Photographic Tone Reproduction for Digital Images

Erik Reinhard Michael Stark Peter Shirley James Ferwerda

University of Utah University of Utah University of Utah Cornell University

Abstract
Linear map
A classic photographic task is the mapping of the potentially high
dynamic range of real world luminances to the low dynamic range
of the photographic print. This tone reproduction problem is also
faced by computer graphics practitioners who map digital images to
a low dynamic range print or screen. The work presented in this pa-
per leverages the time-tested techniques of photographic practice to
develop a new tone reproduction operator. In particular, we use and
extend the techniques developed by Ansel Adams to deal with dig-
ital images. The resulting algorithm is simple and produces good
results for a wide variety of images.
CR Categories: I.4.10 [Computing Methodologies]: Image Pro-
cessing and Computer Vision—Image Representation
Keywords: Tone reproduction, dynamic range, Zone System.
New operator

Radiance map courtesy of Cornell Program of Computer Graphics

1 Introduction
The range of light we experience in the real world is vast, spanning
approximately ten orders of absolute range from star-lit scenes to
sun-lit snow, and over four orders of dynamic range from shad-
ows to highlights in a single scene. However, the range of light
we can reproduce on our print and screen display devices spans at
best about two orders of absolute dynamic range. This discrep-
ancy leads to the tone reproduction problem: how should we map
measured/simulated scene luminances to display luminances and
produce a satisfactory image?
A great deal of work has been done on the tone reproduction
problem [Matkovic et al. 1997; McNamara et al. 2000; McNamara
2001]. Most of this work has used an explicit perceptual model to
control the operator [Upstill 1985; Tumblin and Rushmeier 1993;
Ward 1994; Ferwerda et al. 1996; Ward et al. 1997; Tumblin et al. Figure 1: A high dynamic range image cannot be displayed directly
1999]. Such methods have been extended to dynamic and interac- without losing visible detail using linear scaling (top). Our new
tive settings [Ferwerda et al. 1996; Durand and Dorsey 2000; Pat- algorithm (bottom) is designed to overcome these problems.
tanaik et al. 2000; Scheel et al. 2000; Cohen et al. 2001]. Other
work has focused on the dynamic range compression problem by
spatially varying the mapping from scene luminances to display lu- Using perceptual models is a sound approach to the tone repro-
minances while preserving local contrast [Oppenheim et al. 1968; duction problem, and could lead to effective hands-off algorithms,
Stockham 1972; Chiu et al. 1993; Schlick 1994; Tumblin and Turk but there are two problems with current models. First, current mod-
1999]. Finally, computational models of the human visual system els often introduce artifacts such as ringing or visible clamping (see
can also guide such spatially-varying maps [Rahman et al. 1996; Section 4). Second, visual appearance depends on more than simply
Rahman et al. 1997; Pattanaik et al. 1998]. matching contrast and/or brightness; scene content, image medium,
and viewing conditions must often be considered [Fairchild 1998].
To avoid these problems, we turn to photographic practices for in-
spiration. This has led us to develop a tone reproduction technique
designed for a wide variety of images, including those having a very
high dynamic range (e.g., Figure 1).
Copyright © 2002 by the Association for Computing Machinery, Inc.
Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or 2 Background
distributed for commercial advantage and that copies bear this notice and the full
citation on the first page. Copyrights for components of this work owned by
The tone reproduction problem was first defined by photographers.
others than ACM must be honored. Abstracting with credit is permitted. To copy
otherwise, to republish, to post on servers, or to redistribute to lists, requires prior Often their goal is to produce realistic “renderings” of captured
specific permission and/or a fee. Request permissions from Permissions Dept, scenes, and they have to produce such renderings while facing the
ACM Inc., fax +1-212-869-0481 or e-mail permissions@acm.org.
© 2002 ACM 1-58113-521-1/02/0007 $5.00

267
 Darkest Brightest
textured textured
shadow Dynamic range = 15 scene zones highlight
Middle grey
2x L 2x+1 L 2x+2 L 2x+3 L 2x+4 L . . . 2x+15 L 2x+16 L
Middle grey maps to Zone V

0 I II III IV V VI VII VIII IX X

Print zones

Figure 4: The mapping from scene zones to print zones. Scene zones
at either extreme will map to pure black (zone 0) or white (zone X)
if the dynamic range of the scene is eleven zones or more.
Light

Middle-grey: This is the subjective middle brightness region of

the scene, which is typically mapped to print zone V.
Dark
Dynamic range: In computer graphics the dynamic range of a
scene is expressed as the ratio of the highest scene luminance
to the lowest scene luminance. Photographers are more inter-
Figure 2: A photographer uses the Zone System to anticipate po- ested in the ratio of the highest and lowest luminance regions
tential print problems. where detail is visible. This can be viewed as a subjective
measure of dynamic range. Because zones relate logarithmi-
cally to scene luminances, dynamic range can be expressed
as the difference between highest and lowest distinguishable
scene zones (Figure 4).

Key: The key of a scene indicates whether it is subjectively light,

normal, or dark. A white-painted room would be high-key,
and a dim stable would be low-key.

Dodging-and-burning: This is a printing technique where some

light is withheld from a portion of the print during develop-
Figure 3: A normal-key map for a high-key scene (for example con- ment (dodging), or more light is added to that region (burn-
taining snow) results in an unsatisfactory image (left). A high-key ing). This will lighten or darken that region in the final print
map solves the problem (right). relative to what it would be if the same development were
used for all portions of the print. In traditional photography
this technique is applied using a small wand or a piece of pa-
limitations presented by slides or prints on photographic papers. per with a hole cut out.
Many common practices were developed over the 150 years of pho-
tographic practice [London and Upton 1998]. At the same time A crucial part of the Zone System is its methodology for predicting
there were a host of quantitative measurements of media response how scene luminances will map to a set of print zones. The pho-
characteristics by developers [Stroebel et al. 2000]. However, there tographer first takes a luminance reading of a surface he perceives
was usually a disconnect between the artistic and technical aspects as a middle-grey (Figure 2 top). In a typical situation this will be
of photographic practice, so it was very difficult to produce satis- mapped to zone V, which corresponds to the 18% reflectance of the
factory images without a great deal of experience. print. For high-key scenes the middle-grey will be one of the darker
Ansel Adams attempted to bridge this gap with an approach he regions, whereas in low-key scenes this will be one of the lighter re-
called the Zone System [Adams 1980; Adams 1981; Adams 1983] gions. This choice is an artistic one, although an 18% grey-card is
which was first developed in the 1940s and later popularized by often used to make this selection process more mechanical (Fig-
Minor White [White et al. 1984]. It is a system of “practical sensit- ure 3).
ometry”, where the photographer uses measured information in the Next the photographer takes luminance readings of both light
field to improve the chances of producing a good final print. The and dark regions to determine the dynamic range of the scene (Fig-
Zone System is still widely used more than fifty years after its in- ure 2 bottom). If the dynamic range of the scene does not exceed
ception [Woods 1993; Graves 1997; Johnson 1999]. Therefore, we nine zones, an appropriate choice of middle grey can ensure that all
believe it is useful as a basis for addressing the tone reproduction textured detail is captured in the final print. For a dynamic range of
problem. Before discussing how the Zone System is applied, we more than nine zones, some areas will be mapped to pure black or
first summarize some relevant terminology. white with a standard development process. Sometimes such loss
of detail is desirable, such as a very bright object being mapped to
Zone: A zone is defined as a Roman numeral associated with an pure white (see [Adams 1983], p. 51). For regions where loss of
approximate luminance range in a scene as well as an approxi- detail is objectionable, the photographer can resort to dodging-and-
mate reflectance of a print. There are eleven print zones, rang- burning which will locally change the development process.
ing from pure black (zone 0) to pure white (zone X), each The above procedure indicates that the photographic process is
doubling in intensity, and a potentially much larger number of difficult to automate. For example, determining that an adobe build-
scene zones (Figure 4). ing is high-key would be very difficult without some knowledge

268
about the adobe’s true reflectance. Only knowledge of the geometry
Key value 0.09 Key value 0.18
and light inter-reflections would allow one to know the difference
between luminance ratios of a dark-dyed adobe house and a normal
adobe house. However, the Zone System provides the photogra-
pher with a small set of subjective controls. These controls form
the basis for our tone reproduction algorithm described in the next
section.

Radiance map courtesy of Paul Debevec

The challenges faced in tone reproduction for rendered or cap-
tured digital images are largely the same as those faced in conven- Key value 0.36 Key value 0.72
tional photography. The main difference is that digital images are in
a sense “perfect” negatives, so no luminance information has been
lost due to the limitations of the film process. This is a blessing in
that detail is available in all luminance regions. On the other hand,
this calls for a more extreme dynamic range reduction, which could
in principle be handled by an extension of the dodging-and-burning
process. We address this issue in the next section.

Figure 5: The linear scaling applied to the input luminance allows

3 Algorithm the user to steer the final appearance of the tone-mapped image.
The dynamic range of the image is 7 zones.
The Zone System summarized in the last section is used to develop
a new tone mapping algorithm for digital images, such as those cre-
ated by rendering algorithms (e.g., [Ward Larson and Shakespeare high luminances [Mitchell 1984; Stroebel et al. 2000]. A simple
1998]) or captured using high dynamic range photography [De- tone mapping operator with these characteristics is given by:
bevec and Malik 1997]. We are not trying to closely mimic the
actual photographic process [Geigel and Musgrave 1997], but in- L(x, y)
stead use the basic conceptual framework of the Zone System to Ld (x, y) = . (3)
1 + L(x, y)
manage choices in tone reproduction. We first apply a scaling that
is analogous to setting exposure in a camera. Then, if necessary, Note that high luminances are scaled by approximately 1/L, while
we apply automatic dodging-and-burning to accomplish dynamic low luminances are scaled by 1. The denominator causes a graceful
range compression. blend between these two scalings. This formulation is guaranteed
to bring all luminances within displayable range. However, as men-
tioned in the previous section, this is not always desirable. Equa-
3.1 Initial luminance mapping
tion 3 can be extended to allow high luminances to burn out in a
We first show how to set the tonal range of the output image based controllable fashion:
on the scene’s key value. Like many tone reproduction meth-  
ods [Tumblin and Rushmeier 1993; Ward 1994; Holm 1996], we L(x, y) 1 + L L(x,y)
2
view the log-average luminance as a useful approximation to the Ld (x, y) = white
(4)
1 + L(x, y)
key of the scene. This quantity L̄w is computed by:

 where Lwhite is the smallest luminance that will be mapped to pure
1  white. This function is a blend between Equation 3 and a linear
L̄w = exp log (δ + Lw (x, y)) (1)
N mapping. It is shown for various values of Lwhite in Figure 6. If
x,y
Lwhite value is set to the maximum luminance in the scene Lmax
or higher, no burn-out will occur. If it is set to infinity, then the
where Lw (x, y) is the “world” luminance for pixel (x, y), N is the
function reverts to Equation 3. By default we set Lwhite to the
total number of pixels in the image and δ is a small value to avoid
maximum luminance in the scene. If this default is applied to scenes
the singularity that occurs if black pixels are present in the image. If
that have a low dynamic range (i.e., Lmax < 1), the effect is a subtle
the scene has normal-key we would like to map this to middle-grey
contrast enhancement, as can be seen in Figure 7.
of the displayed image, or 0.18 on a scale from zero to one. This
suggests the equation: The results of this function for higher dynamic range images is
shown in the left images of Figure 8. For many high dynamic range
a images, the compression provided by this technique appears to be
L(x, y) = Lw (x, y) (2) sufficient to preserve detail in low contrast areas, while compress-
L̄w
ing high luminances to a displayable range. However, for very high
where L(x, y) is a scaled luminance and a = 0.18. For low-key dynamic range images important detail is still lost. For these im-
or high-key images we allow the user to map the log average to ages a local tone reproduction algorithm that applies dodging-and-
different values of a. We typically vary a from 0.18 up to 0.36 and burning is needed (right images of Figure 8).
0.72 and vary it down to 0.09, and 0.045. An example of varying is
given in Figure 5. In the remainder of this paper we call the value 3.2 Automatic dodging-and-burning
of parameter a the “key value”, because it relates to the key of the
image after applying the above scaling. In traditional dodging-and-burning, all portions of the print poten-
The main problem with Equation 2 is that many scenes have pre- tially receive a different exposure time from the negative, bringing
dominantly a normal dynamic range, but have a few high luminance “up” selected dark regions or bringing “down” selected light re-
regions near highlights or in the sky. In traditional photography gions to avoid loss of detail [Adams 1983]. With digital images we
this issue is dealt with by compression of both high and low lumi- have the potential to extend this idea to deal with very high dynamic
nances. However, modern photography has abandoned these “s”- range images. We can think of this as choosing a key value for ev-
shaped transfer curves in favor of curves that compress mainly the ery pixel, which is equivalent to specifying a local a in Equation 2.

269
 L = 0.5 1.0 1.5 3 ∞
white

Radiance map courtesy of Greg Ward (top) and Cornell Program of Computer Graphics (bottom)
Ld

0 1 2 3 4 5
World luminance (L)

Figure 6: Display luminance as function of world luminance for a

family of values for Lwhite .

Simple operator Dodging−and−burning

Input Output

Figure 7: Left: low dynamic range input image (dynamic range

is 4 zones). Right: the result of applying the operator given by
Equation 4. Simple operator Dodging−and−burning

Figure 8: The simple operator of Equation 3 brings out sufficient

This serves a similar purpose to the local adaptation methods of the
detail in the top image (dynamic range is 6 zones), although ap-
perceptually-driven tone mapping operators [Pattanaik et al. 1998;
plying dodging-and-burning does not introduce artifacts. For the
Tumblin et al. 1999].
bottom image (dynamic range is 15 zones) dodging-and-burning is
Dodging-and-burning is typically applied over an entire region
required to make the book’s text visible.
bounded by large contrasts. For example, a local region might cor-
respond to a single dark tree on a light background [Adams 1983].
The size of a local region is estimated using a measure of local
contrast, which is computed at multiple spatial scales [Peli 1990]. where center V1 and surround V2 responses are derived from Equa-
Such contrast measures frequently use a center-surround function at tions 5 and 6. This constitutes a standard difference of Gaussians
each spatial scale, often implemented by subtracting two Gaussian approach, normalized by 2φ a/s2 + V1 for reasons explained below.
blurred images. A variety of such functions have been proposed, in- The free parameters a and φ are the key value and a sharpening
cluding [Land and McCann 1971; Marr and Hildreth 1980; Blom- parameter respectively.
maert and Martens 1990; Peli 1990; Jernigan and McLean 1992; For computational convenience, we set the center size of the next
Gove et al. 1995; Pessoa et al. 1995] and [Hansen et al. 2000]. After higher scale to be the same as the surround of the current scale. Our
testing many of these variants, we chose a center-surround function choice of center-surround ratio is 1.6, which results in a difference
derived from Blommaert’s model for brightness perception [Blom- of Gaussians model that closely resembles a Laplacian of Gaussian
maert and Martens 1990] because it performed the best in our tests. filter [Marr 1982]. From our experiments, this ratio appears to pro-
This function is constructed using circularly symmetric Gaussian duce slightly better results over a wide range of images than other
profiles of the form: choices of center-surround ratio. However, this ratio can be altered
by a small amount to optimize the center-surround mechanism for
 
1 x2 + y 2 specific images.
Ri (x, y, s) = exp − . (5) Equation 7 is computed for the sole purpose of establishing a
π(αi s)2 (αi s)2
measure of locality for each pixel, which amounts to finding a scale
These profiles operate at different scales s and at different image sm of appropriate size. This scale may be different for each pixel,
positions (x, y). Analyzing an image using such profiles amounts and the procedure for its selection is the key to the success of our
to convolving the image with these Gaussians, resulting in a re- dodging-and-burning technique. It is also a deviation from the orig-
sponse Vi as function of image location, scale and luminance dis- inal Blommaert model [Blommaert and Martens 1990]. The area to
tribution L: be considered local is in principle the largest area around a given
pixel where no large contrast changes occur. To compute the size
Vi (x, y, s) = L(x, y) ⊗ Ri (x, y, s). (6) of this area, Equation 7 is evaluated at different scales s. Note that
V1 (x, y, s) provides a local average of the luminance around (x, y)
This convolution can be computed directly in the spatial domain,
roughly in a disc of radius s. The same is true for V2 (x, y, s) al-
or for improved efficiency can be evaluated by multiplication in the
though it operates over a larger area at the same scale s. The val-
Fourier domain. The smallest Gaussian profile will be only slightly
ues of V1 and V2 are expected to be very similar in areas of small
larger than one pixel and therefore the accuracy with which the
luminance gradients, but will differ in high contrast regions. To
above equation is evaluated, is important. We perform the integra-
choose the largest neighborhood around a pixel with fairly even lu-
tion in terms of the error function to gain a high enough accuracy
minances, we threshold V to select the corresponding scale sm .
without having to resort to super-sampling.
Starting at the lowest scale, we seek the first scale sm where:
The center-surround function we use is defined by:
V1 (x, y, s) − V2 (x, y, s) |V (x, y, sm )| <  (8)
V (x, y, s) = (7)
2φ a/s2 + V1 (x, y, s) is true. Here  is the threshold. The V1 in the denominator of Equa-

270
 Center sm too large causes dark rings to form around bright areas (s3 in
Scale too small (s 1) the same figure), while choosing the scale as outlined above causes
Surround the right amount of detail and contrast enhancement without intro-
ducing unwanted artifacts (s2 in Figure 9).
Center Using a larger scale sm tends to increase contrast and enhance
Surround Right scale (s 2) edges. The value of the threshold  in Equation 8, as well as the
choice of φ in Equation 7, serve as edge enhancement parameters
and work by manipulating the scale that would be chosen for each
pixel. Decreasing  forces the appropriate scale sm to be larger.
Center Increasing φ also tends to select a slightly larger scale sm , but only
Surround at small scales due to the division of φ by s2 . An example of the
effect of varying φ is given in Figure 10.
Scale too large (s 3) A further observation is that because V1 tends to be smaller than
L for very bright pixels, our local operator is not guaranteed to keep
the display luminance Ld below 1. Thus, for extremely bright areas
some burn-out may occur and this is the reason we clip the display
luminance to 1 afterwards. As noted in section 2, a small amount
of burn-out may be desirable to make light sources such as the sun
look very bright.
Scale too small (s 1) Right scale (s 2) Scale too large (s 3)
In summary, by automatically selecting an appropriate neigh-
borhood for each pixel we effectively implement a pixel-by-pixel
dodging and burning technique as applied in photography [Adams
1983]. These techniques locally change the exposure of a film, and
so darken or brighten certain areas in the final print.

4 Results
Radiance map courtesy of Paul Debevec

We implemented our algorithm in C++ and obtained the luminance

values from the input R, G and B triplets with L = 0.27R +
0.67G + 0.06B. The convolutions of Equation 5 were computed
using a Fast Fourier Transform (FFT). Because Gaussians are sepa-
rable, these convolutions can also be efficiently computed in image
space. This is easier to implement than an FFT, but it is somewhat
slower for large images. Because of the normalization by V1 , our
method is insensitive to edge artifacts normally associated with the
computation of an FFT.
Figure 9: An example of scale selection. The top image shows cen- The key value setting is determined on a per image basis, while
ter and surround at different sizes. The lower images show the re- unless noted otherwise, the parameter φ is set to 8.0 for all the im-
sults of particular choices of scale selection. If scales are chosen ages in this paper. Our new local operator uses Gaussian profiles
too small, detail is lost. On the other hand, if scales are chosen too s at 8 discrete scales increasing with a factor of 1.6 from 1 pixel
large, dark rings around luminance steps will form. wide to 43 pixels wide. For practical purposes we would like the
Gaussian profile at the smallest scale to have 2 standard deviations
overlap with 1√ pixel. This is achieved by setting the scaling param-
tion 7 makes thresholding V independent of absolute luminance eter α1 to 1/2 2 ≈ 0.35. The parameter α2 is 1.6 times as large.
level, while the 2φ a/s2 term prevents V from becoming too large The threshold  used for scale selection was set to 0.05.
when V approaches zero. We use images with a variety of dynamic ranges as indicated
Given a judiciously chosen scale for a given pixel, we observe throughout this section. Note that we are using the photographic
that V1 (x, y, sm ) may serve as a local average for that pixel. Hence, definition of dynamic range as presented in Section 2. This results
the global tone reproduction operator of Equation 3 can be con- in somewhat lower ranges than would be obtained if a conventional
verted into a local operator by replacing L with V1 in the denomi- computer graphics measure of dynamic range were used. However,
nator: we believe the photographic definition is more predictive of how
challenging the tone reproduction of a given image is.
L(x, y) In the absence of well-tested quantitative methods to compare
Ld (x, y) = (9)
1 + V1 (x, y, sm (x, y)) tone mapping operators, we compare our results to a representative
set of tone reproduction techniques for digital images. In this sec-
This function constitutes our local dodging-and-burning operator. tion we briefly introduce each of the operators and show images of
The luminance of a dark pixel in a relatively bright region will sat- them in the next section. Specifically, we compare our new operator
isfy L < V1 , so this operator will decrease the display luminance of Equation 9 with the following.
Ld , thereby increasing the contrast at that pixel. This is akin to pho- Stockham’s homomorphic filtering Using the observation that
tographic “dodging”. Similarly, a pixel in a relatively dark region lighting variation occurs mainly in low frequencies and hu-
will be compressed less, and is thus “burned”. In either case the mans are more aware of albedo variations, this method op-
pixel’s contrast relative to the surrounding area is increased. For erates by downplaying low frequencies and enhancing high
this reason, the above scale selection method is of crucial impor- frequencies [Oppenheim et al. 1968; Stockham 1972].
tance, as illustrated in the example of Figure 9. If sm is too small,
then V1 is close to the luminance L and the local operator reduces Tumblin-Rushmeier’s brightness matching operator . A model
to our global operator (s1 in Figure 9). On the other hand, choosing of brightness perception is used to drive this global operator.

271
 Radiance map courtesy of Paul Debevec
φ=1 φ = 10 φ = 15

Figure 10: The free parameter φ in Equation 7 controls sharpening.

We use the 1999 formulation [Tumblin et al. 1999] as we have

found it produces much better subjective results to the earlier
versions [Tumblin and Rushmeier 1991; Tumblin and Rush-
meier 1993].

Chiu’s local scaling A linear scaling that varies continuously is

used to preserve local contrast with heuristic dodging-and-
burning used to avoid burn-out [Chiu et al. 1993].

Ward’s contrast scale factor A global multiplier is used that aims

to maintain visibility thresholds [Ward 1994].

Ferwerda’s adaptation model This operator alters contrast, color

saturation and spatial frequency content based on psy-

Radiance map and top image courtesy of Cornell Program of Computer Graphics
chophysical data [Ferwerda et al. 1996]. We have used the Pattanaik
photopic portion of their algorithm.

Ward’s histogram adjustment method This method uses an im-

age’s histogram to implicitly segment the image so that sep-
arate scaling algorithms can be used in different luminance
zones. Visibility thresholds drive the processing [Ward et al.
1997]. The model incorporates human contrast and color sen-
sitivity, glare and spatial acuity, although for a fair comparison
we did not use these features.

Schlick’s rational sigmoid This is a family of simple and fast

methods using rational sigmoid curves and a set of tunable
parameters [Schlick 1994].

Pattanaik’s local adaptation model Both threshold and supra-

threshold vision is considered in this multi-scale model of lo- New operator
cal adaptation [Pattanaik et al. 1998]. Chromatic adaptation is
also included.
Figure 13: Desk image (dynamic range is 15 zones).
Note that the goals of most of these operators are different from our
goal of producing a subjectively satisfactory image. However, we
compare their results with ours because all of the above methods
do produce subjectively pleasing images for many inputs. There Next, we do not compare with the “layering” method because it re-
are comparisons possible with many other techniques that are out- quires albedo information in addition to luminances [Tumblin et al.
side the scope of this evaluation. In particular, we do not compare 1999]. Finally, we consider some work to be visualization methods
our results with the first perceptually-driven works [Miller et al. for digital images rather than true tone mapping operators. These
1984; Upstill 1985] because they are not widely used in graphics are the LCIS filter which consciously allows visible artifacts in ex-
and are similar to works we do compare with [Ward 1994; Ferw- change for visualizing detail [Tumblin and Turk 1999], the mouse-
erda et al. 1996; Tumblin et al. 1999]. We also do not compare with driven foveal adaptation method [Tumblin et al. 1999] and Pardo’s
the multiscale-Retinex work because it is reminiscent of Pattanaik’s multi-image visualization technique [Pardo and Sapiro 2001].
local adaptation model, while being aimed at much lower contrast The format in which we compare the various methods is a
reductions of about 5:1 [Rahman et al. 1996]. Holm has a com- “knock-out race” using progressively more difficult images. We
plete implementation of the Zone System for digital cameras [Holm take this approach to avoid an extremely large number of images. In
1996], but his contrast reduction is also too low for our purposes. Figure 11 eight different tone mapping operators are shown side by

272
 Ward’s contrast scale factor Chiu Schlick Stockham

Ferwerda Tumblin−Rushmeier Ward’s hist. adj. New operator

Rendering by Peter Shirley

Figure 11: Cornell box high dynamic range images including close-ups of the light sources. The dynamic range of this image is 12 zones.

Schlick Stockham Ferwerda

Tumblin−Rushmeier Ward’s hist. adj. New operator

Radiance map courtesy of Paul Debevec

Figure 12: Nave image with a dynamic range of 12 zones.

273
side using the Cornell box high dynamic range image as input. The Accurate implementation Spline approximation
model is slightly different from the original Cornell box because we
have placed a smaller light source underneath the ceiling of the box
so that the ceiling receives a large quantity of direct illumination, a
characteristic of many architectural environments. This image has
little high frequency content and it is therefore easy to spot any
deficiencies in the tone mapping operators we have applied. In this
and the following figures, the operators are ordered roughly by their
ability to bring the image within dynamic range. Using the Cornell
Radiance map courtesy of Cornell Program of Computer Graphics
box image (Figure 11), we eliminate those operators that darken the
image too much and therefore we do not include the contrast based
scaling factor and Chiu’s algorithm in further tests. Figure 14: Compare the spline based local operator (right) with the
more accurate local operator (left). The spline approach exhibits
Similar to the Cornell box image is the Nave photograph, al-
some blocky artifacts on the table, although this is masked in the
though this is a low-key image and the stained glass windows con-
rest of the image.
tain high frequency detail. From a photographic point of view, good
tone mapping operators would show detail in the dark areas while
still allowing the windows to be admired. The histogram adjust- Algorithm Preprocessing Tone Mapping Total
ment algorithm achieves both goals, although halo-like artifacts are Image size: 512 × 512
introduced around the bright window. Both the Tumblin-Rushmeier Local 1.23 0.08 1.31
model and Ferwerda’s visibility matching method fail to bring the Spline 0.25 0.11 0.36
church window within displayable range. The same is true for Global 0.02 0.03 0.05
Stockham style filtering and Schlick’s method. Image size: 1024 × 1024
The most difficult image to bring within displayable range is Local 5.24 0.33 5.57
presented in Figures 1 and 13. Due to its large dynamic range, Spline 1.00 0.47 1.47
it presents problems for most tone reproduction operators. This im- Global 0.70 0.11 0.18
age was first used for Pattanaik’s local adaptation model [Pattanaik
et al. 1998]. Because his operator includes color correction as well Table 1: Timing in seconds for our global (Equation 3) and local
as dynamic range reduction, we have additionally color corrected (Equation 9) operators. The middle rows show the timing for the
our tone-mapped image (Figure 13) using the method presented approximated Gaussian convolution using a multiscale spline ap-
in [Reinhard et al. 2001]. Pattanaik’s local adaptation operator pro- proach [Burt and Adelson 1983].
duces visible artifacts around the light source in the desk image,
while the new operator does not.
The efficiency of both our new global (Equation 3, without that most of the images in this figure present serious challenges to
dodging-and-burning) and local tone mapping operators (Equa- other tonemapping operators. Interestingly, the area around the sun
tion 9) is high. Timings obtained on a 1.8 GHz Pentium 4 PC in the rendering of the landscape is problematic for any method
are given in Table 1 for two different image sizes. While we have that attempts to bring the maximum scene luminance within a dis-
not counted any disk I/O, the timings for preprocessing as well as playable range without clamping. This is not the case for our oper-
the main tone mapping algorithm are presented. The preprocessing ator because it only brings textured regions within range, which is
for the local operator (Equation 9) consists of the mapping of the relatively simple because, excluding the sun, this scene only has a
log average luminance to the key value, as well as all FFT calcu- small range of luminances. A similar observation can be made for
lations. The total time for a 5122 image is 1.31 seconds for the the image of the lamp on the table and the image with the streetlight
local operator, which is close to interactive, while our global oper- behind the tree.
ator (Equation 3) performs at a rate of 20 frames per second, which
we consider real-time. Computation times for the 10242 images is
around 4 times slower, which is according to expectation.
5 Summary
We have also experimented with a fast approximation of Photographers aim to compress the dynamic range of a scene in a
the Gaussian convolution using a multiscale spline based ap- manner that creates a pleasing image. We have developed a rela-
proach [Burt and Adelson 1983], which was first used in the con- tively simple and fast tone reproduction algorithm for digital im-
text of tone reproduction by [Tumblin et al. 1999], and have found ages that borrows from 150 years of photographic experience. It
that the computation is about 3.7 times faster than our Fourier do- is designed to follow their practices and is thus well-suited for ap-
main implementation. This improved performance comes at the plications where creating subjectively satisfactory and essentially
cost of some small artifacts introduced by the approximation, which artifact-free images is the desired goal.
can be successfully masked by the high frequency content of the
photographs. If high frequencies are absent, some blocky artifacts
become visible, as can be seen in Figure 14. On the other hand, Acknowledgments
just like its FFT based counter-part, this approximation manages to
bring out the detail of the writing on the open book in this figure Many researchers have made their high dynamic range images
as opposed to our global operator of Equation 3 (compare with the and/or their tone mapping software available, and without that help
left image of Figure 8). As such, the local FFT based implementa- our comparisons would have been impossible. This work was sup-
tion, the local spline based approximation and the global operator ported by NSF grants 89-20219, 95-23483, 97-96136, 97-31859,
provide a useful trade-off between performance and quality, allow- 98-18344, 99-77218, 99-78099, EIA-8920219 and by the DOE
ing any user to select the best operator given a specified maximum AVTC/VIEWS.
run-time.
Finally, to demonstrate that our method works well on a broad
range of high dynamic range images, Figure 15 shows a selection
of tone-mapped images using our new operator. It should be noted

274
 10 zones 7 zones

9 zones 11 zones 5 zones

Radiance maps courtesy of Paul Debevec

5 zones 4 zones
Renderings by Peter Shirley

9 zones 8 zones
Radiance map courtesy of Greg Ward Radiance map courtesy of Cornell Program of Computer Graphics

4 zones 7 zones 11 zones

Photograph courtesy of Franz and Ineke Reinhard Radiance map courtesy of Jack Tumblin, Northwestern University Radiance map courtesy of Greg Ward

Figure 15: A selection of high and low dynamic range images tone-mapped using our new operator. The labels in the figure indicate the
dynamic ranges of the input data.

275
References O PPENHEIM , A. V., S CHAFER , R., AND S TOCKHAM , T. 1968. Nonlinear filtering
of multiplied and convolved signals. Proceedings of the IEEE 56, 8, 1264–1291.
A DAMS , A. 1980. The camera. The Ansel Adams Photography series. Little, Brown
and Company. PARDO , A., AND S APIRO , G. 2001. Visualization of high dynamic range images.
Tech. Rep. 1753, Institute for Mathematics and its Applications, University of Min-
A DAMS , A. 1981. The negative. The Ansel Adams Photography series. Little, Brown nesota.
and Company.
PATTANAIK , S. N., F ERWERDA , J. A., FAIRCHILD , M. D., AND G REENBERG , D. P.
A DAMS , A. 1983. The print. The Ansel Adams Photography series. Little, Brown and 1998. A multiscale model of adaptation and spatial vision for realistic image dis-
Company. play. In SIGGRAPH 98 Conference Proceedings, Addison Wesley, M. Cohen, Ed.,
B LOMMAERT, F. J. J., AND M ARTENS , J.-B. 1990. An object-oriented model for Annual Conference Series, ACM SIGGRAPH, 287–298.
brightness perception. Spatial Vision 5, 1, 15–41. PATTANAIK , S. N., T UMBLIN , J., Y EE , H., , AND G REENBERG , D. P. 2000. Time-
B URT, P. J., AND A DELSON , E. H. 1983. A multiresolution spline with application dependent visual adaptation for fast realistic display. In SIGGRAPH 2000 Con-
to image mosaics. ACM Transactions on Graphics 2, 4, 217–236. ference Proceedings, Addison Wesley, K. Akeley, Ed., Annual Conference Series,
ACM SIGGRAPH, 47–54.
C HIU , K., H ERF, M., S HIRLEY, P., S WAMY, S., WANG , C., AND Z IMMERMAN ,
K. 1993. Spatially nonuniform scaling functions for high contrast images. In P ELI , E. 1990. Contrast in complex images. J. Opt. Soc. Am. A 7, 10 (October),
Proceedings of Graphics Interface ’93, 245–253. 2032–2040.

C OHEN , J., T CHOU , C., H AWKINS , T., AND D EBEVEC , P. 2001. Real-Time high P ESSOA , L., M INGOLLA , E., AND N EUMANN , H. 1995. A contrast- and luminance-
dynamic range texture mapping. In Rendering techniques 2001, S. J. Gortler and driven multiscale network model of brightness perception. Vision Research 35, 15,
K. Myszkowski, Eds., 313–320. 2201–2223.

D EBEVEC , P. E., AND M ALIK , J. 1997. Recovering high dynamic range radiance R AHMAN , Z., J OBSON , D. J., AND W OODELL , G. A. 1996. A multiscale retinex
maps from photographs. In SIGGRAPH 97 Conference Proceedings, Addison Wes- for color rendition and dynamic range compression. In SPIE Proceedings: Appli-
ley, T. Whitted, Ed., Annual Conference Series, ACM SIGGRAPH, 369–378. cations of Digital Image Processing XIX, vol. 2847.

D URAND , F., AND D ORSEY, J. 2000. Interactive tone mapping. In Eurographics R AHMAN , Z., W OODELL , G. A., AND J OBSON , D. J. 1997. A comparison of
Workshop on Rendering, 219–230. the multiscale retinex with other image enhancement techniques. In IS&T’s 50th
Annual Conference: A Celebration of All Imaging, vol. 50, 426–431.
FAIRCHILD , M. D. 1998. Color appearance models. Addison-Wesley, Reading, MA.
R EINHARD , E., A SHIKHMIN , M., G OOCH , B., AND S HIRLEY, P. 2001. Color
F ERWERDA , J. A., PATTANAIK , S., S HIRLEY, P., AND G REENBERG , D. P. 1996. A transfer between images. IEEE Computer Graphics and Applications 21 (Septem-
model of visual adaptation for realistic image synthesis. In SIGGRAPH 96 Confer- ber/October), 34–41.
ence Proceedings, Addison Wesley, H. Rushmeier, Ed., Annual Conference Series,
ACM SIGGRAPH, 249–258. S CHEEL , A., S TAMMINGER , M., AND S EIDEL , H.-P. 2000. Tone reproduction for
interactive walkthroughs. Computer Graphics Forum 19, 3 (August), 301–312.
G EIGEL , J., AND M USGRAVE , F. K. 1997. A model for simulating the photographic
development process on digital images. In SIGGRAPH 97 Conference Proceedings, S CHLICK , C. 1994. Quantization techniques for the visualization of high dynamic
Addison Wesley, T. Whitted, Ed., Annual Conference Series, ACM SIGGRAPH, range pictures. In Photorealistic Rendering Techniques, Springer-Verlag Berlin
135–142. Heidelberg New York, P. Shirley, G. Sakas, and S. Müller, Eds., 7–20.

G OVE , A., G ROSSBERG , S., AND M INGOLLA , E. 1995. Brightness perception, S TOCKHAM , T. 1972. Image processing in the context of a visual model. Proceedings
illusory contours, and corticogeniculate feedback. Visual Neuroscience 12, 1027– of the IEEE 60, 7, 828–842.
1052. S TROEBEL , L., C OMPTON , J., C URRENT, I., AND Z AKIA , R. 2000. Basic photo-
G RAVES , C. 1997. The zone system for 35mm photographers, second ed. Focal Press. graphic materials and processes, second ed. Focal Press.

H ANSEN , T., BARATOFF , G., AND N EUMANN , H. 2000. A simple cell model T UMBLIN , J., AND RUSHMEIER , H. 1991. Tone reproduction for realistic computer
with dominating opponent inhibition for robust contrast detection. Kognitionswis- generated images. Tech. Rep. GIT-GVU-91-13, Graphics, Visualization, and Use-
senschaft 9, 93–100. ability Center, Georgia Institute of Technology.

H OLM , J. 1996. Photographics tone and colour reproduction goals. In CIE Expert T UMBLIN , J., AND RUSHMEIER , H. 1993. Tone reproduction for computer generated
Symposium ’96 on Colour Standards for Image Technology, 51–56. images. IEEE Computer Graphics and Applications 13, 6 (November), 42–48.

J ERNIGAN , M. E., AND M C L EAN , G. F. 1992. Lateral inhibition and image process- T UMBLIN , J., AND T URK , G. 1999. LCIS: A boundary hierarchy for detail-preserving
ing. In Non-linear vision: determination of neural receptive fields, function, and contrast reduction. In Siggraph 1999, Computer Graphics Proceedings, Addison
networks, R. B. Pinter and B. Nabet, Eds. CRC Press, ch. 17, 451–462. Wesley Longman, Los Angeles, A. Rockwood, Ed., Annual Conference Series,
83–90.
J OHNSON , C. 1999. The practical zone system. Focal Press.
T UMBLIN , J., H ODGINS , J. K., AND G UENTER , B. K. 1999. Two methods for
L AND , E. H., AND M C C ANN , J. J. 1971. Lightness and retinex theory. J. Opt. Soc.
display of high contrast images. ACM Transactions on Graphics 18 (1), 56–94.
Am. 63, 1, 1–11.
U PSTILL , S. 1985. The Realistic Presentation of Synthetic Images: Image Processing
L ONDON , B., AND U PTON , J. 1998. Photography, sixth ed. Longman.
in Computer Graphics. PhD thesis, University of California at Berkeley.
M ARR , D., AND H ILDRETH , E. C. 1980. Theory of edge detection. Proceedings of
WARD , G., RUSHMEIER , H., AND P IATKO , C. 1997. A visibility matching tone
the Royal Society of London, B 207, 187–217.
reproduction operator for high dynamic range scenes. IEEE Transactions on Visu-
M ARR , D. 1982. Vision, a computational investigation into the human representation alization and Computer Graphics 3, 4 (December).
and processing of visual information. W H Freeman and Company, San Fransisco.
WARD L ARSON , G., AND S HAKESPEARE , R. A. 1998. Rendering with Radiance.
M ATKOVIC , K., N EUMANN , L., AND P URGATHOFER , W. 1997. A survey of tone Morgan Kaufmann Publishers.
mapping techniques. In 13th Spring Conference on Computer Graphics, W. Straßer,
WARD , G. 1994. A contrast-based scalefactor for luminance display. In Graphics
Ed., 163–170.
Gems IV, P. Heckbert, Ed. Academic Press, Boston, 415–421.
M C NAMARA , A., C HALMERS , A., AND T ROSCIANKO , T. 2000. STAR: Visual
W HITE , M., Z AKIA , R., AND L ORENZ , P. 1984. The new zone system manual.
perception in realistic image synthesis. In Eurographics 2000 STAR reports, Euro-
Morgan & Morgan, Inc.
graphics, Interlaken, Switzerland.
W OODS , J. C. 1993. The zone system craftbook. McGraw Hill.
M C N AMARA , A. 2001. Visual perception in realistic image synthesis. Computer
Graphics Forum 20, 4 (December), 211–224.
M ILLER , N. J., N GAI , P. Y., AND M ILLER , D. D. 1984. The application of computer
graphics in lighting design. Journal of the IES 14 (October), 6–26.
M ITCHELL , E. N. 1984. Photographic Science. John Wiley and Sons, New York.

276