DESIGN OF A CUSTOMIZED PATTERN FOR IMPROVING COLOR
CONSTANCY ACROSS CAMERA AND ILLUMINATION CHANGES
Hazem Wannous
IMS Laboratory, University of Bordeaux, Talence, France
Sylvie Treuillet, Yves Lucas
PRISME Institute, University of Orleans, Orleans, France
Alamin Mansouri, Yvon Voisin
Le2i Laboratory, University of Burgundy, Auxerre, France
Keywords:
Color imaging, Color checker design, Chromatic adaptation.
Abstract:
This paper adresses the problem of color constancy on a large image database acquired with varying digital
cameras and lighting conditions. Automatic white balance control proposed by an available commercial cam-
era is not sufficient to provide reproducible color classification. A device-independent color representation
may be obtained by applying a chromatic adaptation transform, from a calibrated color checker pattern in-
cluded in the field of view. Instead of using the standard Macbeth color checker, we suggest to select judicious
colors to design a customized pattern from contextual information. A comparative study demonstrates that
this approach insures a stronger constancy of the interesting colors before the vision control.
1 INTRODUCTION
The human visual system ensures color constancy,
so that the perceived color of objects remains rela-
tively constant under varying illumination conditions.
This ability, described by the retinex theory, involves
both the eye and the brain (Land, 1977). But no
digital camera has this ability and color coordinates
of pixels are highly depending on acquisition condi-
tions and camera tuning (Barnard and Funt, 2002).
Color constancy generally addresses the illumination
changes, also called white balancing (Barnard et al.,
2002). But, a second problem is caused by chang-
ing the camera. While each digital camera provides a
device-dependent RGB color coordinates system and
embeds a custom color adjustment processing, a more
complete chromatic adaptation transform is required
for minimizing the color differences between sev-
eral cameras. The radiometric response function of
the camera is generally non linear, different for each
color channel and depends on exposure settings such
as aperture, focal length and shutter speed. In addi-
tion, each commercial digital camera embeds some
hidden secret color processing like color demosaic-
ing, white balance adjustment and illumination color
correction automatically applied in JPEG format. So,
some precautions have to be taken with commercial
digital cameras to ensure a consistent color analysis
from JPEG images.
In the case of our medical application (Wannous
et al., 2007; Wannous et al., 2008), images are ac-
quired in a free manner in several care centers with
different types of cameras. In routine, medical staff
have no time and no technical competence for using
RAW format. But, color constancy is the main cue
for a reliable monitoring of tissue classification over
time. If a scene was recorded by a free handled cam-
era under several view points, there would be color
variations between images of the same object. In that
case color constancy may be improved by radiometric
alignment between images.
In some previous works algorithms are proposed
for estimating the radiometric response function from
differently illuminated images taken with the same
camera (Debevec and Jitendra, 1997; Mitsunaga and
Nayar, 1999; Grossberg and Nayar, 2002; Kim and
60
Wannous H., Treuillet S., Lucas Y., Mansouri A. and Voisin Y. (2010).
DESIGN OF A CUSTOMIZED PATTERN FOR IMPROVING COLOR CONSTANCY ACROSS CAMERA AND ILLUMINATION CHANGES.
In Proceedings of the International Conference on Computer Vision Theory and Applications, pages 60-67
DOI: 10.5220/0002835700600067
Copyright
c
SciTePress
Pollefeys, 2004). Then, the color textures of im-
ages may be aligned by the estimated function. Most
of these methods require prior knowledge of expo-
sures on a static scene observed with a fixed camera
(Debevec and Jitendra, 1997; Mitsunaga and Nayar,
1999). Extensions have been proposed for non static
scene (Grossberg and Nayar, 2002) and free move-
ment of the camera or adding vignetting correction
(Kim and Pollefeys, 2008). Such a radiometric align-
ment addresses images acquired with a single digital
camera but is not adapted to a large image dataset ac-
quired over time in several medical centers with dif-
ferent commercial digital cameras.
Indeed, automatic white balancing and color ad-
justments embedded in digital cameras do not ensure
color constancy as the applied algorithms differs from
a trademark to another. Each digital camera provides
a device-dependent RGB color coordinates system.
So, different cameras can exhibit radically different
color responses and can cause significant errors in
scene interpretation (Ilie and Welch, 2005). Further-
more, digital cameras - even of the same type - do
not give consistent response. So, we need to homog-
enize the photos of the data base before the classi-
fier learning stage. A reasonable consistency can be
obtained by pair-wise correlation for modeling trans-
fer function based on image color histogram (Porikli,
2003). But the complexity of this approach increases
quadratically with the number of cameras and may in-
troduce distortions and quantization errors
In this context, a good solution to achieve a color
alignment between images captured with different
cameras is given by introducing a small calibrated
color pattern in the field of view during acquisition.
Then, color values may be converted in a device-
independent coordinate system by estimating the best
adaptation transform that maps the image color mea-
surements to the corresponding target reference co-
ordinates. This on-line calibration process reduces
the color variations due to illumination and camera
changes and consequently ensures the reproducibil-
ity of the automatic segmentation and classification of
the textured color regions. A classical choice for the
color pattern is the 24-sample Macbeth color checker
(Barnard et al., 2002; Haeghen et al., 2000; Ilie and
Welch, 2005; Mansouri et al., 2005). This standard
pattern consists of 24 patches chosen to emulate com-
mon natural colors such as skin, foliage, and sky, in
addition to primary colors and a six step grey scale.
It covers an extensive gamut, adapted to a large range
of images. But a common drawback of a calibration
with such a standard pattern is the decrease in perfor-
mance of constancy when interesting colors are not
represented in the color checker. In fact, many surface
inspection problems are based on a limited palette of
colors. Then, a good strategy for color calibration
would be to minimize the mean square errors in some
judiciously selected areas of the color space, to ensure
stronger constancy of these interesting colors.
In this paper, we propose a methodology for ex-
tracting judicious colors to design a customized pat-
tern by analyzing contextual information in a large
image dataset. This issue deals with the color quanti-
zation problem for automatically extracting dominant
colors in images by some adaptive clustering algo-
rithm (Hsieh and Fan, 2000; Cheng and Yang, 2001;
Sirisathitkul et al., 2004). But here, the aim is not to
classify pixels into their corresponding palette colors,
but to design a specific color chart. So, the domi-
nant colors were selected from a set of images instead
of a single image. Next, a special sorting algorithm
was developed to reduce these colors to a very small
set while saving the best share-out in the color space.
A comparative study between correction results ob-
tained with the standard Macbeth color checker chart
and our customized one proves that the color distance
errors are minimized, to consequently provide a more
robust automatic classification in the area of interest.
The organization of the paper is as follows: in the
next section, we present the proposed methodology
for designing a contextual checker chart. The adopted
correction process for insuring color constancy across
camera and illumination changes is detailed in section
3. In section 4, we present the results of the proposed
approach before to conclude in the last section.
2 DESIGN OF A CUSTOMIZED
COLOR CHECKER
Considering a given surface inspection problem, the
color consistency has to be particularly centered on a
limited palette of interesting colors. In this section,
we propose an automatic method to customize a color
checker pattern using contextual data scene statistics.
The first thing to do is to collect a large dataset
of images under varying illuminations and cameras in
real conditions relative to the aimed application. All
the images have been captured with a small standard
Macbeth color checker pattern included in the field
of view, to be first normalized by applying a standard
color calibration detailed in the next section.
To reduce the combinatorial complexity of the
24-bit color representation, we apply a perceptual
color quantization for automatically extracting dom-
inant colors in this collection of images. The pre-
processing stage includes a Peer group filtering algo-
rithm to remove impulsive noise in the images and to
DESIGN OF A CUSTOMIZED PATTERN FOR IMPROVING COLOR CONSTANCY ACROSS CAMERA AND
ILLUMINATION CHANGES
61
compute a weighting index used after for color quan-
tization (Deng et al., 1999). Homogeneous neighbor-
ing pixels are favored by higher weights. Next, the
generalized-Lloyd algorithm (GLA) is applied in the
perceptive CIELUV space to drastically reduce the
number of colors by clustering. These choices are
justified by the use of a similar pre-processing stage
during color region segmentation (Deng and Manju-
nath, 2001). Among several advanced segmentation
methods, the JSEG algorithm proved to be the most
efficient on our dermatological image database. Nev-
ertheless, this choice can be suited to a large range
of applications as a JSEG method includes a reliable
color quantization stage. The quantization results in
a codebook limited to some dominant colors for each
image (Fig. 1).
(a) (b) (c)
Figure 1: Color quantization by generalized-Lloyd algo-
rithm (GLA): (a) original image, (b) image quantization (c)
color codebook.
We collected 132 colors from a base of 26 images
as the most representative ones for our application
(Fig. 2).
(a) (b)
Figure 2: Codebook of 132 colors selected by GLA on the
image data base compared the 24 Macbeth color patches
(a) corresponding coordinates in the CIELUV chromaticity
plan (b).
As it can be seen on the right part of the Fig. 2,
these dominant colors are not uniformly spread in
the CIELUV chromaticity plan (u*, v*). They form
a stretched cloud of points. A principal component
analysis of these data points shows that the principal
axis is defined by the unit vector [0.97; 0.24]. This
axis of the largest variance (explicated inertia of 0.87)
is highly correlated to the u* axis. This observation
agrees with the examination of the colors of skin
tissues which present hue gradation from pink to
beefy red, yellow and brown. By comparison, the
gamut covered by the Macbeth color checker is
really extensive, but only five patches of the Macbeth
pattern are confined in the area corresponding to the
dominant colors selected by GLA. So, the aim is
to select about twenty judicious colors among the
dominant ones to design a customized pattern. The
proposed selection algorithm is based on two basic
ideas: give greater importance to the colors with the
highest occurrences while preserving the spreading
in the chromaticity plan. The proposed approach for
sorting the most representative colors is the following
iterative algorithm.
Algorithm. Color checker design using E uv dis-
tance between candidate colors.
Input: a data set of candidate colors D={L d,u d,v d,
d = 1,2,...,max} and a predefined K number of repre-
sentative colors to be selected).
Output: a K-color codebook P = {(Lp,up,vp), p =
1,2,...,K}
Method:
1. keep all the candidate colors for u and v values
in range 100
2. compute the E uv distance between all pairs of
the candidate colors
3. sort the candidate colors along the axis of their
increasing distance adjacent in an ordered whole.
Choose the pair, with the smallest distance E uv,
among
i = 1
N
(N i) pairs of N colors as origin of
this axis. Let D j=d
2
(c j,c j 1) be the squared Eu-
clidian distance of adjacent colors in CIELUV space
4. divide this whole of colors into (L < K) consecu-
tive cells where each cell contains the same number
of colors
5. for of each cell, compute the summation of
distances between the adjacent colors dsum i =
j = 1
i
(D j)
6. if dsum i τ collect the median color in each cell
to produce m color codebook, else merge the remain-
ing cells in one and divide the latter in K m cells
where the dsum i in each one equal to (N m)/K
7. collect the median color in each cell to produce
(K m) color codebook.
8. stop
Step 3 of the algorithm guarantee an extended re-
VISAPP 2010 - International Conference on Computer Vision Theory and Applications
62
covery of the workspace (samples), while step 6 takes
into account the distribution of samples (candidate
color) in the uniform perceptual space CIELUV.
We note finally that the splitting threshold τ and
the initial number of cells L have been tuned empiri-
cally in such a way as to obtain K-color codebook.
This algorithm has been applied to the previousset
of 132 candidate colors selected by GLA on the image
data base, with K=22. This choice enables to obtain
a target of similar size than MacBeth. The result is il-
lustrated in Fig.3. Black and white patches are added
to the 22 selected colors to design the customized pat-
tern. The latter has the advantage of proposing a large
palette of representative colors included in chronic
wound images while keeping a compact size. The
two color checkers (the standard Macbeth pattern and
the customized one) have been printed on a calibrated
printer with the coordinates assigned to each patch,
and next used in the following experimental tests.
(a) (b)
Figure 3: Customized color checker and the corresponding
coordinates in the CIELUV chromaticity plan.
3 CORRECTION STRATEGY
Color constancy generally addresses the illumination
changes by so-called white balancing, but this does
not takes into consideration the changing of the cam-
era. While each digital camera provides a device-
dependent RGB color coordinates system and em-
beds a custom color adjustment processing, a more
complete chromatic adaptation transform is required
for minimizing the color differences between several
cameras. The correction strategy is then composed of
two steps.
3.1 White Balancing
The colorimetric data provided by a digital camera
can not be dissociated from the illuminant coordi-
nates. Many algorithms of computer vision have been
developed to model color constancy (McCann, 2004)
under illumination changes. Most of them adopt
retinex models and the assumption of the indepen-
dence of each color channel. The dependence to il-
luminant may then be corrected by a 3x3 linear trans-
form on the RGB coordinates, as follows:
"
R c
G c
B c
#
=
"
R
m
/R
W
0 0
0 G
m
/G
W
0
0 0 B m/B W
#
·
"
R
G
B
#
where (R W, G W, B W) are the white reference co-
ordinates, (R m, G m, B m) the measured coordinates
and (R c, G c, B c) the corrected ones.
White balancing control is realized by digital im-
age processor embedded in cameras. Digital cameras
often propose two modes: automatic or manual white
balancing. The manual mode of cameras requires to
capture an image of a white paper sheet under the am-
bient lighting to store the white source coordinates.
But this is not a convenient use for a routine visit in a
patient room. The automatic white balance control is
based on stored reference coordinates of some stan-
dard illuminants, which are generally considered as
the maximal RGB values present in the image. The
ambient lighting in the patient rooms, composed of
several varying sources, may be indeed far away from
the stored standard illuminants, and because of the
centering, the maximal RGB values in the image may
not be localized on really white object.
For this reason, the channel ratios applied on
the color coordinates of each pixel will be based on
the white patch coordinates (R m, G m, B m) mea-
sured on the checker and the reference ones (R W,
G W, B W) defined during calibration stage under
D65 standard illuminant (CIE, 2008). The above lin-
ear correction allows to obtain white balanced images
whatever the lighting conditions.
3.2 Chromatic Adaptation Transform
In order to ensure the color constancy across mul-
tiple cameras, we apply an on-line calibration pro-
cess based on the known color target included in the
field of view. This calibration allows converting the
measured color values in a device-independent coor-
dinates system (as in sRGB space) by estimating the
following transform Φ with m terms. It can be written
as:
R c
G c
B c
!
= Φ
RGBsRGB
R
G
B
(1)
=
a 1 ·· · a m
b 1 ·· · b m
c 1 · ·· c m
Θ
m
R
G
B
DESIGN OF A CUSTOMIZED PATTERN FOR IMPROVING COLOR CONSTANCY ACROSS CAMERA AND
ILLUMINATION CHANGES
63
Our transformation is based on a polynomial
model. The transformation coefficients are computed
by the closest match between the reference coordi-
nates of the target patches and the measured coor-
dinates in the images. The reference coordinates of
the 24 color patches of the target are provided by
a spectrophotometer (Minolta CS 1000 SPM), under
the standardized D65 illuminant for the 10
field of
view.
The best chromatic adaptation was obtained using
the lower order (m = 3) which is in good agreement
with other works (Haeghen et al., 2000). More gen-
eral polynomial functions have also been tested but
they do not give significant enhancement while they
cause largest distortions and higher computation com-
plexity. Each reference color patch provides 3 equa-
tions, allowing to estimate 3 by N parameters for N
patches. With lower order and 22 patches, the linear
system is over determined and a least mean squares
solution is provided.
Even if the standard calibration with Macbeth sta-
bilizes the segmentation and classification steps, as il-
lustrated in Fig.4, it can be seen that some interesting
colors for medical diagnosis are not optimally cor-
rected. That can be explained by the fact that the mini-
mization of errors is constrained only at some selected
points of the color space represented on the pattern. A
better strategy would be to use a customized checker
by selecting judicious colors to insure the closest con-
stancy of these interesting colors, as presented in the
next section.
4 EXPERIMENTAL RESULTS
In this section, we show the effectiveness of the pro-
posed correction by several experiments. We use two
sets of real images: the first is composed of images
taken under controlled lighting conditions in labora-
tory; the second one includes in vivo images taken
in hospital environment. The Macbeth and the cus-
tomized color charts have been calibrated with the
same spectrophotometer.
The location of the color checker is automatically
detected in the images, and the checker is replaced
in a facial position by applying a robust homographic
transform to remove the perspective effects. Hence
the choice of a specific shape of the color checker tar-
get with four colored balls facilitates automatic detec-
tion. The average RGB coordinates are then extracted
in the squared area centered in each patch.
The color correction is generally applied in sRGB
space for its convergence with the PC-world, but to
make quantitative evaluation, we must use a percep-
(a) (b) (c)
Figure 4: Segmentation of a wound image captured under
automatic white balance control (a) and after color correc-
tion using the calibrated color checker included in the field
of view: (b) Macbeth pattern (c) customized pattern.
tually uniform space such as CIELAB. The individ-
ual inter-sample deviation is computed by the Euclid-
ian distance between the measured coordinates and
the reference ones in CIELAB space for each patch.
Then, we take the average color difference of the 24
patches to measure the differences between the im-
ages and the reference ones.
4.1 Measures under Controlled
Illumination
For comparison purpose, Macbeth and customized
targets have been jointly placed in a lighting test box.
Images has been captured under three different con-
trolled lighting: a cool fluorescent light (Illuminant F
- 4150 K), an incandescent light (Illuminant A - 2856
K), and a daylight (Illuminant D65 - 6500 K). Two
different models of cameras have been tested (canon
EOS 350D, Leica D-Lux 3). The images were stored
in RAW format to avoid embedded custom color pro-
cessing and next to be compared to JPEG compressed
images. The scene has been observed from five dif-
ferent points of view: a centering front view and four
cardinal views with angle about 25
, corresponding to
our clinical protocol for 3D reconstruction (Wannous
et al., 2008). Each view has been repeated 5 times,
i.e. a total of 600 measures (24 patches x 5 views x 5
acquisitions) to compute the average CIEE ab dis-
tance for each camera and each illuminant, before and
after correction. Results are presented in Table 1.
CIELAB color distances are measured simultane-
ously on the Macbeth color checker and on the cus-
tomized pattern before and after correction. The stan-
dard correction’ indicates that the chromatic adapta-
tion is computed from the MacBeth pattern; the ’pro-
posed correction’ corresponds to a chromatic adapta-
tion computed from the customized pattern.
Before correction, we observed very large differ-
ences between the two types of digital cameras under
different illuminants, with an average around 25 units.
A smaller standard deviation has also been observed
on the customized pattern, which may be explained
by a more restricted range of colors.
VISAPP 2010 - International Conference on Computer Vision Theory and Applications
64
Table 1: CIELAB color distances measured on the 24 patches after correction with the standard Macbeth color checker and
the customized designed one under controlled illumination.
CIELAB distances illuminant Canon EOS 350D Leica D-Lux 3
Max Average Std Max Average Std
Macbeth target Cool 39.78 23.35 4.18 51.13 23.11 1.39
before correction A 36.17 26.82 2.19 48.78 21.45 1.59
D65 39.50 28.32 2.21 46.59 24.57 4.44
Customized target Cool 31.40 22.82 3.26 21.08 14.28 0.84
before correction A 32.34 26.21 1.44 23.69 14.12 1.14
D65 36.80 28.93 1.64 45.21 16.52 4.04
Macbeth target Cool 8.45 3.94 0.12 9.89 3.04 0.30
standard correction A 8.97 4.01 0.13 11.05 3.36 0.41
D65 7.57 3.88 0.14 7.89 3.26 0.20
customized target Cool 15.67 5.39 0.24 22.41 5.25 0.58
standard correction A 13.51 5.22 0.29 20.01 5.15 0.54
D65 12.71 5.24 0.33 17.97 6.05 0.35
customized target Cool 5.87 2.76 0.22 5.40 2.00 0.26
proposed correction A 5.46 2.86 0.13 10.15 2.04 0.36
D65 8.90 2.94 0.17 10.37 1.90 0.23
Table 2: CIELAB color distances measured on the 24 patches of the customized color checker, in JPEG and RAW format.
CIELAB distances illuminant JPEG format RAW format
Max Average Std Max Average Std
Cool 25.32 16.83 1.57 21.08 14.28 1.84
before correction A 23.98 17.93 1.00 23.69 14.12 1.14
D65 41.99 21.21 3.18 45.21 16.51 4.09
correction Cool 6.95 2.68 0.34 5.40 2.00 0.26
with A 7.19 2.73 0.44 10.15 2.04 0.36
customized target D65 7.99 2.76 0.30 10.37 1.90 0.23
The correction based on the Macbeth pattern re-
duces the averageCIEE ab by a factor between 5 to
6, reducing the gap to about 3,5-4. But, for the inter-
esting colors on the customized pattern, the color con-
stancy is not sufficient: the differences remain high
with an average around 5-6 and some maximum val-
ues between 12 to 22. Logically, the color constancy
is clearly improved by the proposed correction on the
customized pattern, as the minimization of errors is
constrained on judicious colors confined in the inter-
esting area of the color space.
A second test concerns the influence on the re-
sults of color processing and JPEG compression car-
ried out by the digital camera (see Table 2). The JPEG
compressing format is more convenient for storing a
large image data base, but it implies no control on
the embedded color processing, such as demosaic-
ing or color white balancing specific to each camera
type. However, the test proves that the correction per-
formances are not significantly modified by all these
transformations.
4.2 In Vivo Images
To complete these laboratory experiments, we anal-
ysed in a similar way color shifts on photos taken dur-
ing wound patient examination in several hospitals.
To evaluate the proposed correction in real conditions,
the same tests have been applied to a collection of in
vivo images taken in hospital environment (see Table
3). It confirms that the contextual color correction on
the customized pattern results in a higher degree of
constancy for interesting colors. In last analysis, we
can evaluate the real impact of the color correction at
DESIGN OF A CUSTOMIZED PATTERN FOR IMPROVING COLOR CONSTANCY ACROSS CAMERA AND
ILLUMINATION CHANGES
65
Table 3: CIELAB color distances measured on the 24 patches of the standard Macbeth color checker and the customized one
designed for in vivo images.
CIELAB distances in vivo images (JPEG)
Max Average Std
Macbeth before correction 29.28 15.07 6.56
customized before correction 34.82 20.02 4.05
Macbeth standard correction 9.74 4.24 0.25
customized standard correction 27.88 10.68 1.74
customized proposed correction 8.47 3.37 0.47
the end of the processing chain: the performance of
the classification of the skin tissues is improved and a
better agreement is obtained with the medical expert
(about 15% on classification rate) when including the
correction step (Wannous et al., 2007).
5 CONCLUSIONS
This paper deals with color constancy on a large im-
age database acquired with varying digital cameras
and lighting conditions. Automatic white balance
control proposed by the commercial cameras is not
sufficient to provide reproducible surface inspection
or classification results. The instability in color ren-
dering may be significantly reduced by applying a
color calibration with a color checker pattern included
in the field of view. A device-independent color rep-
resentation may then be obtained by applying a chro-
matic adaptation transform, without the necessity of
re-calibrating the cameras when the lighting condi-
tions change. Instead of using the standard Macbeth
color checker, we suggest to select judicious colors to
design a customized chart using contextual informa-
tion. A comparative study demonstrates that this ap-
proach insures a closest constancy of the interesting
colors for the vision control. We conclude that using
heterogeneous cameras and varying ambient illumi-
nation in patient room is possible for computer aided
diagnosis of skin lesions. While designed for opti-
mum color constancy in cutaneous imaging, the pro-
posed approach to customize a color checker can be
applied to many surface inspection problems requir-
ing a limited palette of colors. Future works will be
centered on the adaptation of the number of patches
inside the color pattern to the quality of the desired
precision.
REFERENCES
Barnard, K. and Funt, B. (2002). Camera characterization
for color research. Color Research and Application,
27(3):153–164.
Barnard, K., Martin, L., Coath, A., and Funt, B. (2002). A
comparison of computational colour constancy algo-
rithms, part ii : Experiments with image data. IEEE
Trans. on Image Processing, 11(9):985–999.
Cheng, S. and Yang, C. (2001). A fast and novel tech-
nique for color quantization using reduction of color
space dimensionality. Pattern Recognition Letters, 22
(8):845–856.
CIE (2008). Colorimetry part 2: Standard illuminants for
colorimetry. Technical report, Joint ISO/CIE Standard
ISO 11664-2:2008(E)/CIE S 014-2/E:2006, CIE Cen-
tral Bureau, Kegelgasse 27, A-1030 Vienna, Austria.
Debevec, P. E. and Jitendra, M. (1997). Recovering high
dynamic range radiance maps from photographs. In
Proc. of the 24th annual conference on Computer
graphics and interactive techniques, pages 369–378.
Deng, Y., Kenney, S., Moore, M., and Manjunath, B. S.
(1999). Peer group filtering and perceptual color im-
age quantization. In IEEE Inter. Symp. on Circ. and
Sys. VLSI (ISCAS’99), volume 4, pages 21–24, Or-
lando, FL.
Deng, Y. and Manjunath, B. S. (2001). Unsupervised
segmentation of colour-texture regions in images and
video. IEEE Trans. on Pattern Analysis and Machine
Intelligence (PAMI ’01), 23:800–810.
Grossberg, M. and Nayar, S. (2002). What can be known
about the radiometric response function from images
? In Lecture Notes in Computer Sciences (ECCV),
volume 2353, pages 393–413.
Haeghen, Y. V., Naeyaert, J. M., Lemahieu, I., and Philips,
W. (2000). An imaging system with calibrated color
image acquisition for use in dermatology. IEEE Trans.
on Medical Imaging, 19(7):722–730.
Hsieh, I.-S. and Fan, K.-C. (2000). An adaptive clustering
algorithm for color quantization. Pattern Recognition
Letters, 21(4):337–346.
Ilie, A. and Welch, G. (2005). Ensuring color consistency
across multiple cameras. In IEEE Int. Conf. on Com-
puter Vision, volume 2, pages 1268–1275.
Kim, S. and Pollefeys, M. (2004). Radiometric self-
alignement of image sequences. In Int. Conf. on Com-
puter Vision and Pattern Recognition, volume 1, pages
645–651.
VISAPP 2010 - International Conference on Computer Vision Theory and Applications
66
Kim, S. and Pollefeys, M. (2008). Robust radiometric cali-
bration and vignetting correction. IEEE Trans. on Pat-
tern Analysis and Machine Intelligence, 30 (4):562–
576.
Land, E. H. (1977). The retinex theory of color vision. Sci-
entific American, pages 108–128.
Mansouri, A., Marzani, F., and Gouton, P. (2005). Develop-
ment of a protocol for ccd calibration: application to
a multispectral imaging system. IJRA, 20(2):pp. 94–
100.
McCann, J. (2004). Mechanism of color constancy. In
IS&T/SID Conf. on Color Imaging, volume 12, pages
29–36.
Mitsunaga, T. and Nayar, S. (1999). Radiometric self-
calibration. In Conf. on Vision and Pattern Recogni-
tion, volume 2,, pages 374–380.
Porikli, F. (2003). Inter-camera color calibration by cross-
correlation model function. In IEEE Int. Conf. on Im-
age Proc, volume 2, pages 133–136.
Sirisathitkul, Y., Auwatanamongkol, S., and Uyyanonvara,
B. (2004). Color image quantization using distances
between adjacent colors along the color axis with
highest color variance. Pattern Recognition Letters,
25(9):1025–1043.
Wannous, H., Lucas, Y., Treuillet, S., and Albouy, B.
(2008). A complete 3d wound assessment tool for ac-
curate tissue classification and measurement. In Proc.
15th IEEE International Conference on Image Pro-
cessing ICIP 2008, pages 2928–2931.
Wannous, H., Treuillet, S., and Lucas, Y. (2007). Su-
pervised tissue classification from color images for a
complete wound assessment tool. In 29th Inter. Conf.
of the IEEE Eng. in Med. and Bio. Soc. EMBS’07,
pages 6031–6034.
DESIGN OF A CUSTOMIZED PATTERN FOR IMPROVING COLOR CONSTANCY ACROSS CAMERA AND
ILLUMINATION CHANGES
67