Professional Documents
Culture Documents
2
School of Electronic and Information Engineering, University of Science and Technology Liaoning, China
EXPERIMENTAL CONDITIONS
A CRT monitor with a 24-bit graphic card was used to
display colour stimuli. Its peak white was set to have the
chromaticity coordinates of CIE Illuminant D65 with a
luminance of 60 cd/m2. The CRT monitor was carefully
characterized using the GOG (gain-offset-gamma)
model.15 The GOG model was used to transform between
FIG. 1. The u’, v’ chromaticities of the 50 colours tristimulus values and RGB digital signals of the display.
selected for investigation. The model was tested using 108 test colours. Colour dif-
and a few cd/m2, vision involves a mixture of rod and ference was calculated between the measured and the
cone activities, which is referred to as mesopic vision.1 It original tristimulus values. A mean DE*ab value of 0.46
requires luminances of at least several cd/m2 for photopic was obtained, indicating a good accuracy of the character-
vision in which only cones are active. ization model.
A model for predicting the colour appearance of related
colours in mesopic, photopic and scotopic vision was pro-
Test Colours
posed by Shin et al.4 This model was built based on their
earlier studies5 in which a haploscopic colour-matching Fifty test colour stimuli were selected to cover a wide
technique was used. The observers were asked to match colour gamut; their chromaticities are shown in Fig. 1.
colours between a test colour chip under various illumi- The SPDs of test colours are shown in Fig. 2.
nance levels and a stimulus presented on a CRT display
under photopic luminance conditions. Another study by
Eloholma and Halonen,6 used a visual task-performance Stimulus Size and Luminance Level
approach. Spectral sensitivity functions were defined for The colours were displayed to cover the full CRT
the luminance range of 0.01 to 10 cd/m2, with particular screen in a darkened room. A piece of black cardboard
reference to night-time driving performance. In another with a hole in the middle was used to mask most of the
study, Kwak7 used related colours with 28 and 108 view- screen. Neutral density filters were used to cover the hole
ing fields under mesopic vision and concluded that it is in the middle to achieving the desired luminance levels.
insufficient to predict colour appearance simply using dif- Figure 3 illustrates the whole viewing field which was
ferent colour matching functions. None of the above stud- seen in a completely dark room. Each colour was meas-
ies was dedicated to the appearance of unrelated colours. ured using a Minolta CS1000 tele-spectroradiometer
New data on unrelated colours is therefore required for (TSR) to obtain its luminance and tristimulus values.
predicting their appearance.
In this investigation,8 only unrelated colours were studied.
Different sizes of stimuli were used in photopic and mesopic
conditions. The results accumulated were used to test some
colour appearance models. Almost all these models, includ-
ing the current CIECAM02 model,9–11 were developed to
consider only related colours.3,12,13 Hunt14 developed a
model to predict the colour appearance for both related and
unrelated colours in different viewing conditions. At a later
stage, he refined the original model to be used for unrelated
colours, known as CAM97u2. However, there has been little
data available to verify the performance of this model.
The aims of this study were: to investigate how the
luminance and size of stimuli affects colour appearance
for unrelated colours under photopic and mesopic vision;
to study how the colour appearance models, CAM97u and
CIECAM02, perform with the visual data; and to develop
a new model for unrelated colours. FIG. 2. The SPD of the 50 colours selected for investigation.
example, airplane pilots look at signal lights which are far X ¼ 6833 pðkÞ ðTðkÞ=100Þ xðkÞ Dk (1)
380
away from them at a small visual angle; therefore, a 0.58
visual angle for stimuli was also used. where p(k): Spectral radiance measurement data W/
To investigate the changes in colour appearance for dif- (srm2nm) T(k): Transmittance of the filter (%)
ferent luminance levels, the same test colours with differ- TðkÞ ¼ 100Ti0 ðkÞ=TW;i
0
ðkÞ (2)
ent stimulus sizes were assessed under four different
luminance levels covering both photopic and mesopic where Ti’(k) is the spectral transmittance of the sample
vision conditions. These levels corresponded to the D65 with filter i and TWi’(k) is the spectral transmittance of
stimulus having luminances of 60, 5, 1, and 0.1 cd/m2, a the peak white with filter i. The tristimulus values Y and
range of 2.77 log units. The range of luminances between Z were computed similarly.
the test colours were 2.0 to 60 cd/m2, 0.27 to 5.09 cd/m2,
0.09 to 1.03 cd/m2, and 0.013 to 0.104 cd/m2 for each of
the four luminance levels. The same batch of test colours Experimental Procedures
was used for all of the phases.
Psychophysical experiments were conducted to obtain
Because the experimental results showed that a large
visual data by a panel of 10 observers (4 females, 6
reduction in colourfulness occurred under low luminance
males) using the magnitude-estimation method.16 Each
levels, especially for smaller stimulus sizes, an additional
colour was assessed in terms of brightness, colourfulness,
two stimulus sizes, 28 and 18, were included to investigate
and hue. No reference white was displayed. The terms
further stimuli at different stimulus sizes under the lowest
lightness and chroma are relative perceptual attributes of
luminance level (0.1 cd/m2). The viewing conditions stud-
colours, and these two attributes do not apply to unrelated
ied were divided into two groups. Group 1 included two
colours, because they do not have a similarly illuminated
stimulus sizes (108 and 0.58) under three luminance levels
(60, 5, and 1 cd/m2) for investigating the changes in col-
our appearance of colours caused by different stimulus TABLE I. Experimental phases in the study.
sizes and different luminance levels. Group 2 included Luminance of Filter Visual angle
four stimulus sizes (108, 28, 18, and 0.58) under a lumi- Phase Name D65 (cd/m2) name of stimuli
nance level of 0.1 cd/m2 to further investigate the colour
Group 1 1 60–0.5 60 0 0.58
appearance of different stimulus sizes under a low lumi- 2 5–0.5 5 1
nance level. 3 1–0.5 1 2
4 60–10 60 0 108
5 5–10 5 1
6 1–10 1 2
Filters Group 2 7 0.1–0.5 0.1 3 0.58
8 0.1–1 18
The four different luminance levels were obtained by 9 0.1–2 28
using neutral density filters over the monitor. 10 0.1–10 108
and 5 there is a tendency for the increased brightness to be Effects of Size of Stimuli
accentuated as the stimuli become of higher luminance.
For the effect of stimulus size, comparisons were made
Figure 9 showed another trend that: for the 0.58 view-
between the appearances of colours having 108 stimulus
ing field data, yellow colours became greener, when lumi-
size with those of the 0.58 stimulus size under four differ-
nance was reduced. This phenomenon is known as the
ent luminance levels. Then comparisons were also carried
Bezold-Brücke Effect.17 Note that this effect did not
out with four stimulus sizes under the luminance level of
occur for 108 viewing field data (Fig. 8).
FIG. 5. Comparisons of the mean brightness visual results between four different luminance-level phases. All samples had
the same size stimuli, 0.58. [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]
0.1 cd/m2 to investigate further the size effect on colour of 5 and 1 cd/m2 (the middle two columns), this tendency
appearance under mesopic vision. is also true but to a smaller extent. For the 0.1 cd/m2
Figure 10 shows the results for the 108 and 0.58 stimu- level (the right hand column), the results for the 0.58
lus sizes at the four different luminance levels (60, 5, 1, stimulus size are about 50% lower for both brightness and
and 0.1 cd/m2 for D65). For luminance level 60 cd/m2 colourfulness (this will be shown more clearly in the top
(the first column), the 108 colours appear brighter and right-hand sections of Figs. 11 and 12). Thus increasing
more colourful than the 0.58 colours. For luminance levels the stimulus size from 0.58 to 108 generally increases both
FIG. 7. Comparisons of the mean colourfulness visual results from four different luminance-level phases. All stimuli were
the same size, 0.58. [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]
brightness and colourfulness, and this is most marked for fovea an area 0.348 is s-cone free. When the very small
mesopic vision (0.1 cd/m2); this could be an important field (0.58) size was used, there is a lack of the s-cones.
factor in recognizing signal lights. The results for hue The above may be speculative, due to the large scatter of
(the bottom row) show that stimulus size has no effect. It the results.
also found that the points fall below the 458 line in Fig. The effects of field sizes of 108, 28, 18, and 0.58, all at
11. Perhaps this is because there are only cones inside the 0.1 cd/m2 for D65 are shown in Figs. 11–13, for bright-
foveola (the central 18 field of the eye); in the central ness, colourfulness, and hue, respectively. In Figs. 11 and
FIG. 9. Comparisons of the mean hue visual results between different luminance-level phases. All samples had the same
size stimuli, 0.58. [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]
FIG. 11. Comparisons of mean brightness visual results between four different stimulus size phases. All samples had the
same luminance level, 0.1 cd/m2. Note that the range is from 0 to 50, instead of from 0 to 300 in Fig. 10. [Color figure can
be viewed in the online issue, which is available at wileyonlinelibrary.com.]
FIG. 12. Comparisons of mean colourfulness visual results between four different stimulus size phases. All samples had
the same luminance level, 0.1 cd/m2. Note that the range is from 0 to 40, instead of from 0 to 100 in Fig. 10. [Color figure
can be viewed in the online issue, which is available at wileyonlinelibrary.com.]
12, most of the points are well above the 458 line. This MODELS’ PERFORMANCE
shows that, at this low (mesopic) luminance level, colours
CAM97u Performance
appear brighter and more colourful when the stimulus size
becomes larger; the effect is greatest when starting with a In CAM97u2, Hunt pointed out that, although unrelated
stimulus size of 0.58. Figure 13 shows that, for hue, most colours are usually perceived in surrounds of luminance
of the points are close to the 458 line, which indicates, very much lower than that of the stimuli, in most practi-
again, that hue is hardly changed when the stimulus size cal situation the adapting luminance, LA, can not be taken
is changed. as zero, because the stimulus being considered provides
FIG. 13. Comparisons of mean hue visual results between four different stimulus size phases. All samples had the same
luminance level, 0.1 cd/m2. [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]
some adapting light. In CAM97u the luminance of the formed badly in predicting colourfulness (an average of
adapting field is calculated in both photopic and scotopic 183 units). This was mainly caused by the large CV val-
forms. The photopic luminance, LA, in cd/m2, is calcu- ues at the high (60 cd/m2) luminance level. Changing the
lated as: stimulus size from 0.58 to 108 resulted in an increase in
gradient for both brightness and colourfulness, especially
LK ¼ L2=3 =200
at 0.1 cd/m2.
where L is the luminance of the sample. Using the mean scaling factors SFm, the average CV
The scotopic luminance of the adapting field (divided values for the 10 phases are 32 and 183 for brightness
by 2.26), LAS/2.26 is calculated as: and colourfulness, respectively. Using the group scaling
factors SFg, the average CV values are 31 and 108 for
LAS =2:26 ¼ ðLS =2:26Þ2=3 =200 brightness and colourfulness, respectively. Thus the
where LS/2.26 is the scotopic luminance of the sample. brightness attribute exhibited almost no change between
LS/2.26 is used instead of L because LS/2.26 ¼ L for the two types of scaling factors used; but for the colour-
the equi-energy stimulus, SE. fulness attribute, smaller mean CV values occurred when
The CV value was again used to indicate the agreement using group scaling factors than when using mean scaling
between visual results and CAM97u predictions. The factors. That means a better model performance for col-
results are summarized in Table III. For comparing the ourfulness might result from using different scaling fac-
model’s brightness and colourfulness predictions with the tors according to whether the luminance level was above
visual results, individual scaling factors were derived for or at 0.1 cd/m2.
each phase using the gradient of a best-fit straight line
that passed through the origin (the ideal black and neutral
CIECAM02 Performance
for brightness and colourfulness respectively).
Table III shows the CV values for brightness (CV-B), As mentioned before, a version especially for unrelated
for colourfulness (CV-M) and for hue (CV-H). Gradient- colours is not present in CIECAM02. To run CIECAM02,
B and Gradient-M were individual scaling factors for certain values are needed that do not exist for unrelated
brightness and colourfulness for each phase, respectively; colours; these values were chosen as follows. The adopted
SFm-B and SFm-M were the mean scaling factors calcu- white was put at Y ¼ 100, x ¼ 1/3, y ¼ 1/3; this is the
lated from the average value of individual gradients for same as the reference white used in CIECAM02, the per-
the 10 phases; CV-B and CV-M were the comparison fect diffuser in the equi-energy illuminant, SE (this results
results in terms of CV values between visual data and in the value of the multiplier of R in the equation for Rc
CAM97u predictions using the mean scaling factors being equal to 1.0, and similarly in the equations for Gc
(SFm). The scaling factors SFg-B and SFg-M were the and Bc, indicating no adaptation, which is appropriate for
average scaling factors which are computed by mean unrelated colours.) The luminance of the adapting field,
value of gradients for each Group; CV’-B and CV’M LA, was taken as one fifth of the luminance of the peak
were the corresponding CV values of comparisons using white for each experimental phase (it was not felt right to
the group scaling factors (SFg). CV-H were calculated set LA at zero because each stimulus would provide some
using the original data without scaling, i.e. 0, 100, 200, level of adapting light). The ratio, Yb/Yw, where Yw is the
300 and 400 corresponding to red, yellow, green and blue luminance factor of the adopted white and Yb that of the
unique hues respectively. background, could not be based on features of the observ-
The results in Table III show that CAM97u predicted ing conditions, because unrelated colours do not have a
the brightness and hue accurately, i.e., the predicted errors background, only a surround; a value for this ratio was
(32 and 14 units, respectively) are smaller than the ob- therefore chosen by trial and error that gave the best pre-
server accuracy between all observers (48 and 15 units dictions of the results of the experimental observations;
respectively, as shown in Table II). However, it per- this value was 0.2. The values used for c, Nc, and F,
were those for a dark surround: c ¼ 0.525, Nc ¼ 0.8, els studied. When further investigations were made in
and F ¼ 0.8. each group using the subgroup scaling factors according
Using the mean scaling factors (SFm), the average CV to different stimulus sizes, it was found that the stimulus
values of the 10 phases are 41 and 99 for brightness and size effect was not very obvious in the Group 1 experi-
colourfulness, respectively. When using group scaling fac- ments, but it had a significant impact in the Group 2
tors (SFg), the average CV values are 31 and 50 for experiments (luminance level at 0.1 cd/m2) for both mod-
brightness and colourfulness, respectively. Both brightness els. In addition, the mean-scale factor for the 0.58 stimu-
and colourfulness attributes showed smaller mean CV val- lus size was smaller than that for the 108 stimulus size for
ues when using group scaling factors compared with those both brightness and colourfulness, which implies that col-
obtained using mean scaling factors. For the colourfulness ours appeared darker and less colourful for a 0.58 size.
attribute in particular there is a large improvement from This tendency was observed in both CAM97u and CIE-
99 to 50 CV values. This means that a better model per- CAM02 models.
formance occurred when using different group scaling In general, for the hue attribute, differences in colour
factors categorized by different luminance levels (above appearance between each model and the visual results
or at 0.1 cd/m2); a similar improvement was also found were small. This indicates that the perceived hue does not
for CAM97u for colourfulness. This confirms that the vis- exhibit significant differences for the different parameters
ual data should be analyzed in groups based on the lumi- investigated. For both colour-appearance models,
nance levels, and this also indicates that the model can be CAM97u and CIECAM02, the predictions to the hue vis-
further improved by adding a parameter related to the ual results were satisfactory, i.e., the predicted errors (14
luminance levels of stimuli. As was the case for and 13 units for CAM97u and CIECAM02, respectively)
CAM97u, changing the stimulus size from 0.58 to 108 are smaller than the observer accuracy (15 units, as shown
resulted in an increase in gradient for both brightness and in Table II). In summary, the results showed that the two
colourfulness especially at 0.1 cd/m2. (Also shown in models gave reasonably satisfactory performance for
Table IV are average colourfulness scaling factors, SFsg, brightness and hue results, in that the errors of predictions
for the subgroups 1, 2, 3 and 4, 5, 6; these will be used are smaller than the typical observer accuracy. CAM97u
in Step 3 of a new model to be described below.) predicted more accurately than CIECAM02 for brightness
data, but much worse for colourfulness. For the hue at-
tribute, the two models perform similarly. This suggests
Comparison of CAM97u and CIECAM02 Performance that an improved model for unrelated colours might be
achieved by combining features of CAM97u and
As described in the previous sections, colour-appear- CIECAM02.
ance models, CAM97u and CIECAM02, were tested
using the unrelated visual data. Comparing their perform-
ance, CAM97u outperforms CIECAM02 for brightness
Unrelated-Colour Modeling
according to the mean-scaled predictions, i.e., the pre-
dicted errors were 32 and 41 units for CAM97u and CIE- Only a few colour-appearance models include a rod
CAM02, respectively. For colourfulness, CIECAM02 out- contribution: Hunt94, CAM97c and CAM97u. In fact, the
performs CAM97u according to the mean-scaled predic- rod contribution modeled in CAM97c and CAM97u is
tions, i.e., the predicted errors were 183 and 99 units for originally based on that of the Hunt94 model which
CAM97u and CIECAM02, respectively. When using the assumes that the achromatic signal consists of cone and
group scaling factors, improvements were made in pre- rod signals. The colour-difference signals, however, are
dicting brightness and colourfulness results for both mod- based on the cone signals only, without any rod contribu-
tion. The new model was developed following the same above or close to 1 cd/m2. Figure 14(b) shows the rela-
approach. tionship between the stimulus sizes and the optimized k
Hunt suggested that the scotopic luminance (LS) di- values for the luminance level at 0.1 cd/m2.
vided by 2.26 is assumed to be equal to the photopic It can be seen that the R2 values are almost unity,
luminance (L) in CAM97u. It is used to calculate the sco- which indicates that the best-fit curves represent the rela-
topic luminance in the new model. Subsequently, the tionships very well between luminance levels and the k
compressed cone signals and rod signals are used to cal- values; and between stimulus sizes and the k values.
culate a cone contribution (Aa) and a rod contribution The brightness correlate (Q) in CAM97u was simplified
(AS) to provide a total the achromatic signal (A). As with and then applied to the new model as shown in the equa-
the cone signal, the scotopic luminance (representing the tion below; the colourfulness, M, is obtained from the
rod signal) was compressed by a power function to com- new-CAM02 model (see Step 4 below).
pute the rod contribution. The exponent (0.42) in the Qnew ¼ A þ Mnew =100 (3)
power function follows that of the dynamic adaptation
function in CIECAM029. The whole achromatic signal A The computation procedure of the new model is as fol-
was calculated as shown below. The cone contribution lows.
(Aa) is obtained from CIECAM02. Step 1: Measure or calculate absolute tristimulus values
for the test colour stimulus corresponding to CIE colour
AS ¼ L0:42
S matching functions (28 or 108).
A ¼ Aa þ kAS
XL ¼ xL=y YL ¼ L ZL ¼ ð1 x yÞL=y
where the constant k is used to determine the ratio
between cone and rod contributions. In other words, the Step 2: Calculate the cone achromatic signal (Aa), col-
proportions of AS are different according to the different ourfulness (M), and hue (Hc) based on the forward CIE-
viewing conditions used. The constant k needs to be CAM02. Since there is no reference white involved with
determined empirically. It was calculated by the unrelated colours, a white point of (Y ¼ 100, x ¼ 1/3, y
‘‘optimal’’ method in Matlab using the visual data, i.e., k ¼ 1/3) is used as the reference white in the calculation
was obtained by minimizing the difference between when CIECAM02 is used. The luminance of the adapting
model predictions and visual data. field is taken as 1/5 of the luminance of the filtered or
It was found that there is a change in k due to lumi- unfiltered D65 white; the surround parameters are set as
nance level. The k values increase with a reduction of those under dark viewing conditions.
luminance levels until reaching 0.1 cd/m2. This can be Step 3: Calculate the scaling factor KM. When the lumi-
explained by the rods becoming more active when lumi- nance level is above or close to 1 cd/m2, KM is obtained
nance level is reduced. As a result, a larger rod contribu- from the sub-group scaling factors for CIECAM02 in Table
tion occurs which is represented by a larger k value at the IV (0.90 for 0.58 and 1.02 for 108. Note that in order to
lower luminance levels until reaching 0.1 cd/m2. At a simplify the model, KM is taken as 1 instead of 1.02 which
luminance level of 0.1 cd/m2, a reduction in k occurs with was obtained from experiment.When L ‡ 1 cd/m2KM ¼ 0.9
smaller stimulus size; this is to be expected because of for 0.58 stimuliKM ¼ 1 for 108 stimuli
the small number of rods in the foveola. When the luminance level is less than 1 cd/m2, KM is
Figure 14(a) shows the relationship between the lumi- obtained from a function of stimulus size based on the
nance levels and optimized k values for luminance levels Fig. 15.
FIG. 17. Comparisons between new model predictions (the mean scaling factor is about 1) and visual colourfulness data
for each phase. [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]
FIG. 18. Comparisons between new model predictions and the visual hue data for each phase.
the points from the 458 line is smallest for hue, followed with a smaller scale than that of the others for legibility.
by brightness, showing that the model is performing well. It can be seen that they have a larger scatter than those of
Although the scatter in the case of colourfulness is larger, the others. This is expected due to larger observer vari-
it is no worse than the scatter in Figs. 6, 7, and 12. ability as found earlier at luminance level of 0.1 cd/m2
The scales of the diagrams in the bottom row of Figs. (see Table II).
16 and 17, the experimentally determined results for Figure 19(a) shows the brightness and colourfulness
brightness and colourfulness, at 0.1 cd/m2, are plotted changes for a red colour of medium saturation (relative to
FIG. 19. (a): The brightness and colourfulness predicted by the new model for a sample varying in luminance level with
28 stimulus size. (b): The brightness and colourfulness predicted by the new model for a sample varying in stimulus size at
0.1 cd/m2 luminance level.