” view camera along with tungsten-balanced chrome

64 ISO film was used for conventional photography.

Color compensation filters were used to achieve a vi-

sual color match between neutral areas of the test tar-

get and its reproduction. The transparency was digitized

Figure 20. Relationship between light, expressed as luminance

factor, and normalized 12-bit raw signals for a digitized pho-

tographic positive transparency.

318 Journal of Imaging Science and Technology®

Berns

using a flat-bed scanner with densitometric spectral

sensitivities similar to Fig. 6. The illumination for all

four systems was tungsten halogen with a correlated

color temperature around 3200 K. Image capture was

performed adhering to the practices outlined in Table

II. All four digital systems have 12-bit analog to digital

converters. The raw signals were recorded as 16-bit

TIFF uncompressed. There was not any signal process-

ing such as gamma correction imposed. Thus, these im-

ages were “linear RGB.” However, because of differences

in spectral sensitivity, light sensitivity (dynamic range),

the spectral characteristics of tungsten illumination,

and flare, the range of digital values were quite differ-

ent for each channel and for each imaging system.

Three targets were imaged: the Kodak Gray Scale, the

GretagMacbeth ColorChecker Color Rendition Chart,

and a custom target of 68 artist oil paints, shown in

Web Fig. 3. Each of the paints was mixed with titanium

white in order to maximize the spectral “fingerprint” of

a given pigment. The Kodak Gray Scale and

ColorChecker chart were used to derive a transforma-

tion from R, G, and B to L*, a*, and b* values. The paint

target was used to evaluate the accuracy of the trans-

formation. The transformation can be thought of as

employing color management principles.

The first step was to develop a transform that linear-

ized the raw data with respect to CIELAB. The general

equation is shown as follows:

L

k

k

d

k

k

g

g

g

o

*

,

,

,

,

=

+

−

(

)











 +

2

1

1

65 535

1

γがんま

(3)

where L* is the measured L* values of the Kodak Gray

Scale, d is either the average raw R, G, or B digital val-

ues of each gray scale sample, and the remaining terms

are model coefficients. This particular equation enables

amplification (kg,1), properties of the ADC (d/65,535),

flare (ko), and differences in measurement geometry

between the spectrophotometer used to measure the

gray scale and the digital capture system (kg,2) to be

taken into account. In most cases, not all of the model

terms were statistically significant and the equation was

reduced in complexity accordingly. This equation can be

thought of as a more analytical form of the exponential

model described by Eq. 1. Nonlinear optimization was

used to estimate the model parameters for each chan-

nel and for each camera. The camera signals, linear-

ized with respect to L*, are referred to as R*, G*, and

B*. Equation 3 also facilitates gray balance. In the case

of digitized conventional photography, a different ap-

proach was taken to linearize the raw signals because

the nonlinearity shown in Fig. 22 cannot be well fit us-

ing Eq. 3. A fifth-order polynomial was used, Eq. 4:

L

d

d

d

d

d

o

*

,

,

,

,

,

=

+











 +



















+











 +











 +



















100

65 535

65 535

65 535

65 535

65 535

1

2

2

3

3

4

4

5

5

βべーた

βべーた

βべーた

βべーた

βべーた

βべーた

(4)

where βべーた 0 – βべーた5 are model coefficients estimated by linear

optimization.

The average R*, G*, and B* values of each color patch

of the ColorChecker were used to develop a linear trans-

formation, Eq. 5:

ˆ*

*

*

*

ˆ*

*

*

*

ˆ*

*

*

*

,

,

,

,

,

,

,

,

,

L

R

G

B

a

R

G

B

b

R

G

B

=

+

+

=

+

+

=

+

+

βべーた

βべーた

βべーた

βべーた

βべーた

βべーた

βべーた

βべーた

βべーた

11

12

13

21

22

23

31

32

33

(5)

where each βべーた term is a model coefficient. Hats (“^”) are

shown over the CIELAB values because these are esti-

mated values. The model coefficients were constrained

via Eq. 6:

βべーた

βべーた

βべーた

βべーた

βべーた

βべーた

βべーた

βべーた

βべーた

11

12

13

21

22

23

31

32

33

100

0

0

,

,

,

,

,

,

,

,

,

+

+

=

+

+

=

+

+

=

(6)

The purpose of Eq. 6 was to maintain gray balance.

The model parameters for each camera were estimated

using least-squares linear optimization. A numerical ex-

ample of this approach to color management is presented

in Ref. 6. This is referred to as a colorimetric transfor-

mation. Although polynomial expansions of Eq. 5 are

often used for color management, these often do not re-

sult in improved performance when independent data

are evaluated.

Before evaluating the independent data, the oil paint

target, it is useful to evaluate the modeling perfor-

mance of the Kodak Gray Scale and ColorChecker

chart. The average CIE94 color difference between

measured and estimated values for the Gray Scale var-

ied between 0.5 and 1.0. The estimation performance

for the ColorChecker chart is shown in Figs. 21 and

22. As expected, the closer the camera’s spectral sensi-

tivities are to the human visual system, the better the

colorimetric estimation. The causes for the large er-

rors for camera C were discussed above. The errors for

cameras A and B are random indicating that this simple

transformation is performing appropriately. It was

surprising that the performance for camera B was not

better. The MARC system, having similar spectral sen-

sitivities, has a reported48 estimation error of 2.5∆E*ab

for the ColorChecker whereas for this analysis, the

average was 6∆E*ab. Further analyses revealed that the

colorimetric performance was very sensitive to the il-

lumination system’s spectral power distribution, spa-

tial uniformity, and illuminance, as well as setting the

optimal exposure time. Prefiltering the source in or-

der to balance the three channels and improve its day-

light characteristics, numerical flat-fielding in order

to improve spatial uniformity, signal averaging, and

more careful attention to exposure reduced the aver-

age estimation error to 3.5∆E*ab. The Marc I camera

employed a Sony ICX021AK sensor whereas camera B

employed a Sony ICX085AK sensor. The latter sensor

has improved light sensitivity but its spectral sensi-

tivities are poorer approximations to the human visual

system. Thus, the reduction in estimation accuracy

from 2.5 to 3.5∆E*ab is consistent with these differences.

All of the camera systems’ colorimetric estimation ac-

curacies were highly dependent on image capture char-

acteristics underscoring the importance of proper

capture procedures listed in Table II. The colorimetric

accuracy of digitized film was intermediate between

cameras B and C, consistent with film’s spectral sensi-

tivities shown in Fig. 9.

The estimation accuracy for the artist oil paint target

is shown in Figs. 23 through 25. The trends for the direct

The Science of Digitizing Paintings for Color-Accurate Image Archives: A Review

Vol. 45, No. 4, July/August 2001 319

Figure 21. CIELAB a*b* projection vector plots showing estimation errors for the ColorChecker chart.

Figure 22. CIELAB L*C*ab projection vector plots showing estimation errors for the ColorChecker chart.

320 Journal of Imaging Science and Technology®

Berns

digital capture systems were the same as the

ColorChecker accuracy: Camera A had the least estima-

tion error while camera C had the greatest estimation

error. The digitized photographic system had the poorest

estimation accuracy of the four imaging systems. For each

imaging system, there were systematic errors; that is,

certain color regions were not accurately estimated. This

is observed by evaluating the error vectors. For example,

the +a*+b* and –a*–b* quadrants had vectors pointed in

similar directions for camera C. Lightness was underes-

timated for the scanned photographic system. Ideally,

error vectors should be largely random and small, such

as those shown for the ColorChecker in Fig. 21. Because

CIELAB is not visually uniform, caution should be taken

when evaluating error-vector plots. The same length vec-

tor corresponds to a different visual difference depend-

ing on the chroma of the measured color. Thus, the overall

accuracy is judged by looking at the histograms plotted

in Fig. 25. Clearly, camera A has the highest estimation

accuracy. One important property is to minimize the

maximum errors. Camera A’s largest error is 6.8 ∆E*94.

The errors form a tight grouping. As a comparison, the

digitized film has errors up to 12 ∆E*94 and the grouping

is broadly dispersed.

The differences in estimation accuracy between the

ColorChecker and the artist paint target, underscores

the importance of independent verification. Achieving

high estimation accuracy for a characterization target

is much simpler than achieving high accuracy when

using the system for imaging paintings.

As an estimation problem, residual errors result, the

magnitude of the errors dependent on the inherent prop-

erties of the camera and also on the method of estima-

tion. A linear method has been described; nonlinear

methods may result in improved performance, shown

in Ref. 6 or using higher order matrices or multi-dimen-

sional look-up tables. This is an evolving area of re-

search. However, no matter how complex the color

management system, the trends described above would

persist: Digitized photographic positive transparencies

and densitometric-type scanbacks (Camera C) will al-

ways result in large estimation errors; the improvement

in performance is correlated with the similarity of a in-

put device’s spectral sensitivities to the human visual

system.

The above analyses lead to the following recommen-

dations: The archival master image should be the raw

data, converted to a standardized format such as 16-bit

TIFF uncompressed. Ideally, the image should include

the work of art, a gray scale, and a color target. The

color target should replace the Kodak Separation Guide.

Because the TIFF format supports image tags (TIFF

stands for Tagged Image File Format), spectral data of

the gray scale and color targets should be a part of the

tag. If possible, the spectral power distribution of the

illumination and the spectral sensitivities of the cam-

era system should also be included. If it isn’t practical

to include characterization targets in the image, the av-

erage digital counts of each target’s color patches, cap-

tured under identical conditions, should be included in

Figure 23. CIELAB a*b* projection vector plots showing estimation errors for the artist oil target.

The Science of Digitizing Paintings for Color-Accurate Image Archives: A Review

Vol. 45, No. 4, July/August 2001 321

the tag. A dictionary of standard image metadata is in

preparation by the National Information Standards Or-

ganization, NISO.52 Images of the characterization tar-

gets should also be archived. If 8 bit data must be stored,

gamma correction between 1.8 and 2.4 should first be

applied. Color management should not be applied to the

digital master. Rather, color management should be

applied to various derivative images. The necessary in-

formation to implement reasonable color management

stems from the characterization targets.

The Future

As color measurement developed during the 20th Cen-

tury, there was a transition from predominantly colo-

rimeters to predominantly spectrophotometers. That is,

trichromatic instruments were replaced with spectral

instruments. The principal advantages were numerous.

The spectral data enabled an object’s color to be calcu-

lated for any illuminant and observer of interest. Thus,

the quality of a color match could be evaluated under

any number of illuminating and viewing conditions.

Second, the spectral data enabled a variety of analyti-

cal analyses such as colorant identification, instrumen-

tal-based color matching, and colorant strength.6 Finally,

and perhaps most importantly, the presence or absence

of metamerism could be readily observed. As an illus-

trative example, Pablo Picasso’s The Tragedy was re-

stored in 1937 and 1997. During 1937, paint losses were

filled and inpainted. See the National Gallery of Art

website: www.nga.gov. The inpainted areas appear pur-

plish when photographed, easily seen in the bottom-

right portion of the web image. Spectral measurements

of untreated and inpainted areas reveal that Picasso

used prussian blue while the conservator used ultra-

marine blue, shown in Fig. 26. Unfortunately, most of

the 1937 inpainting could not be removed during 1997;

hence the pigment mismatch persists. Because the spec-

tral sensitivity of the red sensitive layer of film is shifted

towards longer wavelengths compared with the human

visual system as shown in Fig. 9, the ultramarine was

reproduced as a grayish purple rather than a grayish

blue. Spectral measurements would have alerted the

conservator that a metameric match was produced. This

specific problem with blue pigments has been described

by Staniforth.53 Clearly, an imaging spectrometer would

be a useful analytical tool. Furthermore, the lack of colo-

rimetric accuracy, causing the large color shift in this

example, and the need to standardize an illuminant and

observer, as described above, would be eliminated.

Spectral-based imaging of artwork is currently at a

research stage, summarized in Refs. 49 through 51. Us-

ing camera A, a novel approach to spectral imaging was

evaluated in which a second image, taken by filtering

the camera lens with a light blue filter (Wratten 38),

was combined computationally to estimate the spectral

reflectance of an image area.54–57 The spectral and cam-

era data of the Kodak Gray Scale and ColorChecker

chart were used to derive the necessary transforms.

Following spectral estimation, CIELAB coordinates

Figure 24. CIELAB L*C*ab projection vector plots showing estimation errors for the artist oil target.

322 Journal of Imaging Science and Technology®

Berns

were calculated in the usual way. The ColorChecker was

estimated to an average accuracy of 0.3∆E*94 (1.6∆E*ab).

The artist’s paint target had an average accuracy of

2.0∆E*94, nearly a twofold improvement in colorimetric

accuracy.

One of the very interesting aspects of this spectral es-

timation approach is that spectra are generated from a

conventional digital color camera. The spectra can be used

for a variety of color reproduction58–61 and conservation62–

65 applications. Using only spectral information from the

ColorChecker chart, the estimated spectral reflectance

factor data for four blue pigments from the artist’s paint

target are plotted in Fig. 27. The fits are fairly typical of

spectral-estimation techniques: the estimates tend to

have greater spectral selectivity. The overall shapes are

reasonably predicted. These estimated spectra were

evaluated using a statistical method of pigment identifi-

cation based on reflectance spectrophotometry.65 The fol-

lowing pigments formed the database of possible blue

pigments: cobalt, ultramarine, manganese, prussian, ph-

thalocyanine, cerulean, and indanthrone. The database

was formed from the artist’s paint target. The cobalt and

ultramarine blue spectra were correctly identified. Man-

ganese blue was incorrectly identified as phthalocyanine

Figure 25. CIE94 color difference histogram showing estimation errors for the artist oil target

Figure 26. Spectral reflectance factor measurements of Pablo

Picasso’s The Tragedy (solid line) and 1937 inpainting (dashed line).

The Science of Digitizing Paintings for Color-Accurate Image Archives: A Review

Vol. 45, No. 4, July/August 2001 323

blue. This was due to the secondary peak at 650 nm in

the estimated spectrum. The prussian blue sample was

incorrectly identified as manganese blue. The spectral

estimation was repeated, except that spectral informa-

tion from the entire painted target was used in place of

the ColorChecker. The estimated spectra are also plot-

ted in Fig. 27. In all cases the spectral fits are improved.

The pigment identification was repeated. Only the

prussian blue sample was incorrectly identified, again

as manganese blue. Given the known similarity in spec-

tral properties between prussian and manganese blues,53

these results were very encouraging.

Conclusions

An archival digital image of a work of art should facili-

tate a variety of applications including web-based im-

ages, color reproduction, scientific study, art historical

and other scholarly studies, and most importantly, an

accurate recording of its physical properties. As de-

scribed, achieving an accurate recording is difficult and

requires specialized hardware, optimal imaging prac-

tices, and specialized software.

Two issues need to be emphasized. The first is that

scanback sensors are very common among the museum

community. Because the number of pixels is often the

dominant criterion when specifying an imaging system,

scanbacks fare quite well. However, the majority of tri-

filter scanbacks are densitometric rather than colorimet-

ric. As described above, this results in very poor color

accuracy, even following color management. There are

two solutions: redesign the filter array, e.g., Ref. 66, or

use a monochrome sensor and filter-wheel assembly with

appropriate filter spectral transmittances. In general,

greater emphasis needs to be placed on a sensor’s spec-

tral sensitivities. For accurate color imaging, sensors need

to be closely related to the human visual system. This

aspect of camera design is often overlooked. For example,

Koelling67 listed the five most important technical issues

for digitizing works of art as image resolution, file for-

mat, file storage, file refreshment, i.e., updating storage

media, and copyright. Hernandez68 defined key digital

camera selection criteria as dynamic range, optical reso-

lution, bit depth, scanning area, and software.

The second issue to be emphasized is that image edit-

ing visual techniques to improve color accuracy has not

Figure 27. Spectral estimation of four samples from the artist oil target shown in Fig. 23: measured, solid line, estimated from

spectral data of ColorChecker, dotted line, estimated from spectral data of the entire target, dashed line.

324 Journal of Imaging Science and Technology®

Berns

been included in this article. This was deliberate and

not an omission. Visual matching between objects and

their CRT representation is very complex and is highly

dependent on viewing conditions,26,41 cognitive aspects

given the dissimilar nature of light-emitting and light-

reflecting stimuli,41 the observer,7,8 and monitor calibra-

tion.23–25 Any display will be highly metameric to

reflective works of art. Furthermore, color management

must be used, an evolving area of color technology.6,15,32

Transforming the raw camera signals to a derivative

image such as an sRGB image should be a computational

exercise, not a visual adjustment. The calculations are

straightforward (a numerical example is given in Ref.

6) and eliminate the natural tendency to boost contrast

and color saturation caused by unmatched viewing con-

ditions and color gamut limitations of CRT displays.

For those engaged or soon to be engaged in creating

a digital archive, the recommendations described

above should be seriously considered. At the very

least, standardized (actual or de facto) gray scale and

color targets must be included in each image. The

image metadata or image tag must contain informa-

tion about the targets and details of the image-cap-

ture system.

Acknowledgment. This publication was written while

the author was a Senior Fellow in Conservation Science

at the National Gallery of Art, Washington; the Gallery’s

financial support is sincerely appreciated. The author

gratefully acknowledges the assistance of Janet Bliburg,

Lorene Emerson, and Lyle Peterzell in collecting digi-

tal and conventional images, and Ross Merrill in pro-

ducing the artist oil paint chart, all of the National

Gallery of Art, Washington. Finally, the following imag-

ing professionals were very helpful in reviewing manu-

scripts and providing the benefits of their extensive

experiences: Franziska Frey, Image Permanence Insti-

tute; Connie McCabe, National Gallery of Art; Edward

Giorgianni, Eastman Kodak Company; Mike Collette,

Better Light; John Stokes, Stokes Imaging; Michael

Stokes, Microsoft Corporation; and Francisco Imai,

Munsell Color Science Laboratory. Francisco Imai also

performed the spectral estimations.

References

1. S. Chapman, Guidelines for image capture, Research Libraries Group,

www.rlf.org/preserv/joint/chapman.html (1998).

2. F. S. Frey and J. M. Reilly, Digital Imaging for Photographic Collec-

tions: Foundations for Technical Standards, Image Permanence In-

stitute, Rochester Institute of Technology, Rochester, New York, 1999.

3. Guides to quality in visual resource imaging, Research Libraries

Group, www.rlg.org/visguides/index.html (2000).

4. B. Blackwell, Light exposure to sensitive artworks during digital pho-

tography, Spectra 26(2), 24–28 (2000).

5. N. Lossau and M. Liebetruth, Conservation issues in digital imaging,

Spectra 26(2), 30–36 (2000).

6. R. S. Berns, Billmeyer and Saltzman’s Principles of Color Technol-

ogy, 3rd ed., John Wiley & Sons, New York, 2000.

7. B. A. Wandell, Foundations of Vision, Sinauer Assoc. Inc., Sunderland

MA, 1995.

8. P. K. Kaiser and R. M. Boynton, Human Color Vision, 2nd ed., Optical

Society of America, Washington, DC, 1996.

9. A. Stockman, D. I. A. MacLeod and N. E. Johnson, Spectral sensitivi-

ties of the human cones, J. Opt. Soc. Am. A 10, 2491–2521 (1993).

10. ISO/IEC DIS 10918-4, Information technology—Digital compression

and coding of continuous-tone still images: Registration of JPEG pro-

files, SPIFF profiles SPIFF tags, SPIFF colour spaces, APPN mark-

ers, SPIFF compression types and Registration Authorities

(REGAUT)’, International Organization for Standardization, Geneva,

Switzerland.

11. CIE No. 15.2, Colorimetry, 2nd ed., Commission Internationale de

l’Éclairage, Vienna, Austria, 1986.

12. D. S. Falk, D. R. Brill and D. G. Stork, Seeing the Light: Optics in

Nature, Photography, Color, Vision, and Holography, John Wiley &

Sons, New York, 1986.

13. G. Holst, CCD Arrays, Cameras, and Displays, 2nd ed, SPIE, Intl. Soc.

For Optical Eng., Bellingham, WA, 1998.

14. H. E. Ives, The transformation of color-mixture equations from one

system to another, J. Franklin Inst. 180, 673–701 (1915). Reproduced

in D. L. MacAdam, Ed., Selected Papers on Colorimetry – Funda-

mentals, SPIE Milestone Series, Vol. MS 77, SPIE, Intl. Soc. For

Optical Eng., Bellingham WA, 1993, pp. 57–71.

15. E. J. Giorgianni and T. E. Madden, Digital Color Management Encod-

ing Solutions, Addison Wesley, Reading, MA, 1998.

16. F. Frey and R. Gschwind, Mathematical bleaching models for photo-

graphic three-color materials, J. Imag. Sci. Technol. 38, 513–519

(1994).

17. R. Gschwind and F. Frey, Electronic imaging, a tool for the recon-

struction of faded color photographs, J. Imag. Sci. Technol. 38, 520–

525 (1994).

18. R. S. Berns and M. J. Shyu, Colorimetric characterization of a desk-

top drum scanner using a spectral model, J. Electronic Imag. 4, 360–

372 (1995).

19. http://www.pima.net/standards/iso/tc42/wg18/kp_sfr_measure.htm

from the Photographic and Imaging Manufacturing Association.

20. C. D. Child, Discharge from hot CaO, Phys. Rev. 32, 492–511 (1911).

21. I. Langmuir, The effect of space charge and residual gases on ther-

mionic current in high vacuum, Phys. Rev. 2, 450–486 (1913).

22. T. H. James and G. C. Higgins, Fundamentals of Photographic Theory,

Morgan & Morgan Inc., Hastings-on-Hudson, NY, 1968.

23. IEC 61966-2-1, Colour measurement and management in multime-

dia systems and equipment—Part 2-1: Colour management—Default

RGB colour space—sRGB, International Electrotechnical Commis-

sion, Geneva, Switzerland, 1999.

24. R. S. Berns and N. Katoh, The digital to radiometric transfer function

for computer controlled CRT displays, Proc. CIE Expert Symposium

‘97 Colour Standards for Image Technology, Commission

Internationale de l’Éclairage, Vienna, Austria, 1997, pp 34–37, 1997.

25. T. Deguchi, N. Katoh and R. S. Berns, Clarification of ‘gamma’ and

the accurate characterization of CRT monitors, Proc. SID 30, 786–

789 (1999).

26. R. W. G. Hunt, The Reproduction of Colour in Photography, Printing,

and Television, 5th ed., Fountain Press, England, 1995.

27. ISO 14524, Photography—Electronic still picture cameras—Methods

for measuring opto-electronic conversion functions (OECFs), Inter-

national Organization for Standardization, Geneva, Switzerland. See

also

http://www.pima.net/standards/iso/tc42/wg18/

WG18_POW.htm#14524.

28. ISO 12641, Graphic technology — Prepress digital data exchange —

Colour targets for input scanner calibration, International Organiza-

tion for Standardization, Geneva, Switzerland. (Commonly known as

ANSI IT8.7/1 and 7/2.)

29. ISO 12642 Graphic technology – prepress digital data exchange –

input data for characterization of 4-colour process printing, Interna-

tional Organization for Standardization, Geneva, Switzerland. (Com-

monly known as ANSI IT8.7/3.)

30. C. S. McCamy, H. Marcus and J. G. Davidson, A color-rendition chart,

J. Appl. Phot. Eng. 2, 95–99 (1976).

31. ISO 13655, Graphic technology—Spectral measurement and colori-

metric computation for graphic arts images, International Organiza-

tion for Standardization, Geneva, Switzerland.

32. ICC.1: 1998-09, File format for color profiles International Color Con-

sortium, and ICC.1A: 1999-04, ‘Addendum 2 to Spec. ICC.1:1998-

09,’ www.color.org (1999).

33. CIE No. 116, Industrial colour-difference evaluation, Commission

Internationale de l’Éclairage, Vienna, Austria, 1995.

34. M. Stokes, M. D. Fairchild and R. S. Berns, Colorimetrically quanti-

fied tolerances for pictorial images, TAGA part 2, 757–778 (1992).

35. T. Song and R. Luo, Testing color-difference formulae on complex

images using a CRT monitor, Proc. 8th IS&T/SID Color Imaging Con-

ference, IS&T, Springfield, VA, 2000, pp 44–48.

36. ISO 15739, Photography—Electronic still picture cameras—Noise

measurements, International Organization for Standardization,

Geneva, Switzerland.

37. P. D. Burns and R. S. Berns, Image noise and colorimetric precision

in multispectral image capture, Proc. 6th IS&T/SID Color Imaging Con-

ference, IS&T, Springfield, VA, 1998, pp 83–85.

38. P. D. Burns and R. S. Berns, Quantization in multispectral color im-

age acquisition, Proc. 7th IS&T/SID Color Imaging Conference, IS&T,

Springfield, VA, 1999, pp 32–35.

39. T. Olson, Smooth ramps: Walking the straight and narrow path through

color space, Proc. 7th IS&T/SID Color Imaging Conference, IS&T,

Springfield, VA, 1999, pp 57–64.

40. B. Hill, T. Roger and F. W. Vorhagen, Comparative analysis of the

quantization of color spaces on the basis of the CIELAB color-differ-

ence formula, ACM Trans. Graphics 16, 109–154 (1997).

41. M. D. Fairchild, Color Appearance Models, Addison-Wesley, Read-

ing, MA, 1998.

42. S. Publia and B. Roginski, NARA guidelines for digitizing archival ma-

terials for electronic access, National Archives and Records Adminis-

The Science of Digitizing Paintings for Color-Accurate Image Archives: A Review

Vol. 45, No. 4, July/August 2001 325

tration, College Park, MD (1998). See also http://www.nara.gov/nara/

vision/eap/eapspec.html.

43. F. C. Mintzer, L. E. Boyle, A. N. Cazes, B. S. Christian, S. C. Cox, F.

P. Giordano, H. M. Gladney, J. C. Lee, M. L. Kelmanson, A. C. Lirani,

K. A. Magerlein, A. M. B. Pavani, and F. Schiattarella, Towards on-

line worldwide access to Vatican library materials, IBM J. Research

and Development, 40, 139–162 (1996). Also available at http://

www.almaden.ibm.com/journal/rd/mintz/mintzer.html.

44. F. Mintzer, Developing Digital Libraries of Cultural Content for Internet

Access, IEEE Communications 37(1), 72–78 (1999).

45. F. P. Giordano, G. W. Braudaway, J. Christensen, J. Lee, and F.

Mintzer, Evolution of a high-quality digital imaging system, Proc. SPIE

3650, 110–118 (1999).

46. K. Martinez, J. Cupitt and D. Saunders, High resolution colorimetric

imaging of paintings, Proc. SPIE 1901, 25–36 (1993).

47. D. Saunders and J. Cupitt, Image processing at the National Gallery:

The VASARI projec, National Gallery Tech. Bull. 14, 72–86 (1993).

48. A. Burmester, L. Raffelt, G. Robinson, and S. Wagini, The MARC

project: From analogue to digital reproduction, in A. Burmester, L.

Raffelt, K. Renger, G. Robinson, and S. Wagini, Flämische

Barockmalerei: Meisterwerke der Alten Pinakothek München; Flem-

ish Baroque Painting: Masterpieces of the Alte Pinakothek München,

Hirmer Verlag, München, 1996, pp. 19-26.

49. L. MacDonald and M. R. Luo, Eds., Colour Imaging: Vision and Tech-

nology, John Wiley & Sons, Chichester, 1999.

50. Proc. Intl. Symp. Multispectral Imaging and Color Reproduction for Digi-

tal Archives, Chiba University, Soc. Multispectral Imaging Japan, 1999.

51. Proc. Second Intl. Symp. Multispectral Imaging and Color Reproduc-

tion for Digital Archives, Chiba University, Soc. Multispectral Imaging

Japan, 2000.

52. See http://www.niso.org/PRImagemeta.html.

53. S. Staniforth, Retouching and colour matching: The restorer and

metamerism, Studies in Conservation 30, 101–111 (1985).

54. F. H. Imai and R. S. Berns, High-resolution multi-spectral image ar-

chives—A hybrid approach, Proceedings IS&T/SID Sixth Color Imag-

ing Conference, IS&T, Springfield, VA, 1998, pp 224–227.

55. R. S. Berns and F. H. Imai, Spectral estimation using trichromatic

digital cameras, Proc. Intl. Sym. Multispectral Imaging and Color

Reproduction for Digital Archives, Soc. Multispectral Imaging Ja-

pan, Chiba University, 1999, pp 42–49.

56. F. H. Imai and R. S. Berns, A comparative analysis of spectral re-

flectance reconstruction in various spaces using a trichromatic cam-

era system, Proc. IS&T/SID Seventh Color Imaging Conference,

IS&T, Springfield, VA, 1999, pp 21–25.

57. F. H. Imai, M. R. Rosen and R. S. Berns, Comparison of spectrally

narrow-band capture versus wide-band with a priori sample analy-

sis for spectral reflectance estimation, Proc. 8th IS&T/SID Color Im-

aging Conference, IS&T, Springfield, VA, 2000, pp 234–241.

58. R. S. Berns, Challenges for colour science in multimedia imaging

systems, in L. MacDonald, and M.R. Luo, Eds., Colour Imaging: Vi-

sion and Technology, John Wiley & Sons, England, 1999, pp. 99–

127.

59. R. S. Berns, F. H. Imai, P. D. Burns, and D. Y. Tzeng, Multi-spectral-

based color reproduction research at the Munsell Color Science

Laboratory, Proc. SPIE 3409, 14–25 (1999).

60. F. H. Imai, M. R. Rosen, R. S. Berns, and D. Tzeng, Spectral repro-

duction from scene to hardcopy I: Image input and output, Proc.

SPIE, in press (2001).

61. M. Rosen, F. Imai, X. Jiang, and N. Ohta, Spectral reproduction from

scene to hardcopy II: Image processing, Proc. SPIE, in press (2001).

62. S. Baronti, A. Casini, F. Lotti, and S. Porcinai, A multispectral imag-

ing system for mapping of pigments in works of art by the use of

principal-component analysis, Appl. Optics 37 1299–1309 (1998).

63. A. Casini, F. Lotti, M. Picollo, L. Stefani, and E. Buzzegoli, Image

spectroscopy mapping technique for non-invasive analysis of paint-

ings, Studies in Conservation 44 39–48 (1999).

64. J. Thoma, J. Cupitt and D. Saunders, An investigation of the poten-

tial use of visible-region multispectral imaging for pigment identifi-

cation in paintings, Proc. Colour Image Science 2000, Colour and

Imaging Institute, University of Derby, Derby, UK, 2000, pp 95–106.

65. R. S. Berns, J. Krueger and M. Swicklik, Multiple pigment selection

for inpainting using visible reflectance spectrophotometry, Studies

in Conservation, in press 2001.

66. S. Quan, N. Ohta and N. Katoh, Optimization of camera spectral

sensitivities, Proc. 8th IS&T/SID Color Imaging Conference, IS&T,

Springfield, VA, 2000, pp 273–278.

67. J. M. Koelling, Revealing history – digital imaging, the new photo-

graphic research tool, Spectra 26(2), 10–15 (2000).

68. C. Hernandez and R. Lilien, Building a digital archive – digitizing for

the long term, Spectra 26(2), 16–17 (2000).