Information Preserving Color Transformation for Protanopia and Deuteranopia (IEEE Signal Processing Letter 2007)

IEEE SIGNAL PROCESSING LETTERS, VOL. 14, NO. 10, OCTOBER 2007 711

Information Preserving Color Transformation
for Protanopia and Deuteranopia
Jia-Bin Huang, Yu-Cheng Tseng, Se-In Wu, and Sheng-Jyh Wang, Member, IEEE

Abstract—In this letter, we proposed a new recoloring method and lack of S-cones is referred to as tritanopia. Among these
for people with protanopic and deuteranopic color deficiencies. We three types of dichromats, protanopia and deuteranopia have
present a color transformation that aims to preserve the color in- difficulty in distinguishing red from green, while tritanopia
formation in the original images while maintaining the recolored
has difficulty in discriminating blue from yellow. So far, many
images as natural as possible. Two error functions are introduced
and combined together to form an objective function using the La- research works have been conducted on simulating color-defi-
grange multiplier with a user-specified parameter . This objective cient vision [2]–[5]. These approaches represent color stimuli
function is then minimized to obtain the optimal settings. Experi- as vectors in the three-dimensional LMS space, where three
mental results show that the proposed method can yield more com- orthogonal axes L, M, and S represent the quantum catch for
prehensible images for color-deficient viewers while maintaining each of the three distinct cone types. Since the dichromatic
the naturalness of the recolored images for standard viewers. vision is the reduced form of trichromatic vision, the lack of
Index Terms—Color deficiency, image processing, Lagrange one cone type can be simulated by collapsing one of the three
multiplier, recoloring. dimensions into a constant value.
To enhance the comprehensibility of images for color-defi-
cient viewers, daltonization is proposed in [6] to recolor images
I. INTRODUCTION
for dichromats. In [6], the authors first increase the red/green
UE to the increasing use of colors in multimedia con-
D tents to convey visual information, it becomes more impor-
tant to perceive colors for information interpretation. However,
contrast in the image and then use the red/green contrast in-
formation to adjust brightness and blue/yellow contrast. In [7],
Ishikawa et al. described the manipulation of webpage colors for
roughly around 5%–8% of men and 0.8% of women have cer- color-deficient viewers. They first decompose a webpage into a
tain kinds of color deficiency. Unlike people with normal color hierarchy of colored regions and determine “important” pairs of
vision, people with color deficiency have difficulties discrimi- colors that are to be modified. An objective function is then de-
nating certain color combinations and color differences. Hence, fined to maintain the distances of these color pairs, as well as to
multimedia contents with rich colors, which can be well dis- minimize the extent of color remapping. This approach is fur-
criminated by people with normal color vision, may sometimes ther extended to deal with full-color images in [8]. On the other
cause misunderstanding to people with anomalous color vision. hand, Seuttgi Ymg et al. [9] proposed a method to modify colors
Humans’ color vision is based on the responses to photons for dichromats and anomalous trichromats. For dichromats, a
in three different types of photoreceptors, which are named monochromatic hue is changed into another hue with less sat-
“cones” and are contained in the retina of human eyes [1]. uration, while for anomalous trichromats, the proposed method
The peak sensitivities of these three distinct cones lie in the tends to keep the original colors. In [10], Rasche et al. use a
long-Wavelength (L), middle-wavelength (M), and short-wave- linear transform to convert colors in the CIELAB color space
length (S) regions of the spectrum. Anomalous trichromacy is and enforce proportional color differences during the remap-
frequently characterized by a shift of one or more cone types ping. Based on the same constraint for color deficiency, the au-
so that the pigments in one type of cone are not sufficiently thors further improve the optimization process by using the ma-
distinct from the pigments in others. For example, L-Cones are jorization method [11].
more like M-Cones in protanomaly and M-Ccones are more Basically, all the aforementioned works may generate im-
like L-Cones in deuteranomaly. On the other hand, dichromats ages that are more comprehensible to color-deficient viewers.
have only two distinct pigments in the cones and entirely lack However, recolored images may look very unnatural to viewers
one of the three cone types. Lack of L-cones is referred to with normal vision. From an application viewpoint, images in a
as protanopia, lack of M-cones is referred to as deuteranopia, public place may be simultaneously observed by normal people
and color-deficient people. For example, in a public transporta-
Manuscript received October 15, 2006; revised February 11, 2007. This work tion system, many advertisements and traffic maps are delivered
was supported by the National Science Council of the Republic of China under in colors. Without concerning the needs of deficient observers,
Grant NSC-94-2219-E-009-008. The associate editor coordinating the review color-deficient people may have difficulty in understanding the
of this manuscript and approving it for publication was Dr. Konstantinos N.
Plataniotis. image contents. On the contrary, if only concerning the needs
The authors are with the Department of Electronics Engineering, National of color-deficient people, then these recolored images may look
Chiao Tung University, Hsin-Chu 30050, Taiwan, R.O.C. (e-mail: mysoul- annoying to normal observers. Hence, in this letter, we aim to
foryou.ee91@nctu.edu.tw).
Color versions of one or more of the figures in this paper are available online
develop a recoloring algorithm that can automatically construct
at http://ieeexplore.ieee.org. a transformation to maintain details for color-deficient viewers
Digital Object Identifier 10.1109/LSP.2007.898333 while preserving naturalness for standard viewers.
1070-9908/$25.00 © 2007 IEEE

712 IEEE SIGNAL PROCESSING LETTERS, VOL. 14, NO. 10, OCTOBER 2007

Fig. 1. Rotation operation in the a –b plane.
Fig. 2. (a) Function () with three parameters: ;
;, and
. (b) Func-
tion + ( ) with parameters ;
; and
for a half plane.
II. COLOR REPRODUCTION FOR PROTANOPIA
AND DEUTERANOPIA
and deuteranopic viewers, ) decreases to zero when ap-
A. Color Reproduction Method proaches . In this letter, we define to be
In this letter, we focus on protanopia and deuteranopia,
which are the major types of color deficiency. In order to
mimic the color perception of protanopia and deuteranopia, we (2)
adopt Brettel’s algorithm [2] to simulate the perceived images.
Here, we adopt CIELAB color space as the working domain. for the right half-plane of the – plane, where ranges from
In both protanopia and deuteranopia, there is strong correlation to . Here, represents the maximal change of
between the original colors and the simulated colors in the the included angle and represents the degree of the decreasing
values of and , while there is a weak correlation between rate. These two parameters will be specified by optimizing an
the original and the perceived . That is, the original color objective function based on the contents of the original color
information in gets lost significantly. To retain the infor- image. For the left half-plane , we define the
mation in , a reasonable way is to do some kind of image function in a similar manner but with different and .
warping so that the information of is mapped onto the This is because in practice, we may want the right half and the
axis in the CIELAB color space. left half of the – plane to have different transformations, as
In our approach, we aim to maintain the color differences of shown in Fig. 2(a). Moreover, since approaches zero when
color pairs in the CIELAB color space while keeping the recol- colors are close to the axis, crossover of colors can be avoided
ored images as natural as possible. To keep the recolored image when crossing the b axis.
natural, three premises are adopted. First, the recolored image In Fig. 2(b), we show the plot of the transformed hue
has the same luminance as the original image. Second, colors versus the original hue for the right-half – plane.
with the same hue in the original image still have the same hue If the is positive, then the quadrant with positive
after recoloring. Third, the saturation of the original colors is will be compressed while the quadrant with negative
not altered after recoloring. In our approach, a rotation oper- will be expanded and vice versa. To avoid
ation is adopted in the – plane to transform the informa- colors crossover in the compressed quadrant, we require
tion of onto the axis, as illustrated in Fig. 1. Here, we
assume some color stimuli have the same in-
cluded angle with respect to the axis. The rotation opera- (3)
tion maps these colors to new colors , which lay
on another line with the included angle . If ignoring By combining (2) and (3), we have
the nonlinear property of the iso-hue curves in the CIELAB
color space [13], this rotation process simultaneously changes
the hue of with the same amount of hue. Hence, (4)
the transformed colors still share the same hue
after color transformation. Moreover, the saturation of the orig- Since ranges from to , the LHS of (4) has the
inal color is also preserved. lower bound . Thus, we can obtain the constraint
In mathematics, this rotation operation can be formulated as . On the other hand, the constraint in (4) is not necessary
a matrix multiplication. That is, we have in the expanded quadrant. Hence, we introduce two parameters
and , one for each quadrant. For the compressed quadrant,
the constraint in (4) is required, while for the expanded region,
(1) no constraint is needed for and . In the proposed algo-
rithm, there would be six parameters in total. Their notations
and meanings are listed in Table I.
where and are the CIELAB values of the
recolored color and the original color, respectively. is a B. Optimization Using Detail and Naturalness Criteria
monotonically decreasing function of . Since the color differ- In this section, we introduce two criteria, one for detail pre-
ence along the axis can be well discriminated by protanopic serving and the other for naturalness preserving. For each color

HUANG et al.: INFORMATION PRESERVING COLOR TRANSFORMATION FOR PROTANOPIA AND DEUTERANOPIA 713

TABLE I
PARAMETERS FOR RECOLORING

Fig. 3. (a) Original image. (b) Recolored by the Daltonization method with a
pair in the original color domain, we first calculate the perceived middle-level correction [6]. (c) Recolored by Rasche’s method [10]. (d) Recol-
color difference with respect to a person with normal vision. ored by our proposed method with = 0:1. (e) Recolored by our proposed
method with = 0. (f)–(j): Corresponding color images perceived by people
Then, for the corresponding color pair in the transformed color with deuteranopic color deficiency.
domain, we calculate the perceived color difference with re-
spect to a person with protanopic or deuteranopic deficiencies.
As mentioned above, we follow Brettel’s algorithm [2] to simu-
late the color perception for protanopia and deuteranopia. In our for normal viewers, while a smaller makes the recolored
criterion, we wish these two perceived color differences to be as image more comprehensible for color-deficient viewers.
similar as possible. Hence, we define an error function to be One more thing to mention is about the nonlinear property
of the iso-hue curves in the CIELAB color space [13]. That is,
two colors with the same included angle in the – plane
may not have the same value of hue. Due to this nonlinear prop-
erty, colors with the same hue in the original image may gen-
erate colors with different hues in the recolored image. To solve
(5) this problem, we may simply apply the hue-linearization process
mentioned in [14] as a preprocessing and then apply the delin-
where and range over the colors contained in the images, earization process after the recoloring algorithm.
is a perceptual color difference metric, is our recoloring
function, and denotes the simulated color perception
using Brettle’s algorithm. By minimizing this error function, we III. EXPERIMENTAL RESULTS
can preserve color details of the original image.
On the other hand, we attempt not to dramatically modify the In Fig. 3, we demonstrate some experimental results for the
color perception of the color images since a severe modification “flower” image. Fig. 3(a)–(e) shows the images perceived by
may make the recolored image extremely unnatural for normal normal viewers, while Fig. 3(f)–(j) presents the images per-
viewers. Hence, we define another error function to be ceived by viewers with deuteranopic deficiency. We can ob-
serve that the color contrast between the red flower and the
green leaves is lost for people with deuteranopic deficiency.
(6) We compare our method with the Daltonization method [6] and
Rasche’s method [10], as shown in Fig. 3(b)–(e) and (g)–(j). We
may observe that even though the Daltonization method with
where ranges over all the colors in the original color image. a middle-level correction may also preserve the naturalness of
Minimizing this error function shortens the color distance the recolored image for normal people, the contrast between the
between the original colors and the corresponding remapped flower and leaves looks very poor for deuteranopic people. On
colors. To preserve both details and naturalness, we combine the other hand, even though Rasche’s method may create great
these two error functions using the Lagrange multiplier with a contrast for deuteranopic people, the naturalness of the recol-
user-specified parameter . Here, we further normalize these ored image is extremely poor for people with normal vision. In
two error functions by their arithmetic means to achieve similar comparison, our method may well preserve both details and nat-
order of magnitude. That is, the total error is written as uralness at the same time.
To verify the effect of , we also demonstrate in Fig. 3(e) that
our proposed method will produce an extremely unnatural re-
(7) colored image if . Furthermore, in Table II, we compare
the naturalness error and detail error among different methods,
To minimize the objective function in (7), we roughly based on (5) and (6). In our approach, the naturalness error de-
estimate and in the initialization stage creases while detail error increases when rises. For the Dal-
with , and fixed to 1. Then we use the tonization method, even though its naturalness error is less than
Fletcher–Reeves conjugate-gradient method with the constraint ours, its detail error becomes extremely high. On the other hand,
in (4) to obtain the optimal solution. By choosing different even though Rasche’s method has a smaller detail error, its natu-
values of , users may adjust the tradeoff between details and ralness error is larger. These experimental results show that both
naturalness. A larger makes the recolored image more natural naturalness and detail can be properly preserved by our method.

714 IEEE SIGNAL PROCESSING LETTERS, VOL. 14, NO. 10, OCTOBER 2007

TABLE II
COMPARISON OF NATURALNESS ERROR AND DETAIL ERROR

Fig. 5. Six images for the subjective evaluation.

Fig. 6. Experimental results. (a) Scales from the “naturalness” experiment. (b)
Scales from the “comprehensibility” experiment.

IV. CONCLUSION
We have presented in this letter a new recoloring method for
people with protanopic or deuteranopic deficiency. We propose
a color transformation that can yield more comprehensible im-
ages for protanopic or deuteranopic viewers while maintaining
Fig. 4. (a) Original image. (b) Perceived image by protanopic viewer. (c) Per- the naturalness of the recolored images for standard viewers.
ceived image by deuteranopic viewer. (d) Recolored image for protanopia. (e) The same procedure can be extended to the case of tritanopia,
Perceived image of (d) by protanopic viewers. (f) Recolored image for deuteran-
gopia. (g) Perceived image of (f) by deuteranopic viewers.
in which blue and yellow tones cannot be well distinguished.
The experimental results show that our proposed method per-
forms subjectively better than others, in terms of comprehensi-
bility and naturalness.
In Fig. 4, we show more examples to verify the effectiveness of
the proposed method. REFERENCES
We also used Thurstone’s Law of Comparative Judgment [12] [1] Wandell, Foundations of Vision. Sunderland, MA: Sinauer, 1995.
for subjective evaluation. In our subjective experiments, ten par- [2] H. Brettel, F. Vi’enot, and J. Mollon, “Computerized simulation of
color appearance for dichromats,” J. Optic. Soc. Amer. A, vol. 14, no.
ticipants with normal vision were involved and six represen- 10, pp. 2647–2655, Oct. 1997.
tative color images were chosen, as shown in Fig. 5. All ten [3] G. Meyer and D. Greenberg, “Color-defective vision and computer
participants were graduate students with some background in graphics displays,” IEEE Comput. Graph. Appl., vol. 8, no. 5, pp.
video coding and image processing. Since we have difficulty in 28–40, Sep. 1988.
[4] S. Kondo, “A computer simulation of anomalous color vision,” in Color
finding color-deficient viewers, we adopted Brettel’s algorithm Vision Deficiencies. Amsterdam, The Netherlands: Kugler Ghe-
[2] to mimic the perception of protanopia and deuteranopia. In dini, 1990, pp. 145–159.
the first experiment, each of the six images was, respectively, re- [5] J. Walraven and J. W. Alferdinck, “Color displays for the color blind,”
in Proc. IST/SID 5th Color Imaging Conf., 1997, pp. 17–22.
colored by the Daltonization method, Rasche’s method, and our [6] R. Dougherty and A. Wade, Daltonize. [Online]. Available:
method with . For each image, the original image was http://www.vischeck.com/daltonize/.
first shown to the participants. Then, exhaustive paired compar- [7] M. Ichikawa, K. Tanaka, S. Kondo, K. Hiroshima, K. Ichikawa, S.
isons were performed over the recolored images, and the partic- Tanabe, and K. Fukami, “Web-page color modification for barrier-free
color vision with genetic algorithm,” Lecture Notes Comput. Sci., vol.
ipants were asked to choose the more natural image from each 2724, pp. 2134–2146, 2003.
pair. This experiment is to evaluate the naturalness of the re- [8] M. Ichikawa, K. Tanaka, S. Kondo, K. Hiroshima, K. Ichikawa, S.
colored images from the viewpoint of normal viewers. In the Tanabe, and K. Fukami, “Preliminary study on color modification for
still images to realize barrier-free color vision,” in Proc. IEEE Int. Conf.
second experiment, Brettel’s algorithm was applied over the Systems, Man, Cybernetics, 2004, pp. 36–41.
original images and recolored images to simulate the perceived [9] S. Yang and Y. M. Ro, “Visual contents adaptation for color vision
images for deuteranopia. Exhaustive paired comparisons were deficiency,” in Proc. IEEE Int. Conf. Image Process., Sep. 2003, vol. 1,
performed again over the simulated images, and the participants pp. 453–456.
[10] K. Rasche, R. Geist, and J. Westall, “Detail preserving reproduction
were asked to choose the more comprehensible image from each of color images for monochromats and dichromats,” IEEE Comput.
pair. This experiment is to evaluate the comprehensibility of the Graph. Appl., vol. 25, no. 3, pp. 22–30, May–Jun. 2005.
recolored images from the viewpoint of deuteranopic viewers. [11] K. Rasche, R. Geist, and J. Westall, “Re-coloring images for gamuts of
lower dimension,” EuroGraphics, vol. 24, no. 3, pp. 423–432, 2005.
The results of these two subjective experiments were analyzed [12] W. S. Torgerson, Theory and Method of Scaling. New York: Wiley,
based on Thurstone’s Law of Comparative Judgment [12]. The 1967.
scaling of data is shown in Fig. 6. Fig. 7(a) indicates that both [13] G. Hoffmann, CIELab Color Space. [Online]. Available: http://www.
our method and the Daltonization method produce more natural fho-emden.de/~hoffmann/cielab03022003.pdf.
[14] G. J. Braun, F. Ebner, and M. D. Fairchild, “Color Gamut mapping
images, while Fig. 7(b) indicates that our method may preserve in a hue-linearized CIELab color space,” in Proc. IST/SID6th Color
more details than the other two methods. Imaging Conf., 1998, pp. 163–168.

Information Preserving Color Transformation for Protanopia and Deuteranopia (IEEE Signal Processing Letter 2007)

More Related Content

What's hot (18)

Viewers also liked (20)

Similar to Information Preserving Color Transformation for Protanopia and Deuteranopia (IEEE Signal Processing Letter 2007) (20)

More from Jia-Bin Huang (20)

Information Preserving Color Transformation for Protanopia and Deuteranopia (IEEE Signal Processing Letter 2007)