Rochester Institute of Technology

RIT Digital Institutional Repository

Theses

7-1-1998

Derivation and modelling hue uniformity and development of the

IPT color space
Fritz Ebner

Follow this and additional works at: https://repository.rit.edu/theses

Recommended Citation
Ebner, Fritz, "Derivation and modelling hue uniformity and development of the IPT color space" (1998).
Thesis. Rochester Institute of Technology. Accessed from

This Thesis is brought to you for free and open access by the RIT Libraries. For more information, please contact
[email protected].
 DERIVATION AND MODELUNG OF HUE
UNIFORMITY

And Development of the IPT Color Space

by

Fritz F. Ebner

B.S. Carnegie Mellon University (1986)

M.S. University of Rochester (1990)

A dissertation submitted in partial fulfillment of the

requirements for the degree of PhD.
in the Chester F. Carlson Center for Imaging Science
of the College of Science
Rochester Institute of Technology

July 1998

Signature of the Author

-------------------------

Accepted by Henry E. Rhody

Coordinator, Ph.D. Degree Program
~Wy
Date
 CHESTER F. CARLSON

CENTER FOR IMAGING SCIENCE

COLLEGE OF SCIENCE

ROCHESTER INSTITUTE OF TECHNOLOGY

ROCHESTER, NEW YORK

CERTIFICATE OF APPROVAL

PhD. DEGREE DISSERTATION

The Ph.D. Degree Dissertation of Fritz F. Ebner

has been examined and approved by the
dissertation committee as satisfactory for the
dissertation requirement for the
PhD. degree in Imaging Science

Dr. Mark D. Fairchild, Thesis Advisor

Dr. Roy S. Berns

Dr. Robert Rolleston

Dr. Kathleen Chen

Date
9/z'I/tff

11
 DISSERTATION RELEASE PERMISSION

ROCHESTER INSTITUTE OF TECHNOLOGY

COLLEGE OF SCIENCE

CHESTER F. CARLSON

CENTER FOR IMAGING SCIENCE

Title of Dissertation: Derivation and modeling of hue uniformity and development of the

1FT color space

I, Fritz F. Ebner , hereby grant permission to the Wallace Memorial Library of R.I.T.

to reproduce my thesis in whole or in part. Any reproduction will not be for commercial use

or profit.

Signature:

Date:

III
 DERIVATION AND MODELLING OF HUE
UNIFORMITY

And Development of the IPT Color Space

by

Fritz F. Ebner

Submitted to the

Chester F. Carlson
Center for Imaging Science
College of Science
in partial fulfillment of the requirements

for the Ph.D. degree

at the Rochester Institute of
Technology

ABSTRACT

Metric color spaces have been determined to be significantly non-uniform in the hue attribute
of color appearance. Several independent sources have confirmed the non-uniformity. A data

set was obtained during the course of this thesis work that contains the largest sampling of
color space to date which can be used to compare models of color appearance. The data set
obtained was compared to existing data sets and found to correspond closely.
Lookup table

methods were employed to test significant differences between data sets. A simple
modeling
approach was taken based on commonly understood color space models and knowledge of the
visual system. Several color spaces be derived using the simple model, and one was
can

chosen that models hue uniformity very well and has other desirable attributes. This new color
space is named IPT. Many visual data sets were plotted in the IPT color space and all show

improved performance over

industry standard color spaces. The IPT color space has
applications in color data representation, gamut mapping, and color appearance modeling.

IV
 ACKNOWLEDGEMENTS

I time from
wish to acknowledge all the observers who selflessly (but not without reward) took
their busy schedules to stare at a CRT screen and make hard decisions. Names are listed in
arbitrary order.

Adam Stein Bob McCarthy Brad Smith

Alex Vaysman Eddy Dalai Judy Stinehour

Bob Karz Gaurav Sharma Tom Lynch

Brian Hawkins Brian Wiley Chrissy Larson

Chris VanCampen Dave Birnbaum Dorothy Jaehn
Dave Odgers Hong Li Rob Rolleston

Jackie Holmes Jorge Cajiga Juliet Ong

Karen Braun Kathy Loj Laura Flick

Laura Caruso Liesl Wiley Martha Birnbaum

Maryjo Ebner Me Dianne Steely

Nagesh Narendranath Nancy Goodman Raja Balasubramanian

Phyllis Giambrone Genia Sychtycz Steve Reczek

Rene Rassmussen Gus Braun Jutka Bolgar

Ron Macera Roger Chan Scott Bennett

Steve Bloomberg Skip Steve Linder

Sue Zoltner \^lad Brawe Candice Dobra

Mojgan Rabanni Jim O'neill Nancy Mazur

Dick Losz Ron Johnson Pat Colclough

Marvbeth Poupart Gloria Wimer Pat Donaldson

Doug Kreckel

I would like to acknowledge the following people for going well beyond the call of
duty (or a

free lunch) bv making many observations, providing valuable feedback on the experimental

design and presentation to the observers, helpful discussions, and general support: Jacqueline
Holmes, Karen Braun , Dorothy Jaehn, Lee Cass, and Dave Birnbaum. I would also like to

acknowledge the
help of Dr. Robert Chapman from University of Rochester's Center for
Visual Sciences, for help with the analysis of semantic differential scale data.
 DEDICATION

I would like to dedicate this work to my sister

Meg with whom I shared the first 23 years of

my life. Although she never knew it, she showed me that it is possible to be courageous in the

face of incredible adversity, and that you can keep your sense of humor through it all.

VI
 Table of Contents

INTRODUCTION 1

BACKGROUND 3

2.1 Literature Search 4

2.1.1 Color Order Systems and Color Harmony 4
2.1.1.1 Munsell 4
2.1.1.2 Ostwald 8
2.1.1.3 NCS 10
2.1.1.4 Moon and Spencer 14
2.1.1.5 Others 16
2.1.2 Color Naming and Color Meaning 17
2.1.2.1 Basic Color Names 18
2.1.2.2 Color Meanings 20
2.1.3 Preferred Reproduction of Pictorial Images 23
2.2 Preliminary Experiments 30
2.2. 1 Experiment PI (not completed). Semantic differential scaling of colors in the context of
business graphics 39
2.2.2 Experiment P2. Gamut mapping through the use of color matching (Gamut mapping from
below) 46
2.2.2.1 Perceptually closest color matches (sub-experiment 1) 46
2.2.2.1.1 Abstract: to P2 (sub-experiment 1) 46
2.2.2.1.2 Introduction to P2 (sub-experiment 1) 47
2.2.2.1.3 Experimental 50
2.2.2.1.4 Conclusions P2 (sub-experiment 1)
to 74
2.2.2.2 Preserving meaning in color matches (sub-experiment 2) 76
2.2.5 Conclusions leading to body of thesis work 80

APPROACH AND RESULTS 81

3.1 Experiment El. Finding surfaces of constant hue in color space 86

3.1.1 Abstract 86
3.1.2 Introduction 86
3.1.3 Experimental 88
3.1.3.1 Script 89
3.1.3.2 User interface 90
3.1.3.3 Viewing conditions 92
3.1.3.4 Calibration 92
3.1.3.5 Color selections 92
3.1.3.6 Observers 93
3.1.4 Results and Discussion 95
3.1.4.1 Data set calculation 95
5.7.5 Conclusion 110
3.2 Development of constant hue color spaces ill
5.2.7 Neural Nets 777
3.2.1.1 General overview Ill
3.2.1.2 Modeling the CIELAB transformation 113
3.2.1.3 Modeling the constant hue data set 114
3.2.2 Color space visualization and parameter tuning 776
3.2.2.1 Model Descriptions 117
3.2.2.2 Visualization tool development 122
3.2.2.2.1 Visualization of color spaces 123
3.2.2.2.2 Controls for parameter tuning 128
3.2.2.2.3 Quantitative metrics for hue uniformity and psychophysical data set visualization 138
3.2.2.3 Properties of the chosen color space (IPT) 142
3.2.2.3.1 Engineering Considerations 142

vu
 3.2.2.3.2 Model and Coefficients 146
3.2.2.3.3 Color Functions 147
Matching
3.2.2.3.4 Constant Hue data sets 149
3.2.2.3.5 Munsell Value 5 data 152
3.2.2.3.6 Neutral Lightness Response 153
3.2.2.3.7 Chromatic Lightness Response 154
3.2.2.3.8 MacAdam (PGN) Discrimination Ellipses and Spectral Locus 159
3.2.2.3.9 Suprathreshold Color Difference Ellipses 160
3.2.2.3.10 OSA Color System medium gray (L = 0) constant lightness plane 161
3.3 Experiments E2. Verification Experiments 162
5.5.7 Comparing CIELAB to Hung & Bents and Ebner data sets 762
3.3.1.1 Introduction 162
3.3.1.2 Experimental 163
3.3.1.3 Results 165
3.3.2 Comparing IPT color space to Hung & Berns and Ebner data sets 766

CONCLUSIONS 170

FUTURE WORK 172

APPENDICES 173

6.1 Appendix A: Data plots from experiment PI 173

6.2 Appendix B: Data from experiment P2 175
6.3 Appendix C: Three parameter files corresponding to Figure 57 177

REFERENCES 179

vm
 List of Figures

Figure 1. Progression of thesis research direction 3

Figure 2. Regions of confusion in color space 15
Figure 3. MacAdam's tone compression curves for tone reproduction with limited dynamic
range 28
Figure 4. Taxonomy of business graphics image content 34
"wrong"
Figure 5. Different colors for the face. Face colors that are unnatural look 35
"natural"
Figure 6. Different colors for cars. There is no color for a car 35
Figure 7. Different colors for figuratively representative images 36
Figure 8. Figuratively representative images of automobiles with different colors 36
Figure 9. Reduced version of task and semantic differential scales for experiment PI (abstract

sub-class) 41
Figure 10. Example images from PI experiment. Abstract and Data representative images. .. 41
Figure 11. Variance accounted for by factors from PCA 43
Figure 12. User interface for experiment P2 (sub-experiment 1) 51
Figure 13. Images presented to the observers for experiment P2 (sub-experiment 1) 52
hab=303
Figure 14. Example hue leaf of CRT and Printer gamut shapes at 56
Figure 15. Histogram of
AE*ab for exact match exercise for entire observer
group including
outliers 58
Figure 16. Median match values for 25 colors in a*b*

plane, and in L*-C*ab planes 62

Figure 17. Example gamut
mapping vector with component vectors 63
Figure 18. 2D weighted vector gamut mapping. Open arrow heads indicate observer results.

Filled arrow heads indicate model predictions. Hue leaves shown are for the hue angle

of the observer match 65

Figure 19. 3D weighted vector gamut mapping. Open arrow heads indicate observer results.

Filled arrow heads indicate model predictions. Hue leaves shown are for the hue angle

of the predicted match 66

"corrected"
Figure 20. Hue angle matches on left, and hue angle matches on right (using
Hung's Constant Lightness data) 68
Figure 21. Comparison of Observer results with Katoh and Ito 1:2:2 vectors. Open arrow

heads indicate observer results. Filled arrow heads indicate Katoh and Ito results. Hue
leaves shown are for the hue angle of the predicted match 70
Figure 22. Comparison of Observer results with best 2d enhanced Katoh optimization

parameters. Open arrow heads indicate observer results. Filled arrow heads indicate
model results. Hue leaves shown are for the hue angle of the observer match 72
Figure 23. Comparison of Observer results with best 3d enhanced Katoh model optimization

parameters. Open arrow heads indicate observer results. Filled arrow heads indicate
model results. Hue leaves shown are for the hue angle of the model prediction 73
Figure 24. Images used in experiment P2, sub-experiment 2 77
Figure 25. Histograms of colors where the null hypothesis was rejected between sub-contexts78

Figure 26. Mean mapping vectors for each of the 5 sub contexts 79
Figure 27. Mean mapping vectors of Text sub-context (open arrow
heads) and the mean of all

other sub-contexts pooled (filled arrow heads) 79

Figure 28. Munsell renotation colors of Value 5 plotted on CIELAB a*-b*
plane 81

IX
Figure 29. Munsell renotation colors with extrapolation boundary and CRT gamut projection

onto CIELAB a*-b*

plane 83
Figure 30. Hung and Berns loci of constant perceived hue. Left plot is CL data, right plot is
VLdata 84
Figure 31. User interface for experiment 91
Figure 32. Hue leaf sampling at 0 degrees. The large square is the reference color 94
Figure 33. Projection of test colors onto plane. Large squares indicate reference
a*-b*
colors.94

Figure 34. Histograms of hue matches for 90 observations at reference hue angle of 0 degrees.
Histograms are plotted in the respective locations in the L*-C*ab plane of the reference

colors 95
Figure 35. Average and standard deviation of 95% confidence limits of mean hue matches as a

function of reference hue angle. Units of y axis are in degrees (CIELAB Ahab) 96
Figure 36. Histogram over the entire data set (306 colors X 30 observation means =
9180) of

weights calculated from inverse absolute difference 97

Figure 37. Histogram over the entire data set (306 colors X 30 observation means =
9180) of

weights calculated from inverse variance 98

Figure 38. Weighted mean hue matches in CIELAB space with +-95% confidence limits. ... 100
Figure 39. Projection of constant hue surfaces onto CIELAB a*-b*
plane 101
Figure 40. Loci of constant hue in CIELAB a*-b*
at different lightness levels. Dots represent

reference colors 101

Figure 41 Weighted
. mean hue matches plotted in CIECAM97s color appearance space. The
x-axis is labeled by CIELAB reference hue angle 102
Figure 42. Projection of constant hue surfaces onto CIECAM97s C cos(h)-C sin(h) plane.

Hue leaves are labeled by the CIELAB reference hue angles for comparison 103
Figure 43. Loci of constant hue in CIECAM97s C cos(h)-C sin(h) at different lightness levels.
Dots represent reference colors 104
Figure 44. Quantification of average hue non-uniformity for CIELAB and CIECAM97s 105
Figure 45. Quantification of maximum hue non-uniformity for CIELAB and CIECAM97s. 106
Figure 46. Constant hue surfaces shown in a linear opponent space (Y,X-Y,Y-Z). Numbers
represent CIELAB reference hue angles 107
Figure 47. Constant hue surfaces shown plotted in CIELUV color space. Numbers represent

CIELAB reference hue angles 108

Figure 48. Hung and Berns data plotted on
top of experimental results in CIELAB space. Left
graph compares CL (constant lightness) data. Right graph compares VL (variable
lightness) data. Hung and Berns data are shown in bold dotted lines 109
Figure 49. 3 node, 1 hidden layer neural net 1 12
Figure 50. Tanh function is expansive in the negative range, linear near the origin, and

compressive in the positive range 112

Figure 51. Examples of nonlinear functions. Left plot is function f with parameters al =

0.45, G2 =
0.6. Right plot is f2 with parameters Cl =
1, a2 =
1. Note that at the origin,

f, has a slope of 120

Figures 52. Visualization tool user interface showing the three slices of color dimension,
constant lightness, constant hue, and constant chroma 125
Figure 53. Close up of the controls and constant lightness display 126
Figure 54. Close-up of the constant hue display 127
bars"

Figure 55. Two dimensional "slider that controlled the pretransformation matrix 131
Figure 56. Examples of CIE 1931 color
matching functions transformed through

pretrans formation matrix. Left image is through an

identity matrix. The right image is
through Hunt-Pointer-Estevez matrix 132
Figure 57. Three examples of CMFs along with their influence on the shape of the color

space. Color images Berns CL data Munsell V5 133

space show
Hung and set and

Figure 58. Example hue lines from Guth type model. The first image is rotated at an angle of
0 degrees, the second is at 45 degrees. Note the shape difference of the constant hue
loci 135
Figure 59. Nonlinear controls for visualization tool. Six parameters correspond to the positive

and negative parts of each of the three channels. Functions take the form of

hyperbolic, power function, or chroma compression 136

Figure 60. A line showing the path of a gradient sweep from gray to the blue primary (in CIE
xyY
chromaticity space) 143
Figure 61. Path b*
of a gradient
sweep from gray to blue primary in CIELAB a*
space.

Notice that the mid chroma colors are more purple than the constant hue locus.
Dotted lines are
Hung and Berns constant hue loci 144
Figure 62. Path of a gradient
sweep from gray to blue primary in IPT P-T space. Notice that
the mid chroma colors are more purple than the constant hue locus. Dotted lines are

Hung and Berns constant hue loci 145

Figure 63. Color matching functions of the IPT color space compared to the Hunt-Pointer-
Estevez primaries. Note the larger response on the left tail of the M (green) response
function 149
Figure 64. Constant perceived data sets plotted in IPT (left) and CIELAB (right). Hung and
Berns33
Constant Lightness (CL) data set is in dotted lines. IPT is plotted with P and T
scaled by 100 150
Figure 65. Quantification hue nonuniformity for CIELAB, CIECAM97s,
of and IPT 151
Figure 66. Maximum deviation quantification used in the visualization tool 152
Figure 67. Munsell Value 5 data plotted in IPT deft) and CIELAB (right) 153
L*
Figure 68. IPT lightness correlate as a function of CIELAB 154
L*
Figure 69. Chromatic lightness response of CIELAB. x axis is of chromatic stimulus, y
L*
axis is of neutral stimulus that matched in lightness (mean observer judgement).
L*
RMS error between observed lightness and of chromatic color is 7.12 155
Figure 70. CIELAB matching L* response as a function of predicted L*new= L* + 0.098 C*ab.
RMS error between observed lightness and predicted is 5.97 155
L* L*
Figure 71. Chromatic lightness response of IPT. x axis is of chromatic stimulus, y axis is
of neutral stimulus that matched in lightness (mean observer judgement). ). RMS error
L*
between observed lightness and of chromatic color is 6.73 157
Figure 72. Plot of final lightness predictor for IPT color space for chromatic colors. RMS
error between observed lightness and predicted is 3.5 158
Figure 73. MacAdam ellipses plotted in IPT (left) and CIELAB (right) along with the spectral

locus 159
Figure 74. RIT-DuPont visual color-difference ellipses plotted in IPT (left) and CIELAB
(right) 160
Figure 75. OSA color system medium gray constant lightness plane plotted in IPT (left) and

CIELAB (right) 161

Figure 76. User interface for verification experiment 163
Figure 77. Scale values for judged uniformity between CIELAB and the 2 constant hue data
sets. H&B are Hung and Berns data, E&F are Ebner and Fairchild's data 165

XI
Figure 78. Scale values for judged uniformity between IPT and the 2 constant hue data sets.

H&B are
Hung and Berns data, E&F are Ebner and Fairchild's data 167
Figure 79. Mean scale values for colors and descriptive terms for abstract representative image
type 173
Figure 80. Mean scale values for colors and descriptive terms for data representative image
type 174

List of Tables

Table 1. Bipolar scales arranged

according to their factor loadings 44
Table 2. Model parameters for weighted vector gamut mapping model 65
Table 3. Summary of optimization results for Katoh and enhanced Katoh models 71
Table 4. Transformation error from hue correction LUTs 116
Table 5. Luminance and chromaticities of
viewing setup 175
Table 6. Colors chosen for gamut
mapping color match 176

List of Equations

Equation 1. Slope calculation for weighted component vector gamut

mapping 63
Equation 2. Weighted color difference equation 69
Equation 3. Enhanced weighted color difference model 71
Equation 4. Matrix form of forward computation of 3 node hidden layer neural network. ..113

Equation 5. General form of CIELAB type color model 118

Equation 6. Matrix to rotate second and third row elements for the opponent response 118
Equation 7. Functional forms for nonlinear
step 119
Equation 8. General form of CIELUV type color model 121
Equation 9. Inverse CIELUV type equation 122
Equation 10. Forward IPT model and coefficients 146
Equation 11. IPT inverse model and coefficients 147
Equation 12. Final lightness predictor for chromatic colors from Fairchild and Pirrotta68.
RMS error between observed lightness and L**(L*,C*ab,h) is 4.2 156
Equation 13. Final lightness predictor for IPT color space for chromatic colors from Fairchild
and Pirrotta68. RMS error between observed lightness and predicted is 3.5 158

xn
1 Introduction

The progression of this thesis work that led to the eventual characterization of hue umformity

and development of uniform hue color spaces was not linear. In fact, it was more

evolutionary, in the sense that the fittest paths to completion survived, whereas the paths that

offered less opportunity were abandoned. This dissertation is intended to capture all of the

work that was conducted from the completion of the thesis proposal to the completion of the

work that was deemed necessary and sufficient in order to graduate.

The background section covers the study preceding the research proposal, as well as two

experiments that were performed subsequent to completion of the proposal. These subjects

are related to the study of two things. Firstly, the study of color
meaning in different contexts

as defined by image content, using semantic differential scaling. Secondly, the study of

observer color matches to derive gamut

mapping models.

The approach section, as well as

following sections concentrate on the path that has led to the

culmination of the research efforts, which is the characterization and

modeling of hue

uniformity of color appearance spaces. A large psychophysical experiment was conducted to

find surfaces of constant hue in CIELAB color space. Fifteen equally spaced hue angles in

CIELAB color space were sampled. Thirty observers performed a hue matching task three

times each over 306 colors covering the full gamut of a CRT display. A computer program

was created to model color spaces

using variations of several simple models. Data from two

constant perceptual hue experiments were used to define metrics of

uniformity for optimizing
the metric space. Several other visual data sets were used to provide feedback on
umformity

and
accuracy of color appearance phenomena. A new color appearance space is proposed that

has acceptable hue uniformity, models several other visual attributes as well as or better than

commonly accepted and industry standard color spaces, and is more

closely related to our

understanding of the human visual system.

Figure 1 shows the progression of the work performed throughout the research period. The

first two experiments, labeled PI, and P2 (P stands for proposal, or preliminary), were part of

the initial proposal that was signed off by the advisory committee. PI did not yield results that

were deemed significant enough to follow, thus the path was abandoned. P2 was published in

a Color Research and Application article ,

and part of the results from this experiment showed that

the
non-uniformity in hue in CIELAB color space were so bad, the color name was changed

when
using the color space to perform gamut mapping. The available data sets for constant

perceived hue were found to be insufficient, so the subsequent experiments (El and E2) were

1998.2
designed around hue uniformity. El was presented at EI98 in San Jose in January

Study of hue uniformity has been significant enough to warrant the completion of the degree

based on this work.

 Proposal

P2: Gamut Other experiments

PI: SD scaling of Mapping
From Below from proposal
colors in context
were abandoned

X
El: Finding constant
Dead End
hue surfaces

Derivation of constant

hue color space

E2: Verification
experiments

Graduate

Figure 1. Progression of thesis research direction.

2 Background

The background section is comprised of two parts. First, a literature search is discussed on the

subjects of color order, color meaning, color harmony, and color preference. Secondly,

experiments that led up to the final thesis subject are described.

2.1 Literature Search

2.1.1 Color Order Systems and Color Harmony

Each
A is way to describe the relationship between
color order system a colors. color order

system attempts to create a space within which colors can be specified and compared in an

intuitively clear way. The fact that there are so many such systems implies that there may not

be a most appropriate
way to specify colors for every situation and within
every context.

From a color
harmony point of view, there are a few important color order systems that

respective theories are based upon. Color order systems of Munsell, Ostwald, Hard and Sivik,

and Moon and Spencer are thus described along with the basic foundations of their respective

theories of color harmony. Relevant experiments will also be discussed.

2.1.1.1 Munsell

Munsell3,4
Albert H. was bom in Boston, Massachusetts in 1858. Fie attended college at the

Massachusetts Normal Art School. Following this, he attended the Julien Academy in Paris on

a graduate fellowship. He won second prize in the Beaux Arts competition for his painting

"The Ascension of Elijah". Between 1898 and 1905, he developed his ideas on color order,

notation, and arrangement.

Although many others before Munsell have suggested ways to arrange colors, he was the first

to create an atias of physical samples that is

perceptually uniform. The Munsell color system is

arranged as a three dimensional solid with three dimensions of

Value(Lightness), Chroma, and

Hue. The hue circle is arranged so that there is an equal perceptual distance between each
major hue category. Munsell's major, or simple hue categories were red, yellow, green, blue,

and purple. Purple is included here because Munsell thought that there was a larger perceptual

distance between red and blue than there was between red-yellow, yellow-green, and green-

blue. Intermediate hues, or compound hues, are defined between each of the major hue

categories. These ten hue divisions make

up the hue circle. The hue circle is then divided ten

fold for each simple hue. The value 5 is assigned to each simple hue and 10 to each

compound hue. The notation for a given color is defined by a triplet of hue, chroma, and

value. The value axis goes from 0-black, to 10-white. The chroma scale goes from 0-neutral,

to an open ended high chroma number, although larger values have been assigned
by some.

On the physical atlas of color, some pigments are naturally stronger than others. Therefore,

the strongest pigment of blue-green only goes to chroma of 5 whereas the strongest pigment

of red goes to 10. Furthermore, the maximum chroma

may not happen at value 5. For

example yellow reaches maximum chroma at a value of 7, whereas purple-blue reaches

maximum chroma at value 4.

The Munsell notation for a particular color is <hue designation> <value>/<chroma>. For

example, a pink color might have a Munsell notation of 5R 7/3. Neutral colors are denoted N.

Munsell's system does not directly correspond to any theory of color vision, as do other

systems. Munsell was interested solely in the perception of color and did not attempt to

explain how or why.

Munsell, felt that there was a need for balance in every aspect of one's life. This included art

and color. He attempted, through heunstics and design tenets, to specify rules for creating

harmony and balance in color. Balance, as defined by Munsell, is visual comfort. Balance

means that a color scheme is neither too light or too dark, too weak or too strong, too hot or

too cold. Balance is described by several general laws:

The simplest balance of two colors is a balance about neutral gray. This means that two

colors that sit on a line drawn through N5 and are equidistant from N5 will be balanced.

These colors are considered to be complementary.

Area of color combinations is important. A strong red will balance a larger field of blue

green of small chroma. "The stronger the color we wish to employ, the smaller must be its

2
larger
chroma."

area, while the the area, the grayer or weaker the Area is related to value

and chroma by the rule that the product of value and chroma should be inversely

proportional to the respective areas of the two colors in question. For example, 5R 7/6

will balance 5R 3/3 in a proportion of 9 parts former to 42 parts latter.

Colors of the same hue can be made to balance if they are created with different values

and chromas. It seems always to be best if colors are connected

by a straight line whose

center point falls on or around a value of 5 (mid-lightness).

A small amount of discord is not always undesirable since it makes the scene more

interesting.
There are 9 chief features balance described Faber Birren in "Munsell, A
of as
by grammar of

Color"
. These are:

1. Gray colors harmonize best when neatly and spaced. Its best to center around N5.
evenly

2. Harmonies with
only one hue look best when they are arranged and around a
neady evenly

center point of chroma /5.

3. Opposite colors of medium chroma /5 can be combined in equal area.

4. Opposite colors of equal value but of different chroma should follow the above

mentioned area law.

5. Opposite colors of the same chroma but different value should fall on a straight line

through, and equidistant from, N5.

6. Opposite colors of different value and different chroma should follow the area law.

7. Munsell believed, as did others, that color combinations most liked by people comprise

either colors that are closely related or colors that are in contrast.
"Neighboring colors

should be neatly stepped as to value and should find sequence at middle value 5. They

should be of the same chroma for good balance. One color could have strong chroma and

the other could have weak chroma". The area rule should
apply to these colors.

8. There is harmony in dirninishing sequences. These are sequences that start with high value

and chroma and each successive color is reduced in chroma and value, and
possibly

rotated in principle hue.

9. Munsell liked the path of an ellipse through color space from one color to its compliment.
All Munsell's balance centered around his color system and
of notions about harmony and

simple geometric relations between colors in this space. These relations were drawn from

expenence as an artist and designer, and as such should be viewed to have more
validity than

Spencer13

have been Moon who did not have first

rules that generated
by workers such as and

hand art. Munsell defined of colors as those opposite

experience with
creating complementary

on the hue circle, and characterized them as most strongly contrasting. He also defined them

as two colors that made neutral when spun on a Maxwell disk. He goes on to say that an
gray

admixture of two colors of equal proportions makes a color that on the Munsell diagram
any

lies at the center of a line connecting the colors. This confusion between perception and

stimuli is interesting because it is the only place in his writings that he infers psychological

relationships from indirect observation of different stimuli.

2. 1. 1.2 Ostwald

Ostwald6,7
Wilhelm was around 60 years of age before he became interested in color. Before

this, he had worked in the fields of mathematics, chemistry, and physics. Ostwald's system is

comprised of planes of constant dominant wavelength, each of which is defined by a black

"fullcolor"

point, a white point and a A fullcolor is defined by Ostwald to be an idealized

spectral "block dye". There are enough fullcolor descriptions to span the hue circle, thus

enabling a 3 dimensional solid to be formed. Ostwald created his system before the CIE 1931

standard observer was adopted, and before Stevens started work on his power law (1953) .

Thus he relied on Fechner's law to create equal

spacing of colors. Ostwald created his

physical atlas from the use of disk colonmetry by mixing proportions of black, white and a
physical
analog of a "fullcolor". All colors that sit in a given triangle are compnsed of a

mixture of black, white and a given fullcolor. White and black content are denoted by lower

case letters of the alphabet,

excluding the letter j. Ostwald defined the equation C+W+B=l,

so
only two coordinates need be known to define the position in the triangle. The triangle is

set
up such that colors of equal white content fall on a line that is normal to a line that

intersects the white point, is normal to the equal white content line, and is parallel to the line

connecting black and fullcolor. The analogous rule applies for colors of equal black content.

Colors of equal fullcolor content form vertical lines.

Colorimetrically, the Ostwald system neither spans the entire visual space (as defined by the

CIE xy chromaticity diagram), nor is it

perceptually uniform or
internally consistent.

However, Ostwald's ideas about color

harmony are quite interesting. Because Ostwald

denned all colors as mixtures of black, white, and a pure color, his system fits well with how

graphical artists deal with, and think about, colors i.e. in tints, tones, and shades. Faber Birren,

in his analysis of the Ostwald system, said "When contemplated in the light of average human

appreciation, colors are seen not so much in terms of lightness or darkness, but rather in terms

of whiteness, blackness, grayness. A lavender (white-containing purple) and a sage green

(black-containing yellow) may have equal lightness. Yet they are

hardly to be called

harmonious."
2.1.1.3 NCS

Sivik9
The Natural Color System designed Hard is a Swedish standard
(NCS) was
by and and

for color notation. The NCS is intended to be a system that can be used without a physical

atlas for reference, although one was created for it for illustrative purposes. NCS provides a

way to describe
allows the
color without reference to a particular viewing situation, thus ability

to describe the perception without regard to a particular stimulus. The NCS space is arranged

in accordance to Hering's color

opponency theory, where yellow and blue are opposed, as are

red and green. NCS notation is similar to Ostwald's notation in that a given color lies in a

triangle whose vertices are white black and a maximum chroma color. Colors are described by

having a resemblance to black, white, and to up to 2 of the 4 primary hues. Within the NCS

triangle, whiteness and blackness are described in the same manner as the Ostwald system, and

the equation s+w+c=100 is conspicuously similar, where the letter s denotes. The notation

for the NCS system is a triplet, where the first number is blackness, the second number is

chromaticness, defined as the distance from the achromatic axis, and the third element is an

alphanumeric that denotes hue angle. The hue circle is broken into quadrants of yellow, red,

blue, green, and the hue notation describes the angle from one primary to the next.
They

make the point over and over again that "subjective phenomena can only be meaningfully

measured
using subjective methods... Thus, the NCS, as a system for denoting colors as and

when
they appear to man, is general and not based on the availability of the NCS Color

Adas."

10
Hard and Sivik have criticized
many previous workers in color
harmony for creating long lists
of dogmatic rules that are
mostly opinion and not based on general
rules.10

any underlying

They give as an example the definition of

complementary colors. There are at least five

different, somewhat contradictory, definitions of

complementary colors that are founded

tncks"

entirely on what
they call "stimulus instead of perception (for instance that two colors

mix to on a Maxwell disk makes them complementary).

gray Many of the rules of color

harmony of Geothe, Munsell, Moon and Spencer, and Chevreul have been shown to be too

rigid and not

terribly useful in real world situations.

Hard and Sivik have proposed a model of color combinations in which there are 9

dimensions.11
These are

"subdimensions"

Interval, with of: Distinctness of Border, Kind, and Size

Chord, with subdimensions of: Complexity, Content, and Type

Tuning, with subdimensions of : Color similarities, Area relations, and Rhythm.

Interval indicates the contrast between colors. Distinctness of border describes the borderline

between color surfaces and how sharp the edge is perceptually. Kind denotes where in space

the colors of the pair reside. Size is the distance the colors are from one another. The interval

dimension can be thought of as a more detailed description of color difference. Chord

describes how individual colors are experienced together. Complexity describes the number of

main attnbutes (the six elementary colors) that are represented in a combination. Content says

11
which mam attributes are in the combination. Type denotes what others may have called

complementary colors. It specifies whether colors on the hue circle are opposite one another

"Harmonics"

in the same quadrant, or in the same half-circle. to the of the

or Tuning refers

combination. Color similarities describes what aspects of the colors in combination are the

same, i.e. which attributes of each color are similar. For example, colors could be all of the

same blackness, or whiteness, or chromaticness, etc. Area relations describes the relative sizes

of the colors in question. Rhythm has to do with the regularity of a pattern or texture. A

repeating pattern with a small period would have lots of rhythm.

This nine dimensional approach to color combinations seems excessive, and the dimensions

are
hardly orthogonal in concept. These attributes of perception, however, are
interesting and

should be taken into account when

designing experiments dealing with the subject.

Sivik define "color

gestalt,"

Flard and also what

they term the which is the overall experience

of color and form when viewed in the real world. The basic elements of their color gestalt

12
are:

1. Color elements: Described by the shape of the contour that separates the color from its

surround and the actual color of the shape. There is no mention of smooth gradations

from highlight to shadow, but these presumably are part of the same color element.

2. Line Network: The pattern of lines that make

up the borders around the colors in a

particular area.

12
3. Overall Form: The pattern that the color elements and line network form. There may or

may not be an overall form. This depends on if there is a rhythm in the area in question.

4. Color Character: The color character can change if the color elements change, even if the

distinctness of border remains the same. The color character complements the form

gestalt to make the overall gestalt.

up

5. Balance: This is what

may be known in other spheres as color harmony. It takes into

account color elements, area balance, rhythm balance, etc.

6. Totality: The influence of context on the rest of the gestalt.

Quantification of some of these aspects of color gestalt

may be difficult to obtain.

The study of distinctness of border is of interest to both Hard and Sivik's theory of color

ideas.10
combinations and of their gestalt
They show that distinctness of border between

adjacent color fields can be thought of as a new

way to think about color difference from a

phenomenological point of view. Distinctness of border, or GT as

they abbreviated it (from

Swedish) was shown to be a one dimensional function. The research on the NCS lightness

function yields a blackness that is a function of CIE Y as follows:

1567
s = 100-w =
100
(r+56)

which says that blackness (or whiteness) is a hyperbolic function of CIE Y. Although

functionally different, the shape corresponds roughly to that of Stevens or Fechner. The

13
interesting part of this is that a constant difference in GT along the achromatic scale

corresponded to a constant blackness (or whiteness) difference. This makes NCS blackness

perceptually uniform with both blackness difference and with border distinctness, GT. This

relationship also holds with chromatic color pairs where the hue and chromaticness were kept

constant and the blackness varied. An equation was derived through multiple linear regression

that predicted distinctness of border GT =

f(As,Ac,A(j)) with a
very high correlation

coefficient. Furthermore, it was found that GT was not an additive quantity, i.e. if GT

between A and B is the same as GT between B and C (on the achromatic axis), the GT

between A and C is not AB + BC. This lack of

additivity is not
surprising if one considers that

center-surround antagonism of the visual system enhances contrast sensitivity at high

frequency edges. The assumption is made here also that the borders between colors is sharp.

Very different results would have been found if the borders between the stimuli were even

slightly blurred.

2.1.1.4 Moon and Spencer

Spencer13

Perry Moon and Domina took the CIE specification of the Munsell Color Order

System and created their "metric color space", which is basically just a of the CIE
warping

XYZ space to make it appear more

perceptually uniform.
They sought to create a

mathematical foundation for the rules of color

harmony through the use of their new color

space. Their two basic assumptions were :

14
1. Any arrangement of colors that can be sensed as an orderly combination will be pleasing.

Thus, in the new G5 space, simple geometric figures connecting colors will result in a

pleasing harmony.

2. The interval (Euclidean distance in the W space) between any two colors is unambiguous.

Moon and Spencer considered that

any stimuli that are regarded as
confusing will
necessarily

be disharmonious. From this they created their areas of in their color space. For
ambiguity

the hue circle, the regions are as in figure 2.

similarity

Figure 2. Regions of confusion in color space.

Regions of identity, similarity, and contrast are assumed to be pleasing, whereas the regions of

the first and second ambiguities are assumed to be displeasing. The same regions are found in

the other two planes in the space and are represented

by concentric ellipses in

lightness/chroma at constant hue values. Moon and Spencer (also referred to as M and S) go

a large number of harmonies, in one, two and three variables, with

on to classify

names and denotations in Munsell terms. For harmony in area M and S state
corresponding

15
"A pleasing balance among n color patches is obtained when the scalar moments about the

patches"14

adaptation point in CO -space are equal, for all the

Pope15
in his criticism of the Moon and Spencer work points out that they have greatly

oversimplified the problem, that classification of harmonies into groups of simple figures looks

nice, but needs to be validated

by experimentation. He criticizes their application of Birkhoff s

is divided
measure1

aesthetic M=0/C (aesthetic measure order

by complexity) as a true

mathematical equation, since it was intended more as a basic principle for guidance of design.

On the ideas of regions of

ambiguity Pope has disagreed, but this is dealing with an aesthetic

context more
generally associated with finer art, and not with graphic design, or page layout.

Pope sums
up his feelings in the quote "geometric significance in the symbolic representation

does
significance."13

not mean visual

necessarily

In Granger's experiment on area balance in color harmony, Moon and Spencer's predictions

with regard to area balance showed no predictive value, although Munsell's empirical rule

regarding area balance was shown to account for 35% to 53% of the total variance. Strangely,

there was a very high inter-observer correlation for area balance, between .67 and .73. This

may suggest that there is a predictor for preference of area balance of two color mixes.

2.1.1.5 Others

al.,18

Nayatam, et in 1969 performed some experiments two color harmonies.

on
They
conducted a paired comparison experiment on 102 two color
harmony samples. From this, he

derived, through linear regression, an equation to predict harmonies. Then he predicted the

16
aesthetic measures of 9890 two-color harmonies from the equation. From the predicted

aesthetic measures, his data suggested that two color harmonies of same or similar hues tend

be harmonious harmonize
contrasting hues.
to more than or Also, colors
complementary

better with N-9 (white), than with N-2 (blackish), and that aesthetic measure is affected
mostly

by value and value differences than other dimensions in color space. The experiment was

conducted on pairs of colored patches without reference to particular context.

In summary, it seems that most color harmony models or systems, at best, provide some

feeling to designers for a more systematic approach to synthesizing color harmonies. No one

system has been shown to be general enough to apply tenets to all situations. To try to design

an analysis system that employed one or more of the rules to combinations of color would

both be a prohibitively large undertaking, and would not guarantee any useful results. The

only way to glean useful information about color harmonies is to study observer's reactions to

color combinations, both within the context it is to be used in, and over a wide
sampling of

the space.
Only from results from study of particular context and application can rules be

established for that context.

2.1.2 Color Naming and Color Meaning

Object colors are those colors that we experience in normal photopic

viewing in the real

world. Although viewing conditions can change drastically in the natural environment (and

man made buildings), to a great extent colors of objects appear to remain the same, or at least

quite similar. This phenomenon, called color constancy, is greater if colors are associated with

familiar objects, e.g. people, fruits and vegetables, grass.

17
The ability to see colors as belonging to objects, in spite of the differences of (to a great extent)

the actual physical stimulus reaching the eye allows humans to use color as a tool to help them

survive in their environment. The ability to communicate these colors and to assign

unambiguous names to them is a useful tool.

2.1.2.1 Basic Color Names

Naming of colors in a given language seems to conform to a fairly strict

evolutionary process,

whereby the language must pass through successive stages of evolution to get to the next

level1
It has been shown that of the nearly 100 languages surveyed, a large majority of them

follow the following rule of evolution (for existence of color names):

purple

white green pink

<
[red] < [blue] < [brown] < orange
_black_
yellow

gray

This relationship means that a language most

likely will have words for white and black before

they have the word for red, and that the language will have a word for red before it has a word

for either yellow or green. A language may have, after red, the word for either green or for

yellow, but it must have both before it has a word for blue. These rules sound more like

arbitrary synthesis rules, but they have been derived from study of a large number of languages

in various stages of evolution. In the English language, there are an indefinite number of

words or expressions that denote color.

They are differentiated from the eleven basic color

terms by the following characteristics:

18
1. The meaning of the color is not predictable from the of its parts, e.g. blue-green.
meaning

2. Its significance is not included in that of other color term. e.g. crimson is a kind of
any

red.

3. Its application cannot be restricted to a narrow class of objects, e.g. blonde may be used

to describe perhaps
only hair, complexion, and furniture.

Names20
The ISCC-NBS Dictionary of Color uses a similar attitude when
assigning color

names to volumes in the Munsell color solid. The hue circle is defined in terms of basic color

names descnbed above, with the addition of violet and olive. Intermediate hues are described

by compound color names, e.g. reddish-orange. A slice of constant hue (value vs.
chroma) is

denoted using adjectival modifiers such as

very pale, strong, deep, and light.

Naming of colors is an outcome of the evolution of a language. As the lexicon grows, more

words are assigned to smaller differences in color

(up to a limit). As the ISCC-NBS

Dictionary of Color Names illustrates, in English, there are a huge number of color names.

The meaning associated with, or evoked by a particular color is a less well defined problem,

however. Connotation of
meaning of a particular color depends on context, experience,

culture, time, and other factors. There seems to be one invariant. Black seems to be the

universal color of mourning and death. Most people dislike black. Ancient cultures had

world"1

various and somewhat arbitrary meanings for the colors in their To Leonardo da

Vinci, yellow stood for earth, green for water, blue for air, red for fire, and black for total

darkness. Colors have been used to symbolize royalty. The royal color of the Sung Dynasty

19
(China) was brown, of the
Ming Dynasty was green, and of the Ch'ing Dynasty was yellow.

in India. The Hindu bride her

signify different
Different colors castes wears yellow at

wedding, whereas the Chinese bride wears, and is surrounded by, red.

2.1.2.2 Color Meanings

Physiological and emotional responses to colors have been noted and studied by psychologists.
Deutch21
For instance, Felix found that color brought about reflex actions in the involuntary

actions such as pulse rate, breathing, and blood pressure. He found, however, that red may

calm one person and excite another. There seems to be no general rule to attribute a response

to a given color stimuli. Deutch reasoned that physiological and emotional changes were

based individual experiences,

subjects'

brought about
by associations, so on people reacted to

color stimuli accordingly.

Color preferences have been studied from a number of different viewpoints. In a

survey of

21,060 subjects, H.J. Eysenck reported a mean

ranking of preference of basic colors in the

following order. Blue, red, green, violet, orange, yellow. However, this ranking was for color

in the abstract, without specific context.

Lars Sivik has done a number of studies

regarding the meaning of color. He used semantic

differential scaling, which is a

scaling technique that uses a seven
step bipolar scale of

subjects'

associations to ascertain feelings about stimuli. An example of an SD scale is

Warm OOOOOOO Cold

20
where the subject would fill in the most appropriate circle related to their response. Sivik and

coworkers have used the same scales of

meaning for many such studies23,24,25. Their list of

semantic bipolar variables is: like-dislike, winter-summer, loud-discreet, unappetising-appetising,

joyful-

serious, hygienic-unhygienic, old fashioned-modern, wet-dry, beautiful-ugly, soothing-exciting, positive-negative,

la^y-
sick-healthy, cultured-uncultured, feminine-masculine, complicated-simple, stimulating-dull, cold-hot,

energetic, friendly-hostile, active-passive, shameless-prudish, old-young, expensive-cheap, tense-relaxed, near-far,

secure-anxious. From a factor analysis (principle component analysis and

orthogonalizing) of the

resultant data, five major factors were found. These were denoted Excitement, Evaluation,

Potency, Temperature, Activity. Since the experiment was performed to be "context free", the

connotations from subjects was general and

possibly more
arbitrary than would be the case if

the stage was set in a more narrow sense. Given this, the inter-observer variance showed to be

fairly low. Further studies have been performed that set the context. In "Color connotations

difference
colors,"26

of exterior the above experiment was repeated with the being that subjects

were shown pictures of colors on buildings instead of abstract patches. The list of variables

were, for expediency, reduced in number to thirteen: warm-cold, vulgar-cultured, friendly-hostile,

hilarious-serious, pleasant-unpleasant, uncommon-common, soothing-exciting, masculine feminine, open-enclosed,

unclearly-ckarly demarcated, strong-weak closure, spacious-cramped. Two types of buildings were

shown, and 67 colors were used that fairly evenly sampled the color space. 168 observers took

part in the study. Analysis on the data showed three major factors; emotional (or pleasantness)

evaluation, social evaluation, and a spatial factor. The spatial factor is somewhat obvious, as

there are four variables that are very closely related to spatial aspects. In addition to factor

21
 in NCS
mapping"

analysis, Sivik employed "isosemantic of the meanings the perceptual color

space. This technique is interesting as it allows visualization of the meanings in planes of the

space, although the details as to the mechanics of creating the mappings are somewhat vague.

Whereas temperature was separated as a factor {warm-cold and la%y-energetic) in the context free

color study, it was found to be included in the factor called emotional evaluation. For color in

the building context the concepts of beauty and warmth are highly correlated, but for color in

the absence of context

they are independent. This was discussed to be partly a cultural bias,
"cool"
because Sivik states that there are very few colored buildings m Sweden, although this is

not the case in many other parts of the world. The issue that meanings of words are context

dependent was also discussed. If the meaning of a semantic variable changes as a function of

context, then there is

nothing limiting the SD pairs to
being opposite in meaning, and the

validity of the variable is questionable. This illustrates the need for well chosen variables that

are
unambiguously understood as opposite within the context of study. For the comparison of

individual variables of
meaning between color in and out of context, warm-cold, hilarious-serious,

found have Vulgar-

soothing-exciting, masculine-feminine were all to fairly high positive correlation.

cultured, beautiful-ugly, and

friendly -hostile showed
fairly low correlation. The factors that were

named evaluative in both studies were found to have very low correlation. Thus, this factor

was interpreted to have the highest sensitivity to context.

Scale"
Kobayashi has created a "Color Image that assigns semantic descriptors to perceptual

dimensions of color. He uses warm-cool for hue, soft-hard for value, and clear-grayish for

chroma. These attributes were derived from a factor analysis. The color space was also

22
segmented by hue, and
by tone, which is a name
corresponding to the ISCC-NBS color

dictionary tone descriptors of a given hue. Kobayashi states that the three adjectival bipolar

scales of warm-cool, soft-hard, and clear-grayish are

relatively insensitive to differences of

context, personal taste, and other environmental factors, and therefore can be used in a general

setting. He goes on to
map the space and assign different meanings to different parts of space.

This seems to be a gross overgenerakzation, and although the adjectives that re-describe the

color space
may be fairly invariant, there is no reason to believe that this is an exhaustive use

of the dimensions of color meaning.

2.1 .3 Preferred Reproduction of Pictorial Images

Colorimetry enables color matches to be predicted when original and reproduction are

measured under identical viewing conditions. Color appearance models attempt to enable

color matches to be predicted under

varying viewing conditions. Even if color appearance

models could
exactly predict matches across differing viewing conditions and media types,

they may or
may not
specify a preferred reproduction. Additionally, if viewing conditions are

constrained to be identical, and the original image is free from defect, a colorimetnc

reproduction
may not necessarily be preferred. Evidence suggests that, for several types of

image reproduction technologies and modalities, a colorimetnc reproduction is not preferred.

Froelich28

In 1948 Buck and studied human complexion under a number of standard light

sources.
They determined the average complexion of the three major races (Caucasian,

Negroid). Since human complexion is a

strong memory color and is used
by people to
Asiatic,

23
assess the
quality of color reproduction, it is a very important volume of color space. They

found that "...the preferred source emphasizes the material in brightness and saturation

without marked change in hue. Thus, in each case, the preference was for the lamp whose

This led
viewed."

spectral distribution and color most

closely matched the material
being to

the design of the soft white fluorescent lamp, which has a spectral response that is close to

average Caucasian complexion. This may explain

why a warmer color balance is often

Bartleson29
preferred when
viewing portraits of people. In 1960 studied the memory colors of

familiar objects
using color patches. He found that the memory color of flesh was closest to

the measured colors than for other familiar objects such as grass, sky, and brick. It also had

the least interobserver variance. His overall results indicated that memory colors were

generally more saturated than the real object and that the hue shifts of the memory colors were

toward the most impressive chromatic attribute of the color in question. This result was also

found by Newhall et al. in an experiment on a comparison between color

matching methods.

He found that memory matching, that is successive matching, yielded higher variability, shorter

matching times, higher remembered purities and somewhat higher remembered luminances

than the test stimulus.

On preferred reproduction of colors there is a

body of evidence that has been reconfirmed
by

different people over the years. These data stem

primarily from photographic reproduction

Bartleson31,32
experiments by both and
by Hunt33. Primarily, they were interested in the most

prevalent colors that existed in natural scenes, specifically grass, sky and skin color.

24
 '
Bartleson showed that the preferred color of skin in a reproduction was close to the
very

memory color of skin, and that it departs significantly from the true color of skin when

measured. The mean measured skin color from his subjects was Munsell 10R with a chroma

of around 4, while both the preferred and

memory colors of skin were about 7YR with a

Bartleson32
3.5. later
chroma of showed that the
relationship between memory and preferred

color did not generalize to grass and

sky colors. While preferred skin hue was different than

that of the measured values, preferred grass and

sky colors were of the same
approximately

hue. Preferred sky was higher in chroma, while preferred grass was of lower chroma.

Bartleson observed that while the preferred grass was of lower chroma, it may have been due

to the fact that the available gamut of the process they used could not
adequately produce

saturated enough greens at the time.

Hunt33
picked
up on this and performed an experiment similar to Bartleson's in 1974 when the

process capability was greater for making good greens. He found that the preferred color of

skin tended to be yellower and of around the same chroma as did Bartleson. For sky,

preferred and measured colors were of similar hue, but the preferred chroma tended to be

higher. For grass, preferred color had similar chroma as mean measured, but tended to be

yellower. Thus Hunt showed that Bartleson's results for green grass were
probably due to

process limitations rather than true preference.

25
 have investigated it
al.34

Recently, Asada, et preferred skin color and compared to skin color

from offset publications. They found that skin colors in publications were very similar to

fall in
u'v'

preferred skin color from their own experiments. Plotted in space, skin colors

ellipses that have small radii in hue and larger radii in chroma. As expected, Negro skin color

is darker than Asiatic, which is darker than Caucasian, while the hue angle of all skin tones

were quite close. Preferred skin color as a function of relative size of faces and color

background were studied. There was little difference in the results from differing sizes in the

preferred color, but preferred face chroma increased as a function of

lightening background

color.

In amateur color photography, and more

recently in color
copying it has been found that

"punchier,"
reproductions that are i.e. more colorful and
possibly lighter and higher in

contrast, tend to be preferred over reproductions that attempt to faithfully render color. An

Fedorovskaya.35
attempt to
quantify this was done by It was found that for four natural scenes

as viewed on a CRT monitor, changing chroma, both by scaling and

by translation (in

CIELUV space) affected preferred image quality. Images where the average chroma was

increased by 5 to 10 units were found to be the most preferred. Gamut mapping was done by

clipping in chroma.
They found a
very close
relationship between naturalness and
quality as

well, which is
hardly surprising.

Tone reproduction in color

printing has been given
relatively less importance in the literature

"colorimetric"
than reproduction. This may stem from the fact that most of the tone research

was done within the photographic field in a time when control over processes was determined

26
solely by exposure and development and chemistry. When the mode of reproduction

expanded to allow more manipulation as in the case digital the

arbitrary of reproduction and

specification of color changed from densities and absorbance to CIE based metrics, it
may

have been easy to reject or ignore much of the work

previously done. Consequently, some

workers stave to make colorimetric matches to source data without

taking into consideration

Jones36
the source and the output luminance ranges. As pointed out in 1944, a proportional

reduction of luminance contrasts to fit the dynamic range of the output system creates a far

from preferred reproduction. When doing cross platform "device independent

color"

typically

it is suggested that the

colorimetry be relative. That is, a CIE Y of 1.0 should correspond to

brightest white and a CIE Y of 0.0 should correspond to the darkest black attainable. This, in

effect, is exactiy what Jones warned was inappropriate; a linear scaling of the available

luminance range. Unfortunately, without absolute data about the source luminance range, it is

impossible to know exactly how to tune the tone reproduction. However, some assumptions

may be made to guess what a typical source luminance range might be for an average outdoor

scene and an average indoor scene. Jones states that the average brightness scale, which is the

ratio of maximum to minimum object luminance, is 160, or 2.2 log units. If this is used as the

luminance range for digitized pictorial images, then we can estimate the amount of luminance

compression for a given output process. For instance, the output luminance range of a typical

color laser printer is close to 60, or about 1.78 log units. This suggests that a tone curve that is

more sigmoidal might give a more preferred reproduction than a linear scaling. Figure 3

shows the curve from MacAdam.

27
 Fig. 9. Tone reproduction
y^y curves t
iS/y?

,y//
>/// A, "exact
B, optimum
reproduction"

attainable with

semimatte paper
////
C, proportional reduction of
./'
luminance contrasts to fit
8*

s/ /7* density range of semi-

matte paper
'y '7
D, tone reproduction suitable

^'/
~
for
lights
scene in which

are predominant
high

/
E, tone reproduction suitable
/ when only shadow details
are important.
2 9

*)0 0

Figure 3. MacAdam's tone compression curves for tone reproduction with limited
dynamic range.

Roufs38
approached this problem in a somewhat circuitous manner in a
study on perceptual

image quality. He used gamma functions instead of sigmoidal tone curves. Perceptual quality

as a function of gamma was studied. It was found that for different images, the gamma

corresponding to the preferred print was found to be different. He then scaled subjective

brightness contrast as a function of gamma and then related perceptual quality to subjective

contrast. He found that perceptual

quality was more
closely a function of subjective contrast

and so then, less dependent on image content. I think that subjective contrast of the image

was a function of where the detail was in the tone scale, as suggested by Jones. The images

with the most detail in the shadow regions would find perceptually best reproduction
using a

gamma function that gave the most dynamic range to that part of the tone scale and where the

28
slope of the tone scale was similar to that of the luminance gradient of the original scene. As

part of the same study, perceived sharpness was scaled as a function of both cutoff frequency

and gamma. It was found that within a reasonable range of gamma functions, there is a

negligible effect on perceived sharpness. However, when blurred with a gaussian filter, the

perceived sharpness is reduced. The study did not include sharpening the images using digital

methods such as laplacian filtering, unsharp masking, or error diffusion. Both tone

reproduction and sharpness have a significant impact on the preferred reproduction of natural

images, and as such are worthy of more consideration.

29
2.2 Preliminary Experiments
The preliminary experiments were the intended set of experiments from the thesis proposal.

There were to be four experiments centered around the meaning of color in the context of

business graphics, and the relationship of color meaning to gamut mapping for preferred

The descnbes the justification for

reproduction.
following text the approach and such an

approach.

That is,
above."

The problem of gamut

mapping has traditionally been approached "from

various
mapping techniques have been hypothesized and implemented, images were then

pumped through the mappings, and

finally some type of subjective judgment was made on the

various hypotheses to determine which approach worked best under the conditions of the

This
experiment.39,40

approach is then iterated upon until some satisfactory results are

is
above"

obtained. One primary reason that gamut mapping work has been done "from that

the images that have been treated have almost

exclusively been scanned images. The task of

explicitly mapping each pixel of a scanned image whose color is out of gamut is a prohibitively

huge task for almost

any observer (unless the image is very carefully chosen or manipulated to

be to the contrary) and would not lend itself to

being able to secure a
statistically significant

number of subjects. Further, even if the number of colors to be explicitly mapped is small

enough to be feasible, the task of color

matching individual pixels to closest in gamut colors is

likely to be image dependent and may not yield generalized results.

30
A fundamental drawback "from is
above"

of an approach that the search space of the problem

is essentially infinite for all practical purposes: It is the set of three dimensional mappings that

can be made from a source gamut to a destination gamut. In order to bound the search space,

"reasonable"
assumptions are made for a
mapping approach
using heuristics based on

These
researchers'

experience. assumptions
may not always be valid for all uses of color

reproduction. Furthermore, preference for a given mapping scheme is likely to be highly

image dependent, both on color content and on the context to which the image belongs.

Color content has been addressed

by adaptively mapping colors based on the number of out

of gamut points found in a given region of the color space of an image. However, image

context has not been taken advantage of as a means to make gamut mapping decisions. This

stems
partially from the fact that most of the work in gamut mapping has been done on

images images from the real it is difficult to

scanned (typically world) and quite accurately

segment scanned images into recognized objects.

Such may not be the case with computer synthesized graphical images that are commonly used

in business and the graphic arts. Typical graphical objects can fall into one of a small number

of general categories. One possible taxonomy of computer synthesized graphical objects is

given below. It first makes the distinction between foreground and background (also called

Background take the

ground), then further categorizes page elements. objects can
object and

form or fills. The fills are typically gradient fills and progress from a
fairly
of textures, patterns,

31
light color to white across the page for paper applications (under bright lighting conditions).

For applications where the medium is viewed under dark surround (overhead projectors or

film projectors) the background is typically quite dark, and the foreground objects are lighter.

Other background images are either

repeating bitmaps, as in realistic texture representation

(marble, wood), or
repeating patterns. Backgrounds can be more complex, for example

comprised of a scanned image. However this is less common as it may be confusing, thus,

detracting from the foreground information.

Foreground images fall into one of at least six categories. The categories heretofore identified

are:

1. Uterally representative images, i.e. images created specifically in reference to the image of a

naturally occurring object such as a face, hands, a cow, a computer, a tiger, etc. Images

from found in many

art"

this
category are typified by the "clip software packages. Images

like"
of this category "look their real life counterparts. This is an
extremely broad category

and spans a similar amount of content as scanned images from the real world. As such,

results found from a given image cannot be generalized to the entire category.

2. Figuratively representative images, i.e. those images that unambiguously represent either

like"

something out of the natural world or a concept (such as "STOP"), but don't "look

the image of that thing. One example of this is the stick figures that denote men's and

32
 ladies'
restrooms. This category is quite broad as well, and further categorization would

need to be done if the results, as in case 1, are to be generalized.

3. Externally imposed associations to familiar graphical objects. These are the images that don't

necessarily represent objects from the natural world, yet are so prevalent that they are

identified and remembered. An example would be a corporate logo, such as Xerox's

digital red X. Another example is a particular corporate color that is in standard use for

internal documents.

4. Data representative images, or the graphical communication of numbers. Charts and graphs

are an example of this type of image.

5. Simple geometrical objects. The colors that are attnbutes of images in this category are the most

abstract in terms of concrete association, and are probably more prone to symbolism from

personal experience.

6. Text Objects. Text is the most prevalent of

any object type in most computer

communication. Highlighted text directs attention to important parts of document.

Colored text may have significantly different connotations than other objects
similarly

colored.

33
 Taxonomy of computer synthesized graphical imagery

Computer Synthesized Graphical Imagery

Foreground Background

Literally representative Figuratively representative Textures and Patterns Gradient fills

images images eg. Marble texture, As in PowerPoint
eg. clip art (faces, hands, eg. stick figures Tiled pattern presentations

cow, computer, etc.)

Externally imposed Data representative

associations to images
familiar graphical objects eg. charts and graphs

eg. corporate logos

Abstract objects Text Images

simple geometrical shapes

eg. a red rec tangle

Figure 4. Taxonomy of business graphics image content

For each of the above mentioned categories of images and for images within each category,

may be different. For for

symbolism and associations of colors attached to objects
instance,

images that are supposed to look like from the natural world
objects
(Literally representative

images) and for images that are familiar and

frequently seen (Externally imposed associations),
colors are
fairly well defined and color per se
may less
meaning important than color
fidelity
(color fidelity in the sense of the
"supposed"
color
looking like it is to). Take, for example, the

image of a woman's face. If the blue

skin is colored or green or even
red, it looks quite

bizarre.

34
Figure 5. Different colors for the face. Face colors that are unnatural look
"wrong"

A study of
meaning of the color of these faces that are so unnatural has dubious value.
clearly

Of course, if the image is of a car then the coloration can take on a much wider range without

such a visceral reaction.

eliciting

"natural"

Figure 6. Different colors for cars. There is no color for a car.

Figuratively representative, data representative, and simple abstract objects have less obvious

constraints on their color attributes. Since figuratively representative images

clearly stand for

some
thing or some idea, but don't look enough like what
they represent to be literal, the

"acceptable"
latitude of colors is greater. Take, for example, the following images of a teacher

lecturing to two students. The fact that the teacher figure is different colors
may have subde

differences in the perceived meaning of the image, but none of the images is any more absurd

than another.

35
 m *_m tm m

Figure 7. Different colors for figuratively representative images.

The subject matter of the

figuratively representative object
may have significant influence on

the associations of the colors contained in them. For example, the same color
may have a

different connotation when it is associated with a stick figure than when it is associated with a

figurative representation of an automobile.

Figure 8. Figuratively representative images of automobiles with difterent colors.

For this reason, one image selected from the of figurative representative images will
category

not be representative of the entire group. Because of the need for further category

segmentation based on image content, the results from an experiment with a small number of

images will not be amenable to generalization. However, it is interesting to investigate the

difference in response to the same concept (e.g. car) when presented both as literal and as

figurative objects.

36
Data representative objects can take
many forms. However, all charts and graphs exist to

graphically commumcate numbers and, as results from

such, a small
sampling of specific

images may be able to be generalized to the entire category.

Both text and abstract objects also be generalized to their

may category more
easily than the

first three object types.

The success of the proposed experiments rested on the of three hypotheses

validity

1. The first hypothesis was that a gamut that is created from

mapping explicit
matching

(minimizing perceptual
distance) of individual colors will create a useful
mapping

transform. As will be shown in experiment P2, it was possible to create useful

mapping

transforms using the matching method. The color matches could be modeled with a

number of simple models that are developed in P2. They remain to be verified with

psychophysical experiments, but the author is confident that with the proper set of

parameters, they will perform as well as most other

clipping type transformations.

2. The change in the meaning of colors across contexts (same color in different image types),

is statistically significant and meaningful. To test this hypothesis, experiment PI was

conducted. This experiment showed that while there are some significant differences in

specific descriptive terms, and the factors found (dimensions of

meaning) were
very similar

to those found in other studies, it was difficult to glean from that useful instances of task

descriptions.

37
 be different (or both)
matching decisions
3. Gamut mapping color will when either the type

of image presented to the observer changes or the task description changes. Pilot

experiments were performed to verify these hypotheses. I concluded from these pilot

experiments (both comprise P2) that while the technique can be used to create interesting

models for gamut mapping, there was not a significant difference between color

matching means across images tested. Because of the failings of the PI pilot study,

the differences in task descnption between the two P2 experiments was the difference

between making a closest perceptual match and

making a match where the meaning of the

image was best preserved (where preservation of

meaning was left up to the observer to

define). On comparing the two experiments where the only difference was task

description (preserving meaning vs.

perceptually closest), the comparison was done

through looking at differences in the parameter values that went into the gamut mapping

model. No statistical
study was done of differences, but all model parameters were
very

close in value.

Although hypotheses 1 2 be valid,

and some parts of were shown to the invalidating of

hypothesis 3 forced a change in direction. From the gamut mapping experiments (P2), it was

clear that hue uniformity was a significant problem. From these data, and the lack of a

significant
body of other data in the field, it was decided that hue uniformity be further

studied.

38
2.2. 1 Experiment P1 (not completed). Semantic differential scaling of colors in the context
of business
graphics

An experiment was designed to test if dimensions of could be derived from a

meaning

Osgood42
semantic differential scaling of colors. did pioneering work in the area of meaning,

and the derivation of multidimensional spaces to represent it. He assumed that a

multidimensional space could be made to represent orthogonal dimensions of meaning, and

that one could derive those dimensions from a factor analysis applied to semantic differential

scaling of the concept in question. A semantic differential scale is one where polar adjectives

(descriptive terms of opposite meaning) sit between a 7 step scale. The task is to judge the

concept against where it falls on the scale between the polar adjectives. Osgood repeatedly

found that three major dimensions of

meaning accounted for a large portion of the variance of

the data (between 60% and 70%), irrespective of the subject

being measured. These

dimensions he called Evaluation (e.g. good-bad), which was almost always accounts for the

largest variance (about twice as much as the next 2), Potency (e.g. hard-soft), and
Activity (e.g.

active-passive).
Potency and activity on average accounted for about the same amount of

vanance. Sivik and others applied Osgood's techniques to the meaning of color (as described

earlier on page 20-22), and found additional major factors, the most notable of them
being

warmth (e.g. warm-cold colors).

The objective of the PI experiment was to validate that significant dimensions of

meaning

could be derived from semantic differential scaling as it applied to the meaning of colors in the

context of business graphics.

39
The abstract and data sub-contexts were chosen to be evaluated as a first attempt. (The

figurative, literal, and text representative images were going to be added following verification

that meaningful factors could be extracted. The experiment was halted before this was

achieved.) Fifteen semantic differential scales were used which were copies of the ones Sivik

used in reference 22. An example script is shown (reduced in size) in figure 9. For the data

representative image, the script read: "Don't judge the color as such, judge it as the coloration

line There The images for data

chart."

of the are 15 scales in figure 9. used abstract and

representative images types are shown in figure 10. Each of the image types was printed using

10 colors:
Sky Blue (printer Cyan), Red, Purple, Lime Green, Gray, Blue, Violet, Yellow,

Green, and Orange. For each image type (abstract and data), for each color (10 of them), 10

people performed the task

by filling out the
scaling page shown in figure 9. Although this

sampling is not sufficient to be significant, it seemed appropriate at the time as a quick test of

the method to see how and if useful information could be drawn from the analysis techniques.

40
Judge each color with respect to the scales shown below.

Interesting 3 ? Q Q 3 3 Q Boring

Old QG3QQQ 3 Young

Easy 3 3 3 3 3 3 3 Difficull

Dangerous 3 3 3 3 3 3 3 Safe

Warm 3 3 3 3 3 0 3 Cold

LikeQ 3 3 3 3 3 3 Dislike

Serious 3 3 3 3 3 3 3 Hilarious

Beautiful 3 3 3 3 3 3 3 Ugly

Soothing 3 3 3 3 Q ? Q Exciting

Positive 3 3 Q 3 3 3 3 Negative

Complicated 3 3 3 3 3 3 3 Simple

Real 3 3 3 3 3 3 3 Fake

Passive Q 3 3 ? ? ? 3 Active

Expensive 3 3 3 3 3 3 ? Cheap

Success 3 3 3 3 3 3 3 Failure

Figure 9. Reduced version of task and semantic differential scales for experiment PI (abstract

sub-class)

12 3 4 5 6 7 8

Figure 10. Example images from PI experiment. Abstract and Data representative images.

41
Means were calculated for each combination of color and semantic differential scale. Because

there are 10 colors and 15 scales per image, there are 150 means that need to be analyzed.

Each mean is comprised of 10 samples. For this paltry subset of the experiment, 3000

observations had to take place. Figures showing the mean and standard deviation for each

color and scale are quite large, so are shown in appendix A: Mean data plots from experiment

PI. As can be seen from the plots, the data are quite noisy as error bars indicate standard

deviation of the mean. The only way that useful results would be gleaned from this

experiment was if the data indicated that for different image types, there was a systematic shift

in meaning that was uniform across all colors. As can be seen from the plots, there are

significant differences in the mean values of some scales for some descriptive terms, but they

are not uniform across all colors. For instance, the mean value for the descriptive term

"interesting", for the color orange is about -0.4 for the abstract image type whereas it is about

-1.6 for the data image type. The significance is judged by the mean of one scale
being

beyond the error bar (1 standard deviation) of the mean from the other image type1. Though

there are some significant differences between means of colors and descriptive terms across

sub-contexts (sub-contexts is referred to as an image belonging to a specific type, e.g. abstract,

or data), the trends are not systematic or uniform enough for the results to be used for

subsequent experimentation.

This is not strictly correct, as the standard error of the mean should have been used to judge significance.. However, the

standard error of the mean would result in a smaller error bar since it is calculated from 1.96 *
standard deviation divided by
the square root of number of observers (10 in this case). Since 1.96/sqrt(10) is less than one, then the standard deviation is
larger than the standard error.

42
To test whether the results from this experiment were similar to those done by Sivik, and
by

Osgood, a principal components analysis

(PCA) was applied to the data set. PCA was done

based on Varimax method

using the SAS statistical software package. PCA was applied to the

entire data set (pooled sub-contexts), and to the abstract and data sub-contexts individually.

Because none of this work was continued, only the pooled results will be discussed.

The PCA found three factors that were significant (eigenvalues >
1) which accounted for

about 63% of the variance of the data. Figure 11 shows the first 10 factors (of 15) and the

amount of variance accounted for. Cumulative and individual variance is shown.

1
3

fl fl
a
"> 0.8
n n
'5

o
8 0.6
? cumulate
c
c a
(0 individual
re 0.4
I >

re
3 0.2
E
3
o 0 __JL-L-lTJiT-i-r-*1--ir-B1--T-n

1 23456789 10

Factor From PCA

Figure 11. Variance accounted for by factors from PCA

Osgood43 Sivik44

These results correspond to those found by both and

by although,
roughly

Sivik's results included 4 factors, and accounted for 85% of the total variance of the system.

dimensional "semantic
space"

The three significant factors can be used to construct a three on

43
which the descriptive terms used in the experiment can be mapped as vectors in the space.

Descriptive terms (bipolar scales) can be grouped

by factor loadings, which is used to

determine along which dimension a scale is most aligned. Table 1 show the scales arranged

according to their factor loadings.

Bipolar scale variable Factor 1 Factor 2 Factor 3 Primary

Factor

Interesting-Boring 0.8256 -0.2699 -0.1856

Old- 0.2240 0.4499

Young -0.7017

Warm-Cold 0.4692 -0.0044 -0.0631

Like-Dislike 0.8254 0.2782 0.0735

Beautiful- 0.8319 0.2129 0.0942
Ugly
Positive-Negative 0.6183 0.3831 -0.1306

Success-Failure 0.53665 0.3830 0.1237

Easy-Difficult 0.2263 0.5622 ! -0.4522 2
Dangerous-Safe 0.1098 -0.7946 0.1700 2

Soothing-Exciting -0.2100 0.7836 0.0900 2

Complicated-Simple 0.2314 -0.6766 0.1889 2
Passive-Active -0.5585 0.5860 0.1115 2
Scnous-I Iilanous -0.1210 -0.0108 0.8380 3
Real- Fake 0.4184 0.2046 0.6589 3
1 ixpensive-Cheap 0.5120 0.1061 0.6268 3-
Table 1. Bipolar scales arranged
according to their factor loadings

To apply Osgood's terminology for the three major semantic dimensions, factor 1 would be

named the Evaluation factor, factor 2 be factor 3

would named the Activity factor, and would

be named the
Potency factor. Quite few loadings
"clean,"
a of the are not that is, they don't

fall neatly into one factor, rather

they project onto more than one dimension of the space. The

scale Passive- Active, for is nearly

instance, split half and half between the first and second

44
factors. This analysis does show, however, that the fifteen scales can be reduced to a small

number of dimensions that account for a significant portion of the variance of the system.

Although somewhat interesting, the applicability of the experimental results to determining the

best task description for gamut was deemed unfeasible, so further analyses and
mapping

experimentation were halted.

45
2.2.2 Experiment P2. Gamut mapping through the use
of color matching (Gamut mappingfrom below)
This experiment was started concurrendy with experiment PI, and was intended to test the

truth of hypotheses 1 and 3. Restated, the objectives of this experiment were:

1 . To create a gamut mapping model where the parameters can be fitted to psychophysical

data derived from a color

matching experiment.

2. To test if there is a significant difference in the mean color matches across sub-contexts

(image types).

3. To test if there is a significant difference in the mean color matches across task description

perceptual"

(from "find closest to "preserve the meaning").

This experiment was split into 2 sub-experiments. The first experiment concentrated on

testing differences between sub-contexts, and the development of a general model to use to

create vector directions that make a color mapping similar to the color
matching mean

matches made by observers. The second experiment was a

copy of the first with the task

description changed to determine if color matches differed significantly.

2.2.2. 1 Perceptually closest color matches (sub-experiment 1)

2.2.2. 1. 1 Abstract: to P2 (sub-experiment 1)
A colorimetrically characterized computer-controlled CRT display was used to determine

closest perceptual color matches of 25 colors when an exact match was not allowed. An

artificial but realistic color gamut was created

by intersecting the display gamut with a gamut of

46
a Xerox 4920 color laser printer. Each of 21 observers performed color matches between out

of gamut colors and those on the artificial gamut's edge. Each observer made color matches

on 4 different images. The images represented some of the categories that business graphic

images can fall into. Between the different image types, there were no multidimensional

(MAN OVA) statistically significant differences at the 10% confidence level in any of the 25

colors tested. The mapping vectors showed that 1) observers don't make simple matches as

assumed
by most gamut mapping experiments done to date, 2) the influence of image content

for simple graphical images tested does not have a large effect when the task is to make closest

perceptual color matches, and

3) CIELAB hue angle is not uniform enough, especially in blue

and cyan regions, to make adequate gamut

mapping transforms. A simple model for clipping

type gamut mapping is proposed. Results are compared to predictions of a new gamut

mapping technique that minimizes weighted color difference between the target color and the

gamut boundary.

2.2.2.1 .2 Introduction to P2 (sub-experiment 1)

The gamut
mapping problem is well known in color industries. There are also
many

definitions of gamut mapping. The definition assumed here is: Gamut mapping is a solution

to the problem encountered when

rendering a color image on a device that cannot accurately

reproduce all the colors in the image, after

accounting for viewing conditions.
Many

approaches have been examined, and some have proved more useful than others. Stone, et

al.45

introduced the concept of gamut mapping, but lumped viewing conditions transforms into

Stone,46
the mapping, and used CIEXYZ space to alter color information. Later, Wallace and

47
Gentile, et
al.,47

and others separated the issues of gamut mismatch from adaptation

transforms somewhat by performing mappings in more perceptually uniform color spaces,

Bouman48

such as CIELAB color space. Wolski, Allebach and have created transforms that

L*
use different types of mappings in different parts of color space (above and below of 50).

Berns49
Hoshino and provide a good summary of
existing techniques for gamut mapping, all

mapping is done in
that the a uniform color appearance type space.
assuming perceptually

Fairchild50

Montag and have recently completed a large systematic evaluation of existing gamut

mapping techniques, including clipping transforms, and compression transforms along lines of

constant lightness, constant chroma and constant saturation. Test images were simple spheres

and the entire test was performed on a CRT screen.

Gamut mapping have (to been defined

authors'

transforms without exception the knowledge)

functionally, then tested. Many functions have been proposed for mapping a point outside a

color gamut onto or into the volume of reproducible colors.

Mapping vectors typically follow

one of five directions (with 4 of 5 rnaintaining metnc hue angle to be constant):

1) Constant

lightness, 2) constant chroma, 3) constant saturation or

analog of saturation, 4) centroid, or

towards a central point in the gamut (e.g. 50,0,0 in CIELAB), and , 5) minimum distance or

minimum weighted distance to the gamut surface. Sometimes a combination of approaches is

taken based on location in color space and in certain sequences. A new gamut mapping

technique using minimum weighted color difference has been proposed

by Katoh and Ito51,

and will be discussed in more detail later. While linear and non-linear compression

48
techniques have been investigated, the direction of the compression either preserves lightness

and hue, or preserves saturation (or it's analog in CIELAB) and hue.

While many of the above transforms work well under certain conditions, the search space for

mapping a three dimensional point to another point in three space is essentially infinite. The

functions for transforming color data are based on heuristics learned through experience, and

from that point are not invalid. However, this does not guarantee that any transform tested is

in any way optimal. What is not tested, cannot be found.

With this in mind, an experiment was designed to find the perceptually smallest distance from

a point in color space to a closed surface of color (gamut boundary). The focus here was on

gamut
mapping of business-graphics images, although the results may be generally applied to

all image types. Additionally the experiment tested whether image type affected observer's

gamut mapping decisions. The task descnption was the same for all trials:

"Adjust the color of the image on the left hand side of the screen to make it look as close as

possible to the image on the right hand of the screen.".

Three different image types were presented: Abstract (a colored square), data (a colored

and figurative (a cartoon car). An additional image was used (a colored square with a
chart),

to test the dependence of a black border on observer's matches. The experiment

border)

49
allowed observers to pick colors from a two dimensional image of a simulated gamut edge to

gamut"

match colors that were presented in an image. Twenty five "out of colors were

21 different An exact color exercise was

selected and mapped
by observers. matching

performed by each observer first to determine subject's skill at color

matching using the

interface four (of 25) had errors that were

provided.
Using the exact match results, subjects

too large and were not used in the subsequent experiments.

2.2.2.1.3 Experimental

There are two major and presumably separable problems

today regarding the reproduction of

colors from CRT to hardcopy print, or from any source color gamut/ viewing condition to any

destination color
gamut/viewing condition. These are
1) accounting for differences in viewing

condition, i.e. viewing mode (self luminous vs. hard copy), chromatic adaptation, luminance

adaptation, etc. and 2) accounting for gamut mismatch. These experiments assume that

problem
1) has been solved
adequately and concentrate on
investigating problem 2).

The following sections explain the details of the experimental design. The user interface tool

was designed and implemented. The test images were selected. The viewing conditions were

set
up and the CRT was calibrated. The simulated gamut edge was created. The task

description and selection of 25 colors was made.

Twenty one adult observers, 9 females and

12 males, performed the exercise.

2.2.2.1.3.1 User Interface

Fig. 12 illustrates the user interface for the experiment. The target image was presented at the

top right. The image that is manipulated to make a match was presented at the
top left. The

50
gamut edge (set of available colors from which to and a zoomed version of the gamut
choose)

edge were presented across the bottom.

White background

Ta rget
Irr tage
(outo f gamut)

next

Gamut surface Zoom

previous
area area

done

Figure 12. User interface for experiment P2 (sub-experiment 1)

"gamut" "zoom"
To use the tool the observer picks a color in the area. This changes the area

"zoom" "gamut"
to make the center color in the area the same as the color picked in the area.

"zoom"
When a point in the area is clicked, the color of the parts of the test image that are

alterable are changed to have the color selected. The colors can be updated
very quickly and

"zoom" "gamut"
change as the user drags the pointer around in the area. The area is a two

dimensional representation of a gamut edge in CIELAB L*C*abh color space. The x axis

represents hue and the y axis represents lightness. The color at the point x,y is the h=x, L*=y,

and C*^ =
the chroma limit at the corresponding L*,h coordinates. There are 36,000 distinct

colors in the gamut area, enough to create the appearance of a continuous gradation of color

around the gamut surface.

51
Any target image can be used in this test. The parts of the image that are updated by the user

correspond to an arbitrary length list of polygons described by an initialization file.

"contexts"

Four test images were selected to represent three different image The images

shown are simple computer generated graphical type images that represent the abstract, data,

and figurative image context groups. Not studied here were representative images from the

literal and text contexts due to limitations in the availability of software tools to generate the

proper polygon sets to represent the alterable pixels. Fig. 13 shows examples of the three

"contexts" "abstract"
selected image that were used in the experiment. Two contexts were

used, one with no border, and one with a black border.

-P
12 3 4 5 6 7 8
Figurative Abstract no border Abstract Data

Figure 13. Images presented to the observers for P2 (sub-experiment

experiment
1)

2.2.2.1.3.2 Viewing Conditions

The experiment was performed under simulated D65 lighting. The decision was made to test

under bright surround conditions to more

closely model real life situations where the results of

52
this work
may be used. The CRT and computer were set
up under controlled
lighting with a

background and surround of

gray matte board. An attempt was made to match both the

luminance and the

chromaticity of the llluminant white and the CRT white point to nullify any

chromatic adaptation issues. A mixture of fluorescent D50 and D75 lamps were used because

the color
rendering index of available simulated D65 source is quite low. A Minolta Chroma

Meter CS-100 was used to make all the color measurements. A PTFE
pressed puck
(Halon)

and paper measurements were taken at the same angle as the CRT screen. The xyY values of

the surfaces are shown in the first table in appendix B. The AE*^ calculated under D65 for

the difference between the measured white of left and right positions on the CRT was 3.5.

2.2.2.1.3.3 CRT Calibration

The CRT calibration was done in two steps. First the white point digital counts were selected,

18"
then the system was characterized. A Sun Sparc 5, CRT was used. There is no
way to

directly alter the gain of the monitor, so the digital count maxima must be altered to affect the

white point. The chosen white point was R=255 G=240 B=220, with xyY value of (.305 .321

79.1). The chromaticity of paper white under the selected illurnination was (.320 .333
80.3)

which is between D65 and D55. The gray background of the matte board at (.307 .321

13.3) is

quite close in chromaticity to the CRT white. Based on these measurements it was thought

that there would be little confounding from differences in surround luminance or chromaticity.

Once the white point was calibrated, the digital count to luminance relationship was measured

for each of the R G and B guns

by measuring .2, .3, .4, .5, .6, .8,
and 1.0 times the maximum

53
digital counts separately. A spline
fitting routine was used to create smooth curves that

mapped digital count to luminance of separate channels. Then XYZ measurements were

taken of R, G, and B maximum values, and a 3x3 matrix was used to create the transformation

to XYZ from RGB primaries. Flare was measured but not included in the CRT model. The

accuracy of this calibration is an average AE*^ of less than 2.

Once the CRT was characterized, forward and reverse models were coded and a 3 dimensional

LUT was made. Each of the test images was processed through this LUT to make the neutrals

in the image the same chromaticity as the white of the CRT.

2.2.2.1.3.4 Gamut Selection

The reverse CRT model was used to find the gamut of the CRT in CIELAB space. A

16x16x16 sampling of CRT device space was converted to 16x16x16 samples in CIELAB

space. The edge of this device cube (16x16x16 14x14x14 values) was used to make a

piecewise planar surface in CIELAB that represents the gamut edge. A two dimensional

gamut
boundary description was created
by using gamut software that performed
mapping

"leaves"
chroma with constant hue and lightness. Flue first
clipping were created
by sampling

CIELAB L*C*abh space at 100 lightness levels and 360 hue partitions with a chroma that is

higher than any device chroma everywhere (128 was used). The gamut software was
mapping

then used to
clip the chroma of each lightness level of each hue leaf. The resulting data can be

represented as a 2D image of L*C*abh values, or as a sequence of hue leaves around the hue

circle.

54
A second gamut was created from CIELAB values obtained from sampling output of a color

laser printer (Xerox 4920). A combination of the CRT and printer gamut was used to create

the simulated destination gamut edge for the experiment. A gamut boundary was created
by

taking the minimum chroma of the two gamut boundaries at each L*,h point. This creates an

"intersection"
gamut. A preliminary gamut edge chosen was an intersection gamut with

chroma scaled
by 90%. The resulting gamut edge was quite under saturated, though and it was

"intersection"
decided to change to an unsealed gamut edge. It was felt also, that the

intersection gamut is the most reasonable one for simulating real-life conditions.

The intersection gamut

boundary was converted to an RGB device image (360x100 pixels) and

"gamut"
scaled to fit in the window space of the color matching tool. Fig. 14 shows an

303

example of the CRT gamut and the printer gamut at a hue angle of degrees.

55
Figure 14. Example hue leaf of CRT and Printer gamut shapes at hab-303.

Because the experiment was performed

entirely on a CRT display, the only type of gamut

mapping that could be tested was where the CRT gamut is larger than the printer gamut. It is

well known that many printers can create colors the CRT is incapable of making, especially in

cyans and greens of mid lightness, and in bright yellows. Because of the limitations imposed

by the display medium, gamut

mapping where the printer has higher chroma was not testable.

2.2.2.1.3.5 Color Selection

Twenty five colors were selected from the surface of the CRT gamut as test colors to match

"intersection"
with colors on the reduced gamut boundary. They were chosen to span the

color space
fairly uniformly. Representatives from the 11 most basic color terms were

chosen, as well as a concentration about blue since it is an

interesting color center (the greatest

hue non
uniformity in CIELAB space is in the blue region). Out of gamut colors were chosen

in parts of the color space where there was a large mismatch in volumes. This means that in

56
blues and magentas, the target colors were
significantly darker than in the greens, cyans,

yellows and reds. Because there is no large volume mismatch in light blues and magentas, nor

in dark greens, cyans, yellows, or reds, the sampling is sparse in those regions.

The colors selected for the gamut match test are shown in the second table in
mapping

appendix B. Twenty five colors seem to be close to an upper limit for a session
matching

without a break. Since most observers performed two tests in a given sitting, 50 colors could

25th
have been chosen with a break at the color. The mapping task is quite a bit more

frustrating than the exact match exercise though, so it was important not to make the task too

overwhelming to the observer.

2.2.2.1.3.6 Exact Match Data and Results

The determine
observers'

exact match experiment was used as a vehicle to individual abilities

to 1) match colors and

2) use the interface. The average AE*^ over the entire data set was

4.25. The median

AE*^ over the data set was 3.08. The standard deviation was 3.61. The

histogram is shown in Fig. 15.

57
 Frequency of Delta E

\- <b- fc> 6' A- % f,? f,? ^-

Or Or <V v r <V

Figure 15. Histogram of AE*b for exact match exercise for entire observer
group including
outliers.

The average error is on the same order as the color error between the left and right areas on

the CRT screen that presented the target and test color patches. Much larger errors can be

seen in low frequency. These may be attributed to carelessness, or to fatigue in observers.

Four observers were found to have average errors over 7. These observers were not allowed

to continue, as their average errors were too high. The remaining 21 observers had average

errors of 7 or less. Most large errors were in

very dark colors where influence from flare on

the CRT (bright surround

conditions) was confounding.

2.2.2.1.3.7 Results and Discussion

Twenty one observers performed the gamut

mapping color match task. Four sets of images

were presented to each observer over four different sessions. The image sets were presented

in the same order for each observer. Observers typically matched two sets of images in one

58
sitting.
Including the exact match set, an average of three sittings was necessary to complete

five sets of
matching tasks. The four sets of images were named abstract, data, fig, and

noborder. These sets corresponded to an image of a square (with a black border), an image of

a data graph, an image of a cartoon car, and an image of a square with no black border.

Pictures of these images are shown in Fig. 13. The list of colors was randomized once and

used for all sets of images.

Each of the 21 observers performed one set of matches for each of the four image types. The

data set contains, for each color, for each observer, for each image type, one color match

vector. Each color match vector is composed of three independent numbers and must be

treated separately. Given this data set, means can be calculated in a

variety of ways.

Interesting means include the mean mapping vector for each color and the mean vector for

each color for each image type. Additionally, it is

interesting to test whether the difference of

the means of color vectors for each image type are statistically significant. If there is no

statistical difference in the means, they can be pooled to create a better estimate of the overall

mean of the matching decisions.

2.2.2.1.3.7.1 STATISTICAL ANALYSIS

For each of the 25 colors, a multivariate ANOVA was performed to test whether mean (three

dimensional) mapping vectors from the 4 image types were significantly different from one

another. The null hypothesis for this test was

=
HO: fj.1 =
(12 =
jX3 (14 where

59
(Il =
mean of context 1 (abstract)

p.2 =
mean of context 2 (data)

p.3 =
mean of context 3 (figurative)

|I4 =
mean of context 4 (no border abstract)

HI: at least one mean is different from another.

b*
There are 21 subjects X 4 contexts of data for each color for each of L*, a*, and vectors.

Since the mapping vectors lie in three dimensions, it is appropriate to use multivariate

techniques to analyze the difference of the means (MAN OVA). Three dimensional

MANOVA was calculated for each of the 25 colors tested. The SYSTAT statistics software

Wilks'
package was used to do all the analysis. Lambda, Pillai trace and
Hotelling-Lawley trace

were all performed to calculate f-statistics. The critical F value for these tests was calculated

based on 9/189, 9/240, 9/230 degrees of freedom respectively. The critical value for all tests

is equivalent to 9/ degrees of freedom. The critical F value at the 10% level is 1.6315.

The null hypothesis was not rejected at the 10 percent level for any of the 25 colors tested for

independent
colors'

any of the above techniques, although univariate tests on some

dimensions did reject the null hypothesis. Based on these data, it was concluded that there is

no significant difference between contexts for color

mapping decisions. The data from all four

sets were pooled and considered to come from one population.

60
Some comments of observers are relevant deserve Subjects
though, and comment. noted that

bright cyan and bright green, both device CRT maximum were the most difficult and
colors,

frustrating colors to match

(indeed, it was impossible to make an adequate match). These

colors also appeared to be self luminous, or "fluorescent", and impossible to match with the

palette of colors allowed to the user in the limited "gamut

edge"

set of colors. This is the part

of color space where there the most mismatch between CRT.

is printer and CRTs can make

very light, brilliant greens and cyans, where printer the colors are
highly chromatic at much

darker colors.

2.2.2.1.3.7.2 MEAN MAPPING VECTORS

Fig. 16 shows the median match vectors in the a*-b*
hue leaves
plane and as in
L*-C*ab planes.

The gamut boundaries shown are for the hue angle that the point mapped to, not the hue leaf

associated with the target color.

"constant"
The general trend seems to be a
(perceptual) hue mapping with lightness and

chroma both being traded off.

Mapping vectors for the most part point toward the center of

L*
the volume, although not
exactly at 50, but quite close in many parts of the color space.

In the yellow-green quadrant, the vectors all point toward greater than L*=50, while in the

L*
yellow red quadrant, the mapping vectors all point toward less than =
50. This may be due

to the yellow-green hue leaves being convex where there are more high chroma colors near the

target lightness. The yellow-red hue leaves are all concave, so more lightness must be

sacrificed to get reasonable chroma. The centroid behavior of the mapping vectors is not

since the task description was to make a closest perceptual match.

surprising especially

61
 ao 100 120

Figure 16. Median match values for 25 colors in a*b*

plane, and in L*-C*ab planes.

"match"
If the task description was to make a preferred or to make a match that preserves the

intent or
meaning of the color in the context of the specific image shown, this behavior may

change significantly.

2.2.2.1.3.7.3 A MODEL OF MAPPING COLOR VECTORS TO THE GAMUT

SURFACE:

Every color match vector can be thought of as a combination of component vectors that

originate at the target color and end somewhere on the surface of the gamut. A reasonable set

of three component vectors is:

L*
1) constant vector,

62
2) centroid vector (points toward 50,0,0),

3) minimum distance to gamut surface vector in the L*-C*ab plane (at fixed hue angle).

Fig. 17 shows an example of a

mapping vector with it's constituent component vectors. Note

that, as is the case in the experimental data set, the match vector is a combination of at least

two of the three component vectors and cannot be represented

satisfactorily by any one

component.

perceptually smallest

match vector

gamut

edge

Figure 17. Example gamut vector with component vectors.

mapping

The slope P of the line (in the L* C*^ plane) from the target color to the gamut boundary is a

combination of weighted component vectors. This is shown in equation 1.

* *
P =
(Lw *
lightness.y + Cw centroid.y + Mw
*
mindist.y) / (Lw
lightness.x + Cw * centroid.x + Mw * mindist.x)
for weighted component vector gamut mapping.
Equation 1. Slope calculation

63
Lightness, centroid, mindist are two dimensional unit vectors calculated for each target color.

The .x and
.y
suffix for each unit vector denote x and
y axes. The x axis corresponds to

chroma and the y axis corresponds to lightness. Lw, Cw, and Mw are scalar multipliers. The

mapped color C is a point that sits on the gamut boundary and intersects a line with slope P

that passes through the target color.

The three dimensional extension to the model is to find the minimum AE*ab between target

color and mapped color P in a range of hue angles about the hue angle of the target color.

The parameter Hw determines how many hue leaves will be searched for a n^nirnurn AE*^

between the target color and the color P. Hw is an integer that is the extent above and below

the target hue the routine will search for a minimum.

A model was computed in both the two and the three dimensional cases. In order to find

appropriate parameters, a local optimization was set

up that minimized AE*^ between mapped

model vectors and observer matches, as a function of the scalar parameters, Lw, Cw, Mw, and

"amoeba" C52
FIw (in the 3d case). The routine from Numerical Recipes in was used to find

minima. The routine implements an N dimensional simplex that crawls toward minima by

using reflection, expansion and contraction. Table 2 shows the results for two and three

dimensions. The errors calculated in 2 dimensions are

only for
L*
C*^ plane and don't try to

account for target hue angle

being different from mapped hue angle. This case would be

appropriate if the color mapping space was

perfectly uniform in hue angle, and observer's

64
choices preserved perceptual hue. Figs. 18 and 19 show the resultant vectors compared to the

observer matches. Note that the regions of greatest error are in the blues and in the concave

hue leaves of the red-yellow region. The hue leaves shown are for the hue angle of the

observer match for the 2d case and for the model prediction in the 3d case.

Minimization Lw Cw Mw Hw Max error Average

error

2d(L*C*J 1 1.04 .25 NA 9.8 3.24

3d 1 1.03 .38 +-3 33.9 6.84
Table 2. Model parameters for weighted vector gamut
mapping model.

40 60 80 100 120 20 40 60 80 100 20 40 60 80 100

20

80 100 120 10 20 30 40 50 60
20 40 60

Figure 18. 2D weighted vector gamut mapping. Open arrow heads indicate observer results.

Filled arrow heads indicate model predictions. Hue leaves shown are for the hue angle of the

observer match.

65
 100

r ^Op<4
\\l 80 jffli 5

^S1 /^53.9
60
e^ fes.

&3 SO 75

40

20

4>v
\

-100

33 20 40 60 80 100 120 20 40 60 80 100

20 40 60 80 100 20 40 60 80 100 120 10 20 30 40 50 60

Figure 19. 3D weighted vector gamut mapping. Open arrow heads indicate observer results.
Filled arrow heads indicate model predictions. Hue leaves shown are for the hue angle of the
predicted match.

2.2.2.1.3.7.4 HUE MAPPINGS

Constant hue loci data from Hung and Berns were used to test hue nonuniformity of

mapping vectors. If the constant hue loci could be used to correct the hue nonuniformity of

the mapping vectors, it may be concluded that observers are performing color matches that

preserve the perceived hue of the target color.

The significant areas of hues not

mapping toward a*=b*=0 are in blues and cyan. Using the

constant-lightness data set from Hung and Berns, a

lookup table was made that mapped

66
incoming hues that sat on constant hue loci, to the reference hues that were used to find the

loci. In between these 12 hue loci, linear interpolation was used to fill the rest of the space.

The constant lightness data set created a that corrected hue vectors
mapping significantly

toward 0 for the blue region, although the cyan region was unchanged and green was made

slightly worse (yellower). Fig. 20 shows the original matches, along with the "hue
correc

matches in the a*-b*

plane. Given that the data used to create this consisted of 36
mapping

data points, this is not

unimpressive, at least in the blue region of color space. Unfortunately,

the cyan region was not helped. This is

probably due to the fact that the Hung and Berns data

for lightness, significantly darker.

are constant whereas the cyan color vector gets
mapping

Based on observer feedback it seems the mappings are done with a

strong emphasis on

maintaining perceived hue.

The data here strongly suggest, at the very least in the blue and cyan regions, that CIELAB

color space is not uniform enough to do constant metric hue gamut mapping. Additional data

must be drawn from constant hue loci experiments to straighten the hue metric before it will

be adequate for gamut mapping.

67
 b plot, L marks arrow tails

a b plot, L marks arrow tails

"corrected"
Figure 20. Hue angle matches on left, and hue angle matches on right (using Hung's
Constant Lightness data).

2.2.2.1.3.7.5 COMPARISON WITH KATOFI AND ITO'S MODEL

Ito9

Recently, Katoh and have proposed and tested a new gamut

mapping model. Since it is

similar in some respects to the above model proposed

by the author, it deserves inspection.

The model defines a gamut

mapping direction for each out of gamut color
by finding the

minimum weighted color difference in CIELAB space. The weighting of the color difference

is similar to that of AEcmc ,

where AL*, AC*,,,, and AH*^ are weighted
by dividing by scalar

values. The weighted color difference equation used is shown in equation 2.

68
 A2 A2
AL*Y (AC* (AH*
?b
AE =
+1 +
ab

Kl Kc Kh
)

Equation 2. Weighted color difference equation.

Kl, Kc, and Kh can be altered to change the shape of an ellipsoid like surface of constant

error.
Mapping vectors are defined by the target color and the point at which the smallest

ellipsoid intersects with the gamut volume (with the target color being at the center of the

ellipsoid). In a discrete sense (and how it was done in this work), the vector is found by

calculating the weighted color difference between the target (out of

gamut) color and each

point on the gamut boundary. 36000 data points were used to represent the gamut boundary,

100 lightness levels at 360 hue angles. The mapped point is that which has the smallest

weighted error. Katoh and Ito tested their model with computer generated images and found

that images mapped

according to a ratio of Kl:Kc:Kh of 1:2:2 respectively best matched the

original images. This may be interpreted as lightness is more important to preserve than

chroma or hue in the images used in the experiment.

Fig. 21 shows a comparison of Katoh and Ito's best parameter mappings against the observer

results found in this experiment. Katoh and Ito's results are based on the gamut of the printer

used and is likely to be somewhat different from that used in this experiment. There are
they

marked differences in hue in the blue and green regions. The largest differences in chroma

also appear in the blue region. This is not surprising as the blue region is most non uniform in

69
hue. However, the L*-C*A fit is quite close in a large part of the space. The total AE*^ error

between the predictions and the observer matches is 318. The average error is 12.72.

L*
b pi jr_, rr,.

100
5*103.

>s ~^~Z>t*iig> j^-*

BO

60

40

20

20 40 60 80 100 120 20 40 60 80 100

00

^^-20 7
30

60 ^K^ ^r-7T9 7

40

20

C
20 40 60 80 100 20 40 60 80 100 120 10 20 30 40 50 60

Figure 21. Comparison of Observer results with Katoh and Ito 1:2:2 vectors. Open arrow
heads indicate observer results. Filled arrow heads indicate Katoh and Ito results. Hue leaves
shown are for the hue angle of the predicted match.

In order to see if other parameters existed that fit the data

observer better, the same

optimization technique as described above was used. In addition, an enhancement to the

model was made that gave better fit to the experimental data. This enhancement allows the

L*
ellipsoid to be rotated in the
C*A plane. The enhanced model has the form in equation 3.

70
 ( (
L*-50 L*-50
A =
AC * Sin Kt ArcTan + AL*
Cos Kt ArcTam
C* C*

( t
B= AC*
L *
-50 Vl L*-50
Cos Kt ArcTan AL *
Sin Kt ArcTan
C* C*

B\2 (AH*A2

AE
Kl) +{Kc) +{ Kh

Equation 3. Enhanced weighted color difference model.

L*
The and
C*^ in equations A and B are elements of the target color. Kt is the parameter

that determines the degree of rotation of the ellipsoid.

Optimized model parameters were found for both Katoh and Ito model and the enhanced

model. The models were also optimized in 2D space

(L*
C*^ planes) to remove error due to

nonuniform hue in CIELAB space. Table II summarizes findings.

Model Kl Kc Kh Kt Max error Average

error

Katoh 3d best 1 2 2 NA 48.46 12.72

parameters

3D Katoh 1 2.38 .64 NA 38.24 7.32

3D enhanced 1 3.22 .61 -.41 35.32 6.48
2D Katoh 1 2.66 NA NA 8.07 3.68
2D enhanced 1 3.33 NA -.45 7.98 2.64
Table 3. Summary of optimization results tor Katoh and enhanced Katoh models.

71
The results from the optimization show that the best fits occur when the coefficient of chroma

is greater than that for lightness, but coefficient for hue is less. This implies that observers are

more sensitive to lightness changes than to chroma changes, but they are more sensitive to

hue changes than to lightness changes. This result is somewhat

contradictory to Katoh and

Ito's results that showed less sensitivity to hue shifts. The rotation parameter Kt shows that

rotating the ellipsoid can fit the observer match data better, but it complicates the

interpretation.

100
100
^103.
100

"VO^^iiS- ^119 ^^24^1

138.
SO ^79.5
^^-2<1.7
80
80

/ J-*53.3

60 60
60 V^^yr-^^STl 9 1 .

40 40
40

20 20
20

- C* c*
. ,
c
20 40 60 80 100 120 20 40 60 80 100 20 40 60 80 100

20 40 60 80 100 120 L0 20 3n 40 50 60

Figure 22. Comparison of Observer results with best 2d enhanced Katoh optimization
parameters. Open arrow heads indicate observer results. Filled arrow heads indicate model
results. Hue leaves shown are for the hue angle of the observer match.

Figs. 22 and 23 show the best results from the enhanced model. Fig. 22 shows the results of

2D matches. Notice that this fit is close in all regions of the space. This
very implies that the

72
model
may work
very well when a
perceptually uniform hue color space is developed.

Whereas this enhanced model, as well as the original model fit the data adequately, there are

some problems inherent in a

mapping approach that seeks to minimize a color difference. For

instance, there is a greater

tendency for many to one mappings when target colors lie beyond

the highest chroma values in a given region (colors lying beyond the tip of the hue leaf). This

"flat"
will tend to create spots on an image. A weighted vector
mapping has less tendency to

create
many to one mappings since it is influenced more by static directions in space. Note

that the weighted vector approach can fit the observer match data better than the Katoh and

Ito model, but performs

slightly worse than the enhanced color difference model.

L*
a b plot, narks arrow calls

100
^103.

SO

^0\gilp-^-.]
\\l ry 80

49
60
*>
.'

"7S V: SO 75
"Iki 40

a 20

-100

Ti] 20 40 60 80 100 120 20 40 60 80 100

100

^S^-20.7
80

60 ^Sr^9-7

40

20

20 40 60 80 100 120 10 20 30 40 50 60
20 40 60 80 100

Figure 23. Comparison of Observer results with best 3d enhanced Katoh model optimization

parameters. Open arrow heads indicate observer results. Filled arrow heads indicate model

results. Hue leaves shown are for the hue angle of the model prediction.

73
These results show that both types of gamut mapping models can be used to fit the observer

data The to the effectiveness of a color match approach in

fairly well. question remains as

determining gamut mapping vector directions

2.2.2.1.4 Conclusions to P2 (sub-experiment 1)

The major points of
summary are:

1. Observers don't make simple matches of the kind that are assumed in-gamut mapping

experiments to date. However, three simple models were optimized against the observer

match data and found to fit the match vectors acceptably, and may be used to create a

clipping or compression type gamut mapping that will create

perceptually closest matches

to out of gamut colors. Although the experiment was performed

using solid color

matching for the most part (not complex imagery), this

may also
apply to pictorial images.

2. This experiment shows that there is no multidimensional, statistically significant difference

in any of 25 gamut
mapping decisions over the four different images shown. The entire

data set has been treated as if it came from the same population.

3. CIELAB, arguably the most popular color space for performing gamut mapping

transforms, is not uniform enough in hue angle to make adequate

(perceptually constant

hue) matches.

The experiment
clearly showed that CIELAB is nonuniform in hue. An experiment needed to

be designed to characterize the nonuniformity over more of the color space in a simple,

74
macroscopic manner that can use results to create a transform to a color space whose metric

hue angle is perceptually constant hue.

75
2.2.2.2 Preserving meaning in color matches (sub-experiment 2)
The above work was repeated almost exactly, with the exception of the addition of two new

image (sub-contexts text images), and the change in the

types of
literally representative, and

observer task. The script for this experiment was as follows:

"In the last experiment, if you did it, you attempted to find the closest perceptual match between two colored

images. The criterion for matching was to find the color that made the images look as close to one another as
possible.

In this experiment, you are presented with the same images and colors, but instead of attempting to make a
color match between images, you are to adjust the color of the image on the left to try to make the image

appear to convey the same meaning as the image on the right.

For instance, you may be making a presentation in Microsoft PowerPoint, and the color you want to use is the
one in the image onthe right. Your task is to pick a color from the available color palette that will convey the
same message, or the most similar message as that conveyed in the image on the right. Please don't make an
"name"

adjustment that is too radical to the color of the image on the left, i.e. the basic color or color category
shouldn't change (reds should remain reds, greens should remain greens, etc.).

As an example, I may be presented with a red color as a target color to match. But, the perceptually closest
color is a bit less saturated. In order to convey what I might want to convey with a saturated red (for instance,
danger or attention), I may want to find a color near red that doesn't match in hue as much, but is more
saturated.

In this experiment, you are presented with the same images and colors, but instead of attempting to
make a color match between images, try to find a best substitute for the target color. This may or may
not be the same as the closest perceptual color.

One approach to the problem is to first find a closest perceptual match. Once this is done, try to find a

color that you would use in place of the target color without
changing the color category (or color
name)."

Some observers had difficulty understanding the task. I discussed the intention of the task

with all observers who did not understand what

they were to do. Even when some observers

understood what they were to do, performing the task in a

way that is different than a

perceptually closest match was difficult.

The set of images used in this experiment is shown in figure 24.

76
 12 3 4 5 6 7 8

Literal Figurative Abstract Data Text

Figure 24. Images used in experiment P2, sub-experiment 2.

Twenty five observers performed the

meaning preservation task for each of 5 sub-contexts

and over the same 25 colors as in the previous experiment.

The results from statistical analysis showed that for the multivariate ANOVA (MANOVA),

the null hypothesis of all means

being the same was rejected for 17 of the 25 colors at the 10%

level. Further analysis showed that the text sub-context was the most different from the other

contexts. When the MANOVA analysis was re-run with the text data set removed, the null

hypothesis was rejected for only 5 of the 25 colors. A univariate analysis on differences of data

between individual sub-contexts shows that most univariate rejections happen between the

text sub-context and others. Figure 25 shows a graphical depiction of the colors rejected

across sub-contexts.

77
 12

10

n
Abstract Data

Figure 25. Histograms of colors where the null hypothesis was rejected between sub-

contexts

Although the null hypothesis was rejected for several of the colors across sub-contexts, as

figure 26 shows, the actual value differences of the mean mappings for each context does not

show a
readily identifiable systematic trend. However, when
plotting the pooled mean of all

sub-contexts except text against the text sub-context, several comments can be made. Figure

27 shows mean mapping vectors for text sub-context (open arrow

head) and the mean of the

rest of the subcontexts pooled (filled arrow heads). For many colors, the text
mapping is to a

lighter color than the pooled mapping, and in the green color, the text
mapping is directly

pointing at the origin, whereas the pooled mapping is yellower. The lighter mappings of the

"T"
text sub-context may be due to assimilation effects where the image of the is blended with

the background white color, thus making it appear lighter.

78
 100

^^13.7
90

/ 5
^^-48.

60 y^^f?o.6

40

20

20 40 60 80 100 120 20 40 60 80 100

100

^\^.-21.2
80

60 &efrsi;l->>
6~'

40 ""-* -39.5

20

C
20 40 60 80 100 20 40 60 80 100 120 10 20 30 40 50 60

Figure 26. Mean mapping vectors for each of the 5 sub contexts.

a b plot, l marks arrow tails

00

~'
>^*U
80 -5^"73
'5

60

40

20

20 40 60 80 100 120 20 40 60 80 100

40 7 60

70.4
40

54.7

-56 5 20

20 40 60 80 100 20 40 60 80 100 120 10 20 30 40 50 60

Figure 27. Mean mapping vectors of Text sub-context (open arrow heads) and the mean of all

other sub-contexts pooled (filled arrow heads).

79
2.2.3 Conclusions hading to body of thesis work
Results from these experiments (the P2 experiments) showed that, although color
matching is

an
interesting approach to the gamut mapping problem, it is likely limited in the granularity of

the exercise. Restated, this approach was useful to derive a general gamut mapping model and

to show that CIELAB color space is very non-uniform in hue in the blue region. However,

using the color

matching approach was limited in value when
trying to derive general and

systematic differences in mappings across either image type or task description.

Based on the conclusions drawn from experiments PI and P2, it was felt that attacking the hue

non-uniformity problem had more promise than

following the course set out in the proposal.

Thus, experiments were designed to more fully explore the nature of hue uniformity in color

space, and the meaning and color

matching approach the gamut mapping problem was

abandoned.

80
3 Approach and Results

A search on data sets of constant perceptual hue was performed to sec if there were bodies of

data available that

adequately characterized hue uniformity well enough to not have to perform

the data set generation. Several sources were researched, but very little data was found that

was useful. The most obvious place to start is to investigate the Munsell
renotatio

colors.

Figure 28 shows the Munsell renotation colors of Value 5, plotted on the

a*-b*
plane of the

"circles"
CIELAB color space. Concentric denote colors of
visually equal chroma and curves

emanating from the origin like the spokes on a wheel denote colors of
visually equal hue. At

first glance, this plot indicates that CIELAB color space is

fairly uniform in perceived hue.

-2QQ

400-

Figure 28. Munsell renotation colors of Value 5 plotted on CIELAB a*-b*

plane.
However, upon closer inspection, it is seen that a large proportion of the Munsell renotation

colors were not scaled

by humans, but extrapolated. Figure 29 shows the colors that were

actually scaled
by observers. The colors outside of the squiggly line are all extrapolated (in xy

chromaticity space!). Additionally, figure 29 shows the projection of a typical CRT gamut onto

the CIELAB a*-b*

plane to illustrate the amount of color volume that is outside of the actual

scaled data from the Munsell renotation. From these data, it was concluded that the Munsell

data set was deemed insufficient to use to judge perceptual hue uniformity. One explanation

for the low chroma limit is that the Munsell color chips were manufactured to be very durable

and fade resistant, and were made with a matte finish (which limits chroma). High chroma

paints that were available in 1943 were not durable and fade resistant enough to use for the

Munsell color chips.

82
 400-

* -200
.Q

450-

Figure 29. Munsell renotation colors with extrapolation boundary and CRT gamut

CIELAB
projection onto
a"-b"

plane.

There did exist one data set that was specifically derived to measure hue constancy in CIELUV

3
color space. This was the collection of two data sets from Hung and Berns. Two separate

experiments were performed to find loci of constant perceived hue in CIELUV color space.

The data sets were named CL (constant lightness) and VL (variable lightness). Figure 30

Berns hue data, one for each of the CL and VL data

shows two plots of Hung and constant

sets. The CL data comprise 4 points per hue angle

equally spaced in chroma from max

L* L*
chroma to % of max chroma (all of equal CIELUV lightness, with same as the of the

83
max chroma color). The VL data comprise 9 points per hue angle equally spaced in CIELUV

L*
from 20 to 90. In both cases, the reference color was the maximum chroma color of that

hue angle.

a*

(CD
a"
(VL)
-100 -50 0 50 -100 -50 0 50

100 100
1
i
\ i (

50
s
N
\
\
v
*
\
/
/
t
/
/
/

50
N
\ 1
/
/
b*
o
:-:i17 b*
o
11>

-50
-
i \ \
N
s

-50
-P* ^ v V
N
\

-100
N
\
s

-100
H

Figure 30. Hung and Berns loci of constant perceived hue. Left plot is CL data, right

plot is VL data.

Note that the CL data are smooth, while the VL data appear to be noisy. Notice, in particular,

that there is a crossover in the hue loci of blue and cyan-blue (lower left quadrant). This

implies that the data found are inconsistent, and were therefore judged to be unusable. This

conclusion is supported
by the variance measures reported by Hung and Berns in their article.

The VL data had about 3 times as much variance as the CL data.

This analysis of available data from which to derive surfaces of constant hue in color space left

us with about 48 data points. This is a meager number when

considering the volume of color

84
space occupied
by a CRT display. Thus, an experiment was conducted to find surfaces of

constant perceived hue. The goal of the experiment was to create a database large enough to

be able to derive and test a color space which has a metnc hue angle that represents constant

perceived hue.

85
3.1 Experiment El. Finding surfaces of constant hue in color space

3.1.1 Abstract

A colorimetrically calibrated CRT display was used to measure constant perceptual hue

surfaces in color space. 306 points over fifteen equally spaced hue angles
(every 24 degrees) in

CIELAB color space were sampled. An average of 20 lightness-chroma combinations per

reference hue plane was sampled.

Thirty observers performed the matching task three times

each. Intra-observer variation was used to weight mean observer hue matches for each of 306

colors. Analysis of perceived hue uniformity was performed in CIELAB and CIECAM97s

color spaces. Other constant hue experimental results are analyzed and compared to data

obtained here.

3.1 .2 Introduction

The goal of a color appearance space, among other things, is to provide a structured
way to

model perception. That is, it should have appearance attributes that accurately reflect the

perception of that attribute. Attributes include lightness, brightness, colorfulness, chroma,

hue.56
saturation and

The most studied and best-understood perceptual attributes are lightness and brightness,

especially of neutral colors. This is not surprising, as lightness is the perception that defines

shapes and textures to the greatest extent and it is the most basic. Indeed, the first two colors

in evolving lexicons black

white57

most are and Because of this attention lightness and

86
brightness are
fairly well understood phenomena. have
However, other perceptual attributes

not had as much devotion.

All attributes are not

equally important. Depending on the application, it may be relatively

more critical to For example,

accurately model one attribute. when gamut mapping, if one

wishes to preserve the perception of hue while

altering chroma and lightness, it is quite

important not to change the perceptual attributes of that hue. At the least, the hue name
very

of the color must be maintained.

Hung and Berns showed that for four color appearance spaces, none have accurate

representations of hue neither with lightness nor chroma. Additionally, color spaces have

different nonuniformities. For instance, for CIELAB the cyan-blue to blue-purple region is

the least uniform, whereas for Nayatani's space, the red region shows the least uniformity. It

has been observed

by the author that the nonuniformities in CIELAB color space are severe

enough in the blue region to unambiguously cross color name boundaries between blue and

purple at a given metric hue angle.

If surfaces of constant perceived hue that correctly modeled human perception existed,

presumably a color space could be derived that fit the data. Unfortunately, such a set of data

does not exist in the literature. Munsell and NCS color order systems data were scaled
only for

colors with relatively low chroma because of material limitations. Other experiments have

87
been done to find constant hue surfaces over a larger gamut, but have been too cumbersome

and difficult to get significant results. Hung and Berns give a more detailed analysis
statistically

of these drawbacks.

Berns'
While the Flung and experimental results added more data to the available set, it did not

cover enough of the gamut to enable good modeling. Across 12 hue angles, their color

sampling was limited to 3 colors of equal lightness as the reference color plus 8 colors of

varying lightness along the gamut edge. While the experiment showed that perceived hue is a

function of both lightness and chroma, the relationship of hue with the other two dimensions

was not
fully examined.

In order to complete the description of full surfaces of constant perceived hue, an experiment

was performed that sampled an average of 20 colors over 15 equally spaced hue angles in

CIELAB space. The experiment was designed to be easy to perform without

introducing

elements that might sacrifice precision. Thirty observers performed the experiment three

times each.

3.1.3 Experimental

The iterative Hung Berns deemed be consuming for

approach of and was to too time the

number of colors to be tested. An alternate approach was designed that gave the observer

freedom to select hue matches

using a slider bar that controlled the CIELAB hue angle with a

precision of .5 degrees at all lightness and chroma values. This allowed observers to match the

306 colors of one trial run of the experiment in an average of 90 minutes. Observers were
allowed to make as
many or as few matches as
they wanted in a given sitting. Test stimuli were

presented in a shuffled order for each trial run of the experiment.

3.1.3.1 Script

Each observer was given a script with the same instructions (it was part of the user interface

display so they could refer to it as often as

they wanted). The instructions were:

"Make the hue of the test patch the same as the hue of the reference patch. The same hue

blueness."
means the color has the same amount of redness, greenness, yellowness, or

For the first session each observer was allowed to get acclimated with the interface and ask

questions until
they were comfortable with the task. Some observers had difficulty

between hue matching task finding the minimum color difference

discriminating a and

between the test stimuli and the reference. This did not pose a problem in general, as the

minimum distance and

matching hue usually resulted in the same answer, since the user was

confined to the CIELAB hue dimension. There were conditions due to gamut limitations,

the minimum color difference meant

deviating from a hue match. There were
though, where

four subjects who had this confusion despite verbal attempts to descnbe the task and to warn

then-

about the difference. These subjects also had large intra-observer variance, so

contribution to the overall mean was quite small.

89
3.1.3.2 User interface

20"

The experiment was performed on a Sun Sparc Ultra 1 with a

Sony GDM-20E20 monitor.

The resolution was 1152x900 with a refresh rate of 76 Hz and a 0.31mm-phosphor trio pitch.

The CRT display was kept on

during the extent of the experiment (which ran from May

through July, 1997) to minimize drift from power cycling. The layout of the interface is shown

in figure 31. The stimulus squares for both test and reference subtended an angle of 4.2

degrees.

The white border had a luminance of 71 cd/m2. The luminance factor of the background gray

was .35 (25 cd/m2). The anchor colors showed what the test stimulus would look like if the

slider bar was pulled as far as it could go in either direction. The hue extent of all colors was

+-50 degrees. This amount was a good tradeoff between sensitivity and range of hue angle

based on
preliminary experiments. The user was instructed to make as good a match as

possible, then proceed by pressing the next button. If desired, the user could press the

previous button to see matches that were already made. The user could quit at
any time and

the state would be saved, so that the next session would begin where the last one left off (if the

trial was not completed). The entire experiment was conducted

using the mouse as the only

input device. No keyboard input was used.

90
 ^
10.5"

white border

gray background
reference

175"|

slider bar
gray

anchor

-?
'
( next )

C save j ( previous)

Figure 31. User interface for experiment.

The topmost patch is the reference patch. There were 15 such patches, but they changed at

"next"

every color presentation because the colors were randomized for each trial. The middle

patch is the test color. This is the color that changes when the slider bar is moved. The colors

update as soon as the slider bar is moved, and remain where the user leaves the slider. The

"next" "previous"

match decision is recorded when the or button is pressed. The bottommost

patch is a
gray reference patch that has the same lightness as the current test patch. This patch

was included to be more compatible with the Hung and Berns interface, which also had a
gray

patch. The purpose of the gray patch was confusing to most observers and would not be

included if the experiment were to be rerun.

91
3.1.3.3 Viewing conditions

The CRT was set

up to have a white point near D65, with a luminance of 71cd/m . The

experiment took place under dark surround conditions. The walls of the room were covered

with black felt to eliminate the possibility of reflection from the CRT.

3.1.3.4 Calibration

Gain-Offset-Gamma58
The CRT was calibrated using a variant of the model. First the white

point was found through a search method that altered the digital count of white (it could not

be altered
by analog methods). Then the maximum red, green, blue channels were measured

to find the chromaticities of the primary colors. The gray scale was then measured to find the

luminance-digital count relationship. Ten-bit tone-reproduction curves were made from spline

fitting data from 20 points

along the neutral axis. Forward and inverse models were created,

and the model was tested for inversion error and colorimetric accuracy. The inversion error of

53
the model was less than .001 AE*^. The average AE*ab for a
sampling of colors was

approximately 1, and the worst case error was

~
3 AE*ab units. The CRT was calibrated three

times over the course of the experiment when the average color error became greater than 2.5.

3.1 .3.5 Color selections

306 colors were selected over 15 equally spaced hue angles in color space. Figure 32 shows an

example
sampling of a hue leaf at 0 degrees.

The reference color was chosen for each hue leaf to have moderate to high chroma, and to be

close to the center of the area of the hue leaf. It was felt that the reference color should not be

92
too far from any test color. This in Berns
is contrast to
Flung and who chose the reference

color to have the highest chroma. Figure 33 shows a projection onto the a*-b*
plane of the

entire test color set (defined in CIELAB).

Colors were chosen to span as much of the available color gamut as possible.
Sampling was

done as
uniformly as possible, but gamut restrictions made on a uniform grid
sampling

impossible.
Instead, within the gamut of a given hue leaf, the colors are uniformly sampled,

but at the edges the samples follow the gamut edge more closely. It was also felt that the

density of colors should be consistent around the hue circle, instead of

having consistent

sample period, with less dense sampling in the cyan region.

3.1.3.6 Observers

30 observers, 18 males and 12 females, between the ages of 27 and 59 participated in the

experiment. The author was

pleasantly surpnsed to learn that the lure of a free lunch would

entice so
many willing and qualified observers. Each observer was given a brief color test with

the Ishihara color vision plates. Using this prescreening, four adult males were identified to

have color vision

deficiency and were not included in the experiment. Subjects ranged from

being very experienced color observers to complete novices. Knowledge of the objective of

observers'

the experiment should not significantly influence choices although, as expected,

experienced observers had much lower intra-trial variability.

93
100-,

90-
?

80 -

70- ? ? ? ?

60- ?

50- ? ? ?
? ?

40 -

30 -

? ? ? ?

20-
?
? ?
10 -

n - 1-
H 1 1 1

0 20 40 60 80 100

Figure 32. Hue leaf sampling at 0 degrees. The large square is the reference color.

^tOO-

89-

60.

f
s .40*r

*. 20
?I I * .?

DO 3D _?/!>. 50 100
-
-^-20

*-40

\
-60

-80

409-

Figure 33. Projection of test colors onto

a'-b"

plane. Large squares indicate reference

colors.

94
 3.1.4 Results and Discussion

3.1.4.1 Data set calculation

Each subject performed the

hue-matching task three times. The data were analyzed in

CIELAB Ahj, space. Histograms of the 90 samples of each color were made. Figure 34

shows histograms for the hue surface at zero degrees.

D c It

3e ;a@ !B
l an C to
ZjM

1
:E=JiJ-
:^ilr at
=jt=

~M :fft\ =^=
j jfl ::|^rfiH
ifillai

.i.l.l.l.lll..
nJr
_ilIILi.

Figure 34. Histograms of hue matches for 90 observations at reference hue angle of 0 degrees.
Histograms are plotted in the respective locations in the L*-C*ab plane of the reference colors.

The histograms show that as expected, there is a central

tendency for color matches, and for

the most part, the tails are symmetric. Confidence limits at 95% (1.96 times standard
error)

were calculated for each color from the entire data set of 90 observations. The standard error

of the sample mean is the standard deviation of the sample mean, and is calculated from the

sample standard deviation divided by the square root of the number of samples (fj/V90). This

95
is an estimate of how different the means would be if the experiment were to be run again.

Figure 35 shows the average and standard deviation of 95% confidence limits for each hue

surface.

-avg. conf lim

-
std dev conf lim

24 48 72 96 120 144 168 192 216 240 264 288 312 336 0 deg
Odeg
deg deg deg deg deg deg deg deg deg deg deg deg deg deg

Figure 35. Average and standard deviation of 95% confidence limits of mean hue matches as a

function of reference hue angle. Units of

y axis are in degrees (CIELAB Ahab).

Observers noted that for the cyan region (192 degrees 264 degrees) there was a large degree

of
uncertainty about where to make a match. Many hue locations looked very similar.
Many

also commented that the purple-blue colors were very easy to match and
they had a high

degree of confidence in making those matches. This can be seen from the average confidence

limit plot. The large average confidence limit regions show the most
variability in the match

data, since it was harder to make a match. The 312-degree hue surface is the blue purple in

color and has very low variability.

Variability of data may also be viewed as a tolerance metric.

If there is more
variability in the data, there is a larger area over which a match is acceptable.

96
These data be in
might useful
helping determine suprathreshold hue tolerances, and is in some

ways similar to suprathreshold tolerance data recently Reniff.

reported
by Berns, Qiao, and

Overall weighted mean data were calculated. Weights were derived from
individuals'
inverse

within-observer vanability. A reasonable measure of

variability was needed that gave more

to observers who were more

weight
internally consistent without
favoring a small
minority of

observers all the time. Two metrics were investigated, sample variance and absolute

difference. A good
weighting metric should include more influence from internally consistent

observers without
completely excluding others. The inverse of absolute difference was chosen

over variance, since it was more equitable in weighting observers. Inverse variance
weighting

gave too few observers (between 1 and 5 for most

colors) most of the weight in determining

the mean. Figures 36 and 37 show the histograms of weights for inverse absolute difference

weighting and inverse variance

weighting respectively.

100.00%

80.00%

60.00%

--
40.00%

20.00%

.00%

o o o oooooooooooo

Figure 36. Histogram over the entire data set (306 colors X 30 observation means -

9180) of

weights calculated from inverse absolute difference.

97
 100.00%

+ 80.00%

60.00%

40.00%

20.00%

.00%

oooooooo^
ooooooooo

Figure 37. Histogram over the entire data set (306 colors X 30 observation means
-

9180) of

weights calculated from inverse variance.

If the mean was calculated without weighting, each sample would be normalized
by the inverse

the intra-observer
samples, in this 30, the mean of
of the number of case since we are taking

mean results. With constant weighting, each observation would be normalized

by
~
0.0333.

For the inverse absolute difference case (figure 36), roughly 45% of weights are above 0.0333,

whereas in the inverse variance case (figure 37), only 27% of the weights are above 0.0333.

Weighted mean data with confidence limits are shown in figure 38. The figure is a 2

dimensional representation of the mean hue matches around the hue circle. The x-axis shows

reference hue angle, and the y-axis represents the hue match with +- 95% confidence limits.

The data for each hue is sorted first by lightness (from dark to light), then by increasing

chroma. This means that the overall slope of the set of points that make up a constant hue

surface represents the variation of hue with lightness. The variation of hue with chroma can

be seen as points that make a sawtooth pattern within a set of constant hue points. Note that

98
the 95% confidence limits are smaller than most of the trends on the constant hue
significantly

surfaces.

Note at zero degrees the slope of the points is essentially 0. From 0 degrees to 72 degrees, the

slope of the points becomes increasingly positive, then decreases back to 0 at 144 degrees.

This CIELAB hue for hue

corresponds to the angle constant surfaces
becoming larger as with

increasing lightness. From 144 degrees to 312 degrees the slope of the constant hue points

becomes increasingly negative, then tends back to zero as 0 degrees is approached. The

constant hue chroma relationships are most apparent at 264 and 288 degrees where the set of

constant data points deviate most from a straight line. Chroma relationships with hue are

visualized in a more intuitive way later. Figure 38 is primarily for visualization of the variation

of lightness with constant hue.

99
 350

72 96 120 144 168 192 216 264 288 312 336

reference hue angle

Figure 38. Weighted mean hue matches in CIELAB space with +-95% confidence limits.

Figures 39 and 40 show constant hue surfaces on the CIELAB a*-b*

axes. Figure 39 shows a

projection of the surfaces, and figure 40 shows slices at

every 10 lightness units from L*=10 to

L*=90. The dots on figure 40 are the reference colors.

100
 100
120 96
144 72
48

50
24
7
^k. ^ itH M jff

168
^^
0
b*
0
"^
192

216

-50
240
64^
W. 336

288
312
-100

-100 -50 0 50 100

Figure 39. Projection of constant hue surfaces onto CIELAB a*-b*

plane.

L* L* L*
=
10 =
20 =
30

L* L* L*
=
40 =
50 =
60

1^
^ ^>
L* L*
L* =
70 =
80 =
90

Figure 40. Loct of constant hue in CIELAB a*-b*

at different lightness levels. Dots represent

reference colors.

101
It is clear that the blue region has the most chroma nonuniformity in CIELAB space. These

Figures 41, 42 43 show the same data

data support results from Hung and Berns. and plotted

CIECAM97s6

in the proposed color appearance model (from the CIE TC1-34). The
recently

dots on figure 43 denote reference colors.

350

300

250 -
*v-
,VA

J3
"
200
a)
E
**s.

150

100

50

o mm/mm
24 48 72 96 120 144 168 192 216 264 288 312 336

CIELAB reference hue

Figure 41. Weighted mean hue matches plotted in CIECAM97s color appearance space. The x-

axis is labeled by CIELAB reference hue angle.

The in calculating
cd/m2

parameters used the CIECAM97s coordinates were: LA=14.2 (20%

of luminance of white), Yb=35, X^=95.01, Yw=100, Zw=108.81, and the conditions for dark

surround: c=0.525, Nc=0.8, FL[=1.0, F=0.9. Note that, similar to but not as drastic as

CIELAB, the blue region has significant chroma non-uniformity. The lightness relationship

with hue in CIECAM97s can be seen to be very similar in trend to the CIELAB data.

102
The degree of
nonuniformity can be quantified in many ways. One way is to calculate the

mean color difference from the mean hue angle for each hue surface. This gives a convenient

and simple metric for cross color space comparison. This assumes that there is a reasonable

similarity between Euclidean distances in both color spaces, and that they are equally

discriminative throughout the entire space. While these assumptions are not strictly correct,

the metric gives a feel for a first order approximation of non-uniformity. Figure 44 shows the

mean distance from the mean hue angle for each hue surface for CIELAB and CIECAM97s

color spaces.

100

C sin(h) 0

-100
100

hue surfaces onto CIECAM97s C cos(h)-C sin(h) plane. Hue

Figure 42. Projection of constant

leaves labeled the CIELAB hue angles for comparison.

are by reference

103
 J=10 J=20 J=30

J=40 J=50 J=60

J==70 J=80 J=90

Figure 43. Loci of constant hue in CIECAM97s C cos(h)-C sin(h) at different lightness levels.
Dots represent reference colors.

The CIECAM mean distance is greater than the CIELAB mean distance for 10 of the 15 hue

surfaces. CIELAB distance is greater than CIECAM97s for 2 hue surfaces, and
they have

approximately the same error on the

remaining 3 surfaces. This may lead one to conclude that

CIELAB is more uniform, but that would ignore the fact that the 288 degree hue surface has

considerably more chroma curvature than CIECAM97s.

104
 -CIELAB

-CIECAM 97s

DOtMtDO-q-OOCMtD
iDGi-<--<t<OfD-T-n
*-i-CMC\IC\1CVJOC1

CIELAB reference hue angle

Figure 44. Quantification of average hue non-uniformity for CIELAB and

CIECAM97s.

Another way to quantify hue non-uniformity would be to measure the hue angle difference

between the greatest hue deviations from the mean in both directions. This would give a

worst case metric that might be more

interesting to workers deciding which color space to use

for hue preserving gamut

mapping operations. Figure 45 shows the maximum difference from

the mean in hue angle for each hue surface.

105
 CIELAB

CECAM97S

JJi_JTW\^JUJWT >JJ .'J JJ

tsi >*
r-
, , , i
<\i <M CsJ <M O CO

CIELAB reference hue angle

Figure 45. Quantification of maximum hue non-uniformity for CIELAB and

CIECAM97s.

Note that in figure 45, the entire region from 192 degrees to 288 degrees has higher maximum

error in CIELAB.

Figure 46 shows the projection of hue surfaces onto the opponent axis of linear opponent

color space as defined by: lightness axis =

CIE Y, opponent axis 1 =
X-Y, opponent axis 2 =

Y-Z. Notice that the curvature for the blue region is significantly reduced in a non-

compressed color space. Although this is interesting, it is not suggested that one work in this

type of space just because the curvature of blue is reduced.

106
 100

*
0 (Y-Z) 100

100

-200 -100 0 100

(X-Y)*
100

Figure 46. Constant hue surfaces shown in a linear opponent space (Y,X-Y,Y-Z). Numbers
represent CIELAB reference hue angles.

Figure 47 shows constant hue surfaces plotted in the u*-v*

plane of CIELUV color space.

Notice that the blue curvature is lessened in this space, although visual inspection of the color

space shows clear color name change from blue to purple along a constant metric hue angle

equal to that of maximum pure blue stimulus for a CRT display.

107
 120 96
4<
100

i 72
48

1 24

-100
:% 28?
\
312
3 16

100
100

Figure 47. Constant hue surfaces shown plotted in CIELUV color space. Numbers represent

CIELAB reference hue angles.

Figure 48 shows
Hung and Berns constant hue loci for constant lightness (CL) and variable

lightness (VL) experiments superimposed over the results found in this research. For the CL

loci, the chroma non-uniformity trends are very similar. The respective CRT gamuts are

somewhat different, but the overall agreement appears to be very good. This should not be

surprising, as both experimental designs and presentation medium were quite close. The

varying lightness hue loci were derived for 8 lightness levels on the edge of the CRT gamut at

each reference hue. The VL data match to current findings is not as good as the CL data. In

the blue-purple region, there is some degree of

overlap between hue surfaces. This may be

due to the fact that Flung and Berns only had 9 observers, and as such had a higher degree of

uncertainty in the mean result.

108
 1100 100
v \
1
/
f J /
1
1
50 \\ VI I'M JV
50

b* b*
0 0
fc^*.^:
*
^tff^^Jffi

-50 -50

>

*
-100
1^ -100

-100 -50 0 50 100 SO 0 50

(CL) a*
(VL)
a*

Figure 48. Hung and Berns data plotted on

top of experimental results in CIELAB space.

Left graph compares CL (constant lightness) data. Right graph compares VL (variable
lightness) data. Hung and Berns data are shown in bold dotted lines.

In order to verify that the surfaces derived here are

truly constant perceived hue, additional

work was done. An experiment was conducted to compare constant metric hue surfaces for

perceptual uniformity. These experiments are described in a later section entitled "verification

experiments."

The results from this experiment

apply stricdy to viewing under dark surround conditions on a

CRT screen. It is not clear if the data also applies to bright surround conditions and
viewing

More should be planned that verify that the constant hue

hardcopy output. experiments

surfaces hold under other viewing conditions.

109
Another concern has to do with the gamut limitations of the experimental medium, namely the

CRT gamut. To make these results general, some method of extrapolation should be done to

map the surfaces out to larger chroma values. It would be desirable to extrapolate to chroma

values at least as large as can be created

by current
hardcopy marking technologies. Again,

with extrapolation comes the need to verify accuracy of the resultant data set.

3.1.5 Conclusion

An experiment was conducted to find surfaces of constant perceived hue in color space. 306

colors were tested over 15 equally spaced hue angles in CIELAB color space. Results show

that there is significant lightness and chroma

non-uniformity in both CIELAB and

CIECAM97s color spaces.

Summary error metrics were developed and shown. Results agree

fairly well with Hung and Berns constant lightness data loci of constant perceived hue.

One of the goals of this work was to gather enough data to enable the modeling of a constant

perceptual hue color space that is functionally related to CIEXYZ or an equivalent color

space. This would allow users to transform their colorimetric data to constant hue color space

before gamut mapping, then transform back for color rendering.

110
3.2 Development of constant hue color spaces

Several approaches were taken to

try to develop models of a color transformation that would

result in a metric hue angle consistent with perceived hue. The two approaches that yielded

useful results were the development of neural net based correction, and the development of a

visualization tool that was used to optimize simple, well known transformations. The neural

net approach was useful for generating three dimensional lookup tables that were used to

create transforms between CIELAB and a constant hue color space. The visualization tool

find for
enough"

was used to parameter values a "good tradeoff between appearance attributes

in general and hue uniformity in particular within a well understood structure for color space

transformations.

3.2. 1 Neural Nets

3.2.1.1 General overview

It was initially thought that neural net ttaining could be used as a mechanism to find a general

purpose, invertable color space description that was uniform in perceived hue. This turned out

not to be the case, but neural networks were useful in creating table based hue correction

approaches for use in verification experiments.

The only completely invertable (with no null space) neural net is one in which the hidden layer

has the same number of nodes as the input and output layers. A neural net with one hidden

layer with 3 nodes, and an input and output layer with 3 nodes each is shown in figure 49.

Ill
 Nodes i Nodes k
Nodes j

Figure 49. 3 node, 1 hidden layer neural net

The equation for computation of the numerical value for each of the hidden layer nodes is the

equation H .
,
=
/( V wi Xt + /? ) .
where i is the index for each of the input nodes, and
j is

the index for each of the hidden nodes. The function /was chosen to be tanh, a sigmoidal

function that has a slope of 1 at the origin. This function can be used to model compressive,

expansive, and linear function types depending on where in the domain the input data lies.

Figure 50. Tanh function is expansive in the negative range, linear near the origin, and
compressive in the positive range.

112
The bias term B can shift the domain to the appropriate part of the curve to match the nature

of the nonlineanty. The output node is computed as a weighted sum of hidden layer nodes,

Ok =
Zj wjk *Hj- Th15 is the general model for a neural net functional form, except that the

special case of
having 3 hidden layers This be
makes it invertable. can shown
by modeling the
neural net
by the equivalent functions in matrix form. The matrix form of the 3 node hidden

layer network is shown in equation 4.

w, wn W13 x. B,
H2 w, w22 W23
*
*2 +
B2
w
31 W32 w33 *, B,
Ox W, w\2 W'n "#,
o2 W
21 W'l2 W'23 H2
O, w
31 w\2 "V H3

Equation 4. Matrix form of forward computation of 3 node hidden layer neural network.

The inverted form is obvious from equation 4, so is not shown. The matrices W and W must

be inverted, then the model is run backward.

3.2.1.2 Modeling the CIELAB transformation

To test whether the neural net approach was viable, it was tested on a known transform set.

The 306 color points that were the starting hue colors for the constant hue surfaces

experiment (El) were used to test if the neural network approach could be used to model the

transform from CIEXYZ (D65) to CIELAB color space. The same colors were used to train

and test the network. After training the network for several minutes, the training was

113
completed, and the network was tested. The methods used to build the network and
training

1.59,60'61

algonthms are discussed in detail in references 59-6 The AE^ error from testing the

trained data was 0.07, with a maximum error of 1.52. This shows that a neural network of the

type shown in figure 49 can effectively model the CIEXYZ to CIELAB transformation. This

should not be surprising since the transformation from CIEXYZ can be generalized to be a

3x3 followed followed by another 3x3. The bias term in this trained
by a nonlinearity case was

to fall on the part of the tanh curve that models a compressive function.

3.2.1 .3
Modeling the constant hue data set

The problem in which one knows the function to use to create the transformation between the

tristimulus values and the uniform color space (e.g. CIEXYZ to CIELAB) can be used to test

whether the training algorithms can accurately approximate the function. The problem that

remains once this is shown is to find the coordinates of the uniform color space that

correspond to the tristimulus values of the constant hue data. Immediately the issue arises of

trying to pick appropriate coordinate values of the uniform color space. We know that the

data that are in the same hue surface should lie on a plane, but the correlates of lightness and

chroma are almost

arbitrarily decided upon. It was
initially thought that some feedback

"push"
mechanism could be found for the training algorithm to the data toward a convergent

solution that had good correlates of lightness, chroma and hue. This was not the case for the 3

node hidden layer network. The nature of the neural network is to try to fit the coefficients

(weights) of the model to minimize the error between the coordinates of the input and the

coordinates of the output. But there is no mechanism to tell the output coordinates where

114
they should go to minimize the error.
Additionally, there are no extra degrees of freedom to

use to force the model to account for errors of the output as there would be if the
coordinates,

network had more nodes.

Many iterative approaches were examined that altered the base hue

angle of the hue planes, and used the predicted model to influence the placement of the

coordinates for the next iteration. None of these approaches yielded results. The
interesting

invertable neural network was thus abandoned for use in color space model development.

The more general neural network is not invertable, but is useful in creating a forward

transform that can be inverted using table based

approaches.62

Because neural nets are

for their extrapolation craziness, extrapolation done

notonous was
by hand to C*ab=127 for all

hue angles, for intermediate hue

and angles.
Sixty points were added to the
training data set,

making it 366 points. The neural network consisted of 6 input nodes, three of which were the

input data. The other three input nodes were calculated

using Chebyshev polynomials of the

input nodes. The network had 1 hidden layer with 10 nodes, and the output layer had 3 nodes.

This network was trained and tested. Then a

lookup table was made for both the forward and

inverse directions. The lookup table was made using techniques similar to those in reference

62. These techniques are well known in the industry, and can be purchased off the shelf with a

commercial color correction software package. The AE*ab of the LUTs are shown in table 4.

115
LUT type Average AE*ab error Maximum AE*ab error

Forward transform (CIELAB 0.31 1.79

to hue corrected)

Inverse transform (hue 0.29 1.21

corrected to CIELAB)

Full circle transform 0.06 0.61

Table 4. Transformation error trom hue correction LUTs.

The LUTs were used in experiments

E2(A) and E2(B) to transform data from uniformly

sampled planes in hue corrected space to CIELAB for display on the CRT monitor.

3.2.2 Color space msualisation andparameter

tuning
The neural network model attempted to solve the color space
mapping by hypothesizing the

coordinates of the uniform color space, then to train a simple network to derive the
trying

parameters of the model that fit the transform. The flaw in this approach is that with a simple

mapping between CIEXYZ and a uniform color

space, all three dimensions of each of the

coordinates must be very carefully placed in order to be able to model the transform with no

extra degrees of freedom. This means that the body of points that comprise the coordinates in

the uniform color space must all move together in a systematic way to conform to the

parameters of the transformation.

116
The alternative approach to
solving the problem (creation of a simple, invertable mapping

between CIEXYZ and a uniform hue color

space) is to hypothesize a reasonable transform,

then observe the resultant coordinates after

applying the model. If certain parameters are

altered and the transformed coordinates are visualized in real time, then intuition can be

developed about which parameters make sense to change, and in which directions. This was

the approach that was used in the creation of the uniform hue color space. This is a viable

approach to use to model many color appearance phenomena and attributes, as it gives

immediate feedback, and can be applied to arbitrary sets of transformation data (i.e. data that

has some perceptual characteristic such as constant hue, and corresponding fundamental color

coordinate specification).

3.2.2.1 Model Descriptions

Two major types of color model were approximated with the visualization tool, the CIELAB

type model, and the CIELUV type model. These models are fundamentally different m the

calculation of the opponent data, so they must be separated. The CIELAB type model is

3x3 transforms, a nonlinear and more 3x3 transforms. The

characterized
by successive stage,

CIELUV type model is characterized

by a 3x3 transform, nonlinear compression of the

luminance channel, then calculation followed by subtraction for adaptation, and

chromaticity

multiplication
by the lightness channel.

The CIELAB type model general form is shown in equation 5.

117
 L XD65
M =RxOxPx Y
1
D65

S z 065
_ .

L'=
f{L)
M'=f(M)
S'=f(S)
~P~ "

M'

D = Tx
S'

Equation 5. General form of CIELAB type color model.

Elements of the model R,0,P, and T are all 3x3 matrices with specific goals.
They are

separated in the first multiplication for purposes of clarity and to separate functions. The

function of the P matrix is to provide a transformation from CIEXYZ into a more

fundamental cone tristimulus space. Many of these transformations exist, and most are linear

and can be performed

using a 3x3 matrix. The O matrix is used to support color models

where the opponent step (calculation of luminance and chrominance channels from tristimulus

values) is performed before the nonlinearity, as in Guth's ATD color model63. The R matrix

allows the opponent transformed space to be rotated through 0 degrees through the use of the

matrix in equation 6.

10 0

0 Cos{9) Sin(0)
0 -Sin{9) Cos{9)
Equation 6. Matrix to rotate second and third row elements for the opponent response.

118
 This assumes that the lightness correlate is made to be in the first coordinate, and the

opponent channels are in the second and third

coordinates. The nonlinear functions f can be

in the form of either

fb f2, or f3, and were implemented as a and could take the form of
group,

chroma type compression (convert the second and third coordinates to polar
coordinates, and

compress the radius

value, then convert back), a power function, or a hyperbolic function.

Equations 7 show these functional forms.

fl=Lal;L>=0

/I=-(-L)<T2;L<0

(trt + 1)^
,2
al + L

Ja^H
f2
a2-L

/3=/2(VM2+52)

Equation 7. Functional forms for nonlinear step.

fj and f2 are applied to all channels equally. f3 is applied

by first computing the chroma from

the second and third (M and S) coordinates, compressing with a hyperbolic function, then

converting back to M and S by taking C sm(0) and C cos(0) where C is the chroma value after

transforming through the nonlinear function. Figure 51 shows example plots of functions f,

and f2. It should be noted that none of the nonlinear transforms can
actually model CIELAB

nonlinearities, since
they have the bias term. This was deemed as too complex for the simple

119
model, and since RLAB has been successful without the added complexity, it was decided to

keep it out.

0.5

-1 -0.5 0.5 1 -1.5 -1 -0.5

-0.5

Figure 51. Examples of nonlinear functions. Left plot is function fi, with parameters (Jl =
0.45,
0"2 =
0.6. Right plot is t~2 with parameters al =
1, 0"2 ~
1. Note that at the origin, fi has a slope

of 0.

This model is made more specific depending on whether the opponent stage is performed

after or before the nonlinear compression stage. If the sequence of transforms is like Guth's

model, the relevant transform steps are: P O R [nonlinear], thus the tristimulus values can be

pretransformed to cone values, then the opponent stage can be calculated, then the system can

be rotated (which is not uniform since the nonlinearity follows), and

finally the nonlinear stage

is applied. One could envision further transformmg the output of the nonlinear stage, but no

good function was found for this stage, although it is available.

If the sequence of transforms is like the CIELAB model, the relevant transform steps are: P

[nonlinear] T, thus the tristimulus values are transformed to

primary cone values, then the

120
nonlinear compression is
applied, then the opponent stage is calculated. The opponent, and

rotation matrices are not used.

The CIELUV type model general form is shown in equation 8. The constants krk5 are

parameters that can be changed to alter the shape of the resultant space. The P matrix has the

same effect as in the CIELAB type model in that it can serve to pretransform the XYZ

tristimulus values into a fundamental cone type space. The nonlinearity can be of the form f,

or fr, but not f3, as the second 2 elements of the transformed tristimulus values don't

correspond to the opponent channel data.

L XD65
M = Px Y
1
D65

S 7 D65
_ _

M'=f(M)
*' L

k2L + k3M + kA-S

k5-M
k2L + k3-M + kA S

u
*'

ks
v =

T"

K-y l /Ci Ka

M'-(u'-u'n )
u*
=

M'iv'-v'J
v*
=

M'
L* =

Equation 8. General form of CIELUV type color model.

121
The CIELUV model is somewhat cumbersome from a
complexity standpoint since it does not

fit cleanly into successive stages of linear and nonlinear transformations. Additionally, it is not

a
very intuitively pleasing model because of the translation type of adaptation (although in

hindsight, this was really not needed, as the adaptation is assumed to have been done previous

to the transformation to the uniform color space coordinate system). The inversion of the

CIELUV general equation is shown in equation 9.

]M'=L*

M = f~\M ')
*
v v'

v'= + "
M'

*
,
u
U YU
M'

M{kl-k5-
c
O
k2k5u'-klk3V)
k\-kAV
k5M
u'

k\V
r

XD65 L

Y
1D65
= P_1X M

7 S
- D65.

Equation 9. Inverse CIELUV type equation

3.2.2.2 Visualisation tool development

The visualization tool served three purposes. Firstly, it allowed accurate

colorimetrically

viewing of the uniformity of the color space within the gamut of the CRT used. For this to

work, both the forward and the inverse models described in equations 5 and 8 were

122
implemented. Secondly, it allowed visualization of several psychophysical data sets. Various

data sets were used in order to ensure that the creation of a uniform hue color space did not

adversely affect other attributes of a general purpose uniform color space.

Along with

visualization of both constant hue data sets, quantitative data were calculated. This gave a

more objective
way to compare different prospective spaces. Thirdly, the visualization tool

used a multitude of controls to alter various parameters of the model. The control of

parameters was designed to give intuitive ways to change coefficients, since there is a large

number of them in a given model.

3.2.2.2.1 Visualisation of color spaces

The visualization tool user interface offered three separate 2-dimensional views of color

spaces: Slices of the opponent channels plane at varying lightness levels, slices of the chroma-

lightness plane at varying hue angles, and slices of the hue-lightness cylinder mapped onto a

plane at various chroma levels. The display updates

everything in real time for any change in

any parameter value for the entire model, thus one can watch the space warp as a function of

parameter values. Figure 52 shows examples of the three visualization modes of the interface.

The constant lightness plane can be seen concurrently with either the constant hue plane, or

the constant chroma cylinder. A checkbox is used to control which plane is displayed.

Note, on the constant lightness slice, that there is a circle whose center lies at the origin and a

line that projects from the origin at a 45 degree angle. These are indicators that show at which

hue and chroma values the other 2 planes cut through the space. There are slider bars to

control the lightness level, the chroma value, and the hue angle in increments of 1/100 for

123
chroma and lightness, and 1/360 for hue. There is another slider bar that controls the number

of samples used to display the images. In figure 51, it is set to the maximum, which is 100

square. This does not update

quickly in real time, so the slider bar allows the number to be

varied between 100 and 10 samples per

display plane. There is also an option to
completely

turn off the displaying of color planes altogether.

Figures 53 and 54 show close-ups of the control frames. The constant chroma frame is not

shown in close-up, because it resides in the same space as the constant hue frame. A checkbox

allows the user to toggle between constant hue and constant chroma.

124
Figures 52. Visualization tool user interface showing the three slices of color dimension,
constant lightness, constant hue, and constant chroma.

125
 get it straight

CH 1 + and -

gamma
si
1^ 0 _l

s2 43 0 =
|

CH 2 + and -
gamma

S3 10 0 J

s4 43 0 J
~

CH 3 + and -
gamma
55 10 0 J

SB 43 0 =J
o'

type compression

hyperbolic j
I power

C*

_J
Show CRT primaries

* Show HB dataset
_|
Show big dataset
_|
Show Munsell

_|
Show avg hue

_|
Show map vectors

L*
_J_| Color plane

Lightness 43 0

rotate 0 0 I
prexform opponent

norm XYZ ->

11 NULL NULL J rotate ->
viui\l->
_rj

-rj WB opponer
both scale 100 Xscale 100 \ri

RL 40 0 100CL 40 0 100BL 20
J =
> 10C
Ra;Yd 250 100 J
j
-

reg post 2000 Ga;Zd -250 -2000 -100

Rb;Vn 0 250 ^=s= | 250

reg pre ) Gb;Xn 100 -250 250

J Q1 J Q3 Q5 J Q7 J Q9 Q11 Q13
_J

J Q15
data_.t
_J

Q2 J Q4 J Q6 J Q8 J Q10 Q12
J J Q14
HungBarns

preset xforms load or savei J Constrain Y J Constrain + Ebner

| LAB
max h deviation, max =
26.68, avg =
12.71
LUV

0 24 45 72 96 120 144 168 192 216 240 264 288 312 336

Linearity metric per hue angle.

Qu|t ij

Figure 53. Close up of the controls and constant lightness display.

126
 Show color leaves

360
Hue 238A 0 =

max C 71 20 = 100

samples 35 10 100

L*
linearity I

show ellipses )

show CMFs

J Check for Chroma display 2D sliders j

200
Chroma 0 0 J

Figure 54. the constant hue display.

Close-up of

127
The color planes displays were created by sampling the planes in the resultant color space,

assuming the domain 0;1, -1;1, -1;1. The coordinates that lay on those planes were then

transformed through the inverse model to CIELAB and displayed. Out of gamut colors were

mapped to CIELAB 50,0,0 (neutral gray).

3. 2. 2. 2. 2 Controlsfor parameter tuning

Controls for parameter

tuning fall into buckets according to their use. Some controls
help

tune the pretransform, some controls affect the nonlinear steps, and some controls can change

the posttransform. Each control is described in the following list:

1. Pretransform controls. There is a list of hardcoded pretransforms that are XYZ to LMS

linear approximations to cone primaries from several different authors. Additionally, a set

of 2 dimensional slider bars were used to control parameters for this 3x3 matrix.

1.1. The hardcoded transforms are derived from popular cone fundamental references.

Each matrix assumes as input CIEXYZ D65 values that have been normalized by the

white value of D65 (0.9504 1.0 1.0889) such that the input to the matrices had a range

from 0 to 1 for all channels. This was determined to be the most convenient way to

deal with the pretransform matrix (P), so the rows could be constrained to sum to 1

to maintain response.
gray

1.1.1. FIunt-Pointer-Estevez transform, which was used in a modified form in the

final choice for the final color space, was derived from the D65 normalized 3x3

from RLAB The

model.64

matrix the normalized form of the matrix used in the

128
 model was derived from multiplying the D65 diagonal
matrix
by a matrix whose

elements comprised the D65 white data. Shown in matrix form, this is:

.3804 .7076 -0.0879 .4002 .7076 -0.0808 0.9504 0.0 0.0

-0.2151 1.1653 0.0497 -0.2263 1.1653 0.0457 x 0.0 1.0 0.0
0.0 0.0 1.0 0.0 0.0 0.9182 0.0 0.0 1.0889
The resultant effect on D65 CIEXYZ data is identical. This is to show
simply

how the visualization tool interpreted the data.

1.1.2. Muller and Judd primaries were found in Wyszecki and

Stiles65
on page 634.

For this matrix, all rows summed to 5, so the resultant matrix used was to
simply

divide each element

by 5, so the rows summed to 1. This matrix is:

0.6391 0.4896 -0.128

-0.5091 1.4098 0.0993 This may not be strictly correct, since the

0.0 0.0 1.0

normalized matrix was not derived from a D65 normalized matrix. However,

strictness was not an issue, as these hardcoded matrices were intended to act as

starting points for parameter tuning.

1.1.3. Guth primaries were found in Wyszecki and Stiles on page 646. The matrix was

treated to have rows that sum to 1 by dividing each row by the sum of the rows.

Again, this may not have been stricdy the correct transformation to represent the

129
 data properly, but it was a reasonable transform when visualized. This matrix is:

0.2332 0.8162 -0.0494

-0.4638 1.3656 0.0982

0.0 0.0 1.0

1.1.4. Vos and Walraven primaries were found in Wyszecki and Stiles on page 615.

This matrix was treated the same way as the Guth matrix. This matrix is:

0.2346 0.8213

-0.4681 1.3785 0.0896

0.0 0.0 1.0

1.1.5. Smith and Pokorny primaries were found in Wyszecki and Stiles on page 615 as

well. Unfortunately, they were

incorrectly calculated and seem to come from a

mix of Vos and Walraven and the proper Smith and

Pokorny data. Nevertheless,

0.2346 0.8213 -0.0558

the matrix used was: -0.4637 1.3655 0.0982

0.0 0.0 1.0

1.2. Two dimensional slider bars were implemented to give fine and intuitive control over

the transformation that altered the fundamental LMS space. For each row of the

pretrans formation degrees

matrix, there are two of freedom, since the rows were

constrained to sum to 1 to preserve gray response. Each of the slider bars allowed the

changing of one of the LMS tristimulus values (after the 3x3 matrix). Thus, with this

control one can

independently alter the shape of the color
matching function of the

130
 "primary"
cone response. Figure 55 shows an example picture of the 2D slider bars.

Figure 56 shows an example of the transformed color

matching function, first

through the identity matrix (thus the result are CIE color
matching functions), then

through the Hunt pretransform.

Very fine control can be exercised over the shape of

the transformed curves with the 2D slider bar controls.

rrJ 2D sliders for pre mat

red

green

blue

bars''
the pretransformation matrix.
Figure 55. Two dimensional "slider that controlled

131
 -if
^ Color Matching functions ^ Color Matching functions

Figure 56. Examples of CIE 1931 color

matching functions transformed through
pretransformation matrix. Left image is through an
identity matrix. The right image is
through Hunt-Pointer-Estevez matrix.

The choice of pretransform matrix has a significant effect over the resultant shape of the color

space. Take, for example, the following senes of figures, together labeled figure 57. Each of

the color
matching functions is shown next to the resultant color space shape. The only

difference in each of the color spaces is the pretransform matrix, which in turn affects the

shape of the CMF and color space. Appendix C shows the for
corresponding parameter set

each of the figures respectively.

132
I I

Figure 57. Three examples of CMFs along with their influence on the shape of the

color space. Color space images show

Hung and Berns CL data set and Munsell V5.

133
2. The O matrix controlled the opponent step for the Guth type model. This type of model

performs the opponent calculation before the nonlinear compression stage. The opponent

transform was calculated using only 2 rows of the matrix, as the third was reserved for the

lightness response. Since, for opponent response, the corresponding rows must sum to 0

to ensure grays lie at the origin, only 2 elements of each row need to be controlled. This

control was offered through the use of 4 slider bars that changed the first 2 elements of

each of the opponent rows.

3. The R matrix controlled the rotation

step for the Guth type model. Rotation matrix

elements are shown in equation 6. One parameter, the angle of rotation, is needed, and

was provided for through the use of a slider bar. Rotation of the domain of the Guth type

model is not uniform, since the nonlinear compression

step follows the opponent and

rotation steps. The effect on the shape of the color space in this type of model is to warp

hue lines away from the axes and toward the diagonals (axis plus 45 degrees). This is

because the slope of the nonlinearity is greatest at the origin. This effect is visualized in

figure 58. This type of model was not given much consideration because it worked only

with a nonlinear compression stage that had a slope near 1 at the origin (hyperbolic type),

thus the lightness response deviate quite far from CIELAB. Additionally, it was very

difficult to find rotation and opponent parameters that showed improved hue constancy

over current spaces (CIELAB, CIELUV). Note, in the figure, the blue constant hue locus

(for Flung and Berns data set) has a different shape on the left figure than on the right.

134
 This is due only to the opponent channels through 45 degrees. The
rotating an angle of

shape deformation is due to the

following nonlinear functions being applied.

Figure 58. Example hue lines from Guth type model. The first image is rotated at an angle of 0
degrees, the second is at 45 degrees. Note the shape difference of the constant hue loci.

4. Nonlinear controls were presented

using a series of 6 slider bars, each
controlling the

positive and negative parts of the curves respectively. Figure 59 shows an image of the

nonlinearity controls. Figure 51 shows an illustration of how the controls can make

various curve shapes. For the power function, the parameters would be 45 and 60 (both

are divided by 100) and for the hyperbolic function, the parameters would be 10 and 10

(both args are divided by 10 to make a reasonable scale range). The selection range on the

135
 interface was 0 to 100.

CH 1 +and -gamma

0
=J

si 38A
s2 43 ,
0 =j

CH 2 + and -gamma

S3 37 ^
0 =J

s4 43 ,
0 =J

CH 3 + and -gamma

0
=J-

S5 BO sas

s6 43 ,
0 =J
o'

type compression

hyperbolic |
power

C*

Figure 59. Xonlmear controls for visualization tool. Six parameters correspond to the positive

and negative parts of each of the three channels. Functions take the form of hyperbolic, power

function, or chroma compression.

5. Posttransform controls included 4 opponent channel controls, and 3 lightness controls.

The opponent channel controls are intended to influence the shape of a CIELAB type

model where the opponent calculation is done subsequent to the nonlinear function. Since

rows of the opponent channels sum to 0 to make grays lie on the origin, only 2 parameters

are necessary. The 3 lightness controls allowed experimentation with how much of the

three channels should be used to create the lightness response off the gray axis.
By forcing

the sum to equal 1, the neutral response can be constrained to be simply a function of the

nonlinearity. Thus, the RLAB response can be exactly mimicked on the neutral axis as

136
 long as the
nonlinearity is power function with gamma =
0.43, and rows sum to 1 for the

pretransform, and the lightness row sums to 1 in the posttransform.

6. The last control that interface

through the
apply linear
was offered user was the to
ability

regression to either the pretransform or the posttransform matrix. The data

matrix, used

was the Ebner and Fairchild data set (306 points of constant hue on 15 different hue

surfaces). One could use the interface to include or exclude of the 15 hue surfaces
any

through checkboxes. Of course at least one surface is needed to provide data to the

regression engine. The regression model data needs pairs of triplets to perform the

optimization on. The first set of triplets was made by taking the CIEXYZ input data,

transforming it to the appropriate coordinate system for input. For the pretransform case,

this was simply the CIEXYZ values. For the posttransform case, the data had to be

processed through the pretransform and the nonlinearity. The second set of triplets was

made
by processing the input data through the entire model, then calculating the average

hue angle for each hue surface. The average hue angle then replaced the actual hue angle

for each color, then the data was processed through the inverse model to the appropriate

point. In the pretransform case, the data was transformed through the inverse

posttransform and the inverse nonlinearity. In the posttransform case, the data was used

directly.

Multiple linear regression was applied to the input-output triplet pairs to generate new 3x3

matrices for the model. This technique worked reasonably well if it was applied
only once

or twice, but the 3x3 matrices diverged when the regression was applied successive times.

137
 3.2.2.2. 3 Quantitative metricsfor hue uniformity andpsychophysical data set visualisation

Several data sets were investigated in the development of the color space. These data sets were

used to verify that alterations made to the color space did not
adversely affect other attributes

"best"
of the space. This also makes the decision of which parameters are much more

complex, since it is not easy, or even feasible to assign quantitative metrics to all the data, let

alone optimize a space with respect to all attributes. Each of the data sets will be discussed

here along with the intent and the source. Only the constant hue data sets and the chromatic

lightness data set were quantified

(Hung and Berns, Ebner and Fairchild, Pirrotta and

Fairchild). Quantification of the constant hue data sets took the form of three different

metrics. For the visualization tool, the metric used was the most severe: For each set of

constant hue data, the absolute difference between the minimum and maximum hue angle was

calculated. The other 2 quantifications are holdovers from the calculation method used in the

publication of reference 2. These quantifications found the mean and maximum absolute

difference (in hue angle) from the mean hue angle for each constant hue set. Thus there are 3

ways to compute a hue uniformity metric.

Fairchild2
1. The Ebner and data set was used to derive one of the two
uniformity metrics,

and the 306 points were displayed on the interface. The data were originally derived in

native form as CIELAB D65 data, so

they were converted to CIEXYZ D65, then

normalized
by the white point coordinates (0.9504, 1.0, 1.0889). This data set can be

found on the MCSL web site at

http://www.cis.rit.edu/people/facultv/fairchild/CAM.html.

138
 Berns"
2. The Hung and constant lightness data derive
(CL) set was used to the other

uniformity metric. The 48 points were displayed on the interface, an example of which is

shown in figure 58. The data were copied out of the paper (table III), then normalized by
the white point coordinates for in the The data
use tool. original were derived under D65

surround.

3. Munsell Value 5 data were obtained from the RIT VAX computer. The Munsell data was

"circles"
used to scale the x and and to minimize of the chroma in
y axes, eccentricity the

destination color space. These are the Munsell renotation data that can be found in

Stiles65
Wyszecki and on pages 840-852 as well as in the original reference66. These data are

originally in CIE xyY

chromaticity coordinates under Illuminant C. The data were

RLAB67
converted to CIEXYZ D65 using the most recent equations
assuming average

surround, and
hardcopy viewing (D=l, Yn=200 cd/m2, rj=l/2.3). The data were then

normalized
by the white point for use in the tool.

4. Lightness response
along the neutral axis was chosen to be nearly identical to the RLAB

color space, which is very nearly the same response as CIELAB. Lightness response was

L*
plotted as a function of CIELAB in the visualization tool.

5. Lightness of chromatic object color data were used to make the new color space more

accurately represent lightness of colors off the neutral axis. These data were copied from

the
reference68

(Table I). These 36 points were reported in CIELAB L*C*abh space, with

an illuminant near D65. The data were converted to CIEXYZ D65 assuming D65

139
 illuminant. These data quantify the Helmholtz-Kohlrausch effect, which states that

chromatic colors appear lighter than neutral colors of the same luminance factor. Since

CIELAB uses only the luminance information to derive the lightness response, it cannot,

By combining the from other certain

by definition, predict the effect. responses channels,

aspects of the effect can be reduced, but it was found that this simple model does not

predict the Helmholtz-Kohlrausch dependence on chroma, but can reduce the hue

dependence that CIELAB exhibits.

6. MacAdam Ellipses were calculated for observer PGN's data. These data are

discrimination data that show threshold sensitivity to the dimensions of CIE xyY

chromaticity space. In a
perfectly uniform color space, the ellipses should resolve to

circles, although threshold information is less interesting in a uniform color space than it is

in a color difference space. These data are from Wyszecki and Stiles on page 309. These

data are defined in CIE xyY space under illuminant C. The ellipses are defined using

major and minor axis lengths, center coordinates (x and y, all data were equiluminant), and

rotation in degrees. For each ellipse, 20 samples in CIE xyY of the perimeter were made.

A point on an ellipse can be represented

by x a Cos((j>),and, y =
b Sin((j)) ,
where a

and b are the axes of the ellipse. The point (x'y') can be found by rotating point (x,y)

through 9 degrees by applying the rotation rule

x'=x-Cos(9) +
ySin(Q);y'=-x-
Sin(6) + y-Cos(6). The resultant points can be

x"

translated to sit around the origin of the ellipse

by applying the translation rule =
Ox -

140
 y" y'

x'> Oy By 9 degrees
-

the ellipse, then

-

sampling rotating to then translating to the

origin of the ellipse, samples can be made in the color space. The sampled points are then

transformed to CIEXYZ D65. The data for in

were also normalized
by the white point use

the tool.

7. Suprathreshold difference
ellipses69

from (Table
color were copied the reference II, Cross

Section for AL* =

0). These data are derived from color differences that are more than just

noticeable, so are
possibly more important to get closer to circular than are the MacAdam

ellipses. These 19 ellipses were sampled in a manner similar to the MacAdam ellipses.

Since the data were

already in CIELAB using Illuminant D65, they were converted
simply

to CIEXYZ D65. The data were also normalized

by the white point for use in the tool.

8. OSA Color System medium

gray (L
=
0) constant lightness plane
scaling data is shown.

These data were copied from Wyszecki and Stiles (reference 65), pages 870-871. The data

were expressed in CIE xyY (D65), and converted to normalized XYZ for use in the tool.

functions"
9. In addition to data sets, the "color matching of the CIE color
matching

functions are visualized. What are called color

matching functions are the CIE color

matching functions after

being transformed through the first 3x3 matrix. This response

function can be thought of as the response of some cone fundamentals that are a 3x3

matrix linear transform away from the CIE 1931 (2) color
matching functions.

10. The spectral locus is plotted in the color space for reference.

141
3.2.2.3 Properties of the chosen color space (IPT)
The IPT color space is named such that its coordinates have some degree of
relationship to

the meaning of the dimensions. The lightness dimension is denoted as I, which can be loosely

related to the word intensity which provides a clue to it's meaning. The red-green dimension is

"dominated"
denoted as P, which can be related to the fact that it is
by the red response

(protan) and is the dimension lost by protanopes. The yellow-blue dimension is denoted as T,

using the same argument for the tritan response. IPT is also short for Image Processing

Transform since it is useful for transformations such as gamut mapping.

3.2.2.3. 1 Engineering Considerations

The IPT color space was chosen for a number of reasons. The primary reason is that it has a

metric hue angle that more

closely represents perceived hue. Additionally, it has been designed

with some
engineering constraints in mind. The biggest engineering constraint is that it

should, for the popular CRT gamut, not be discontinuous in hue. That is, it should not, for a

given hue angle, for a CRT type gamut, have in-gamut colors at a larger chroma than out-of-

gamut colors. This situation can happen if the color space is warped appropriately, and it can

happen with a space that accurately models constant perceived hue. It also happens to occur

in the blue region of the color space (I think coincidentally). The following scenario illustrates

the point.

If one were to create a gradient

sweep from neutral
gray to device blue, this would result in a

straight line in chromaticity space. Consider a typical

primary phosphor set with chromaticity

coordinates of Red: 0.630 0.340, Green: 0.310 0.595, and Blue: 0.155 0.070 with a white point

142
of D65. Figure 60 shows a line connecting the blue primary to a point on the neutral axis, lets

say a point that is 0.2 multiplied

by the white point (near L*=50). If one creates a color

gradient
sweep from device blue to a neutral gray, or to white, it will be immediately apparent

that the middle chroma colors are

decidedly purple in color. This is illustrated in figure 61 by

plotting the path of the gradient sweep in CIELAB color space

along with lines of constant

hue. Here, the Hung and Berns loci are used because they are simpler to visualize.

0.10.20.30.40.50.60.7

Figure 60. A line showing the path of a gradient sweep from gray to the blue primary (in CIE
xyY space).
chromaticity

Figure 60 establishes that the path from neutral

gray to primary blue using a linear mixing rule

fact line in makes colors that

(as evidenced
by the that it makes a straight chromaticity space)

appear purple in the mid-chroma region. The explanation of

why this can cause
discontinuity

CRT description in a uniform hue color space is illustrated in figure 62. This
in the gamut

CRT blue in the new IPT color space (to be

shows the path from neutral
gray to primary

described in full shortly).

143
 m

100

N '
'
/ /
W /
0

Vv

100 Vv

-100 100

b*
Figure 61. Path sweep from gray blue primary in CIELAB Notice
a*

of a gradient to space.

that the mid chroma colors are more purple than the constant hue locus. Dotted lines are Hung
and Berns constant hue loci.

Note the prominent leftward curvature of this path. Note also, that it accurately describes the

perception that the path is more purple in the mid-chroma region, whereas if one would

believe that CIELAB was uniform, it would appear that the CRT primary actually becomes

bluer in the mid-chroma region. From Figure 62, it is not hard to imagine a case where the

curvature of the primary blue signal would become so extreme in certain color spaces, that, at

a line projected from the origin toward a direction near the primary blue signal, the colors

could go out of gamut, then come back into gamut as chroma is increased. This is, in fact,

exactly what happens when a color space is designed that exactly straightens the Hung and

Berns CL blue locus. Because of the danger of

discontinuity of gamut at a constant metric hue

144
angle, it was decided that the color space should be explicitly designed to not exhibit the

artifact.

0.5

-0.5

Figure 62. Path of a gradient

sweep from gray to blue primary in IPT P-T space. Notice that
the mid chroma colors are more purple than the constant hue locus. Dotted lines are Hung and
Berns constant hue loci.

Thus, the IPT color space does not exhibit hue discontinuity at specific metric hue angles. It

does this at the expense of getting the Hung and Berns constant blue locus exacdy straight.

However, as will be discussed later, it was judged more uniform than either of the constant

hue data sets, so the slight bend in blue is not observable.

Other design criteria were that the color space should be as simple as possible and still do a

good job. The final model that was decided upon is extremely simple. It is also
very amenable

general transformation such as the PostScript imaging

to transformation by engines, model,

145
which supports conversion
using cascaded sequences of 3x3 matrices and one dimensional

non linear steps. The decision to use a power function, as in RLAB, simplifies the model

when used in a high accuracy scenario (data not cached into quantized
lookup tables), as it

needs no conditional statements to treat low luminance colors.

All the following visualizations and quantifications are shown with respect to CIELAB color

space. This is because CIELAB is the most popular gamut mapping color space, and the

order of
complexity between the two spaces is nearly equivalent.

3.2.2.3.2 Model and Coefficients

The IPT forward model is shown in equation 9. Equation 10 shows the inverse model.

L 0.4002 0.7075 -0.0807

^D65
M =
-0.2280 1.1500 0.0612 Y
1D65

S\ [ 0.0 0.0 0.9184

.
7
D65.

L'=L043;L>=0

L'=-(-L)043;L<0

M'=M0A2;M >=0

M'=-(-M)043;M <0

S'=SOAi;S>=0

S'=-(-S)OA3;S<0
L'
0.4000 0.4000 0.2000
M'
4.4550 -4.8510 0.3960
5'
0.8056 0.3572 -1.1628

Equation 10. Forward IPT model and coefficients.

146
 L 1.8501 -1.1383 0.2385
M'
=
0.3668 0.6439 -0.0107

S'\ [ 0.0 0.0 1.0889

L =
L'23256;L'>=0
L =
-(-L')23256;L'<0

M =M'23256;M'>=0

M=-(-M')23256;M'<0
S =
S'23256;S'>=0
S =
-(-S')23256;S'<0

X D65 1.0 0.0976 0.2052 L

D65
1.0 -0.1139 0.1332 M

JD65 1.0 0.0326 -0.6769 S

Equation 11. IPT inverse model and coefficients.

3.2.2.3.3 Color Matching Functions

What are called color

matching functions of the color space are the CIE color
matching

functions (xbar, ybar, zbar) transformed through the first 3x3 matrix of the model as a

function of wavelength. These can be thought of as the fundamental signals that are then

compressed and coded into opponent signals. The color

matching functions of the IPT model

deviate only slightly from Hunt-Pointer-Estevez

primaries64

the that are described on page 120.

Figure 63 shows the IPT color

matching functions overlaid on
top of the Flunt-Pointer-

Estevez primaries. Note the only difference is in the middle wavelength

(green) response

function. This function has more responsivity at the left tail, and less responsivity at the peak

than does the Hunt-Pointer-Estevez function. Note that the resultant yellow-blue response is

probably very sensitive to the amount of

energy in the tail of the green response because it is

147
 near the wavelength of the peak of the blue response. This results might be evidence that
very

the functions for transformation to the primary cone response might be improved using an

alteration of this nature, as the conversion to a (metric) uniform color space is likely to mimic

the human response function to the degree that it is perceptually uniform.

The alteration was made to the pretransform matrix because it noticeably straightened the

constant hue loci for the blue region in both constant hue data sets. For example, the Hung

and Berns absolute maximum deviation metric for the Hunt primaries (all else
remaining

constant in the model) was 10.3, while the metric was 7.9 for the IPT color space (altered

primaries). Although this may not seem like a large difference from the metric, the visual

stimulus clearly contained purple in mid-chroma with max chroma at the blue primary, while

with the IPT parameters, no purple was observed

(by the author). This result is corroborated

through the visual verification experimental results.

148
 HuntL

Hunt M

Hunt S

IPT L

IPTM

IPTS

co^-^-'^-'^-LnLnLncDCDCDi^t^

wavelength (nm)

Figure 63. Color matching functions of the IPT color space compared to the Hunt-Pomter-
Estevez primaries. Note the larger response on the left tail of the M (green) response function.

3. 2. 2. 3. 4 Constant Hue data sets

Figure 64 shows an overlay of both constant hue data sets. The left image shows constant hue

data sets in IPT, and the right image shows the surfaces in CIELAB. Hung and Berns

constant lightness data is shown as bold dotted lines. The IPT plot appears to have different

scale factors for the blue (lower left) Hung and Berns line vs. the Ebner and Fairchild surfaces.

This is an artifact of the the sampling was done in CIELAB space for the constant hue
way

experiment. Notice on the CIELAB plot, that the Hung and Berns line for blue (the most

farther Ebner
quadrant) lies over, but
curved line in the lower nght extends much than the and

Fairchild surface at the same location. These colors were initially sampled in a more cyan

position than the device blue, so were limited by the gamut of the CRT display.

149
 1 100
I

V
> -. i
il i \
*

50 V N /
V /
^ 1 /
Il I/ y 1
f
> 50
AT *

xv w>y
\A
^-^^^A ==^s="

p** -

T*100 0 0
b*

~^&af Wfck **S!v77^>J

"yt^i

J/, *
-50

1
-50
1
* X
s
/ Ik

-100
-50 0 50
-100 -50 0 50
P*100
(CL) a*

Figure 64. Constant perceived data sets plotted in IPT (left) and CIELAB (right).
Berns53
Constant Lightness (CL) data is in dotted lines. IPT is
Hung and set

plotted with P and T scaled

by 100.

Notice that, in the blue region, the hue loci are straighter in the IPT space than in CIELAB.

Furthermore, the hue loci in the rest of the space appear to be as straight in IPT as in

CIELAB. Quantification is shown for CIELAB, CIECAM97s (just because it was already

done, so I didn't remove it), and IPT for the mean distance from the mean hue and the

maximum absolute distance from the mean hue in figure 65. Clearly, the mean distance from

the mean hue is smaller over the entire space than is either CIELAB, or CIECAM97s. The

maximum absolute deviation from the mean hue angle is significantly smaller in IPT than in

either CIECAM97s or CIELAB in the blue region. There are 4 reference hue angles where

the IPT color space has a larger maximum deviation from the mean hue angle. These

hue 144 168 (greenish cyan), 312 (magenta-

reference angles are (green), (magenta), and 336

150
red). The absolute magnitude of hue nonuniformity is quite small for 144 and 168, and while

the magnitude of the metric is larger at 312 and 336, hue nonuniformity was not observed.

Interestingly, these two reference hue angles had the lowest average confidence limits (see

figure 35). As will be seen later, IPT was not judged less uniform than either constant hue data

set at
any of the reference hue angles.

CIELAB reference hue angle CIELAB reference hue angle

Figure 65. Quantification of hue nonuniformity for CIELAB, CIECAM97s, and IPT.

The visualization tool used a different metric to calculate hue uniformity. It used the most

maximum absolute deviation of hue between all points in a locus. The

stringent metric, the

calculation was done both on Ebner and Fairchild's data set (as in figure 65), and on
Hung and

Figure 66 shows those quanitifications. Because the visualization tool did

Berns (CL) data set.

implement the CIELAB model, it is approximated (and labeled CIELABEQ, for

not exacdy

power of 0.43 instead of the cube root and

CIELAB equivalent) by using a nonlinear

conditional statement for low luminance.

151
 Hung and Berns max. deviation Ebner and Fairchild max. deviation

25-

20-

DIPT a ipt
5 15-
CIB_AB_EQ
a CIB-ABEo|
10-
J 1

TfrhKHh^li TT ^fcffilfflii
V <0 (M IO O
O
R RY Y VG G GC C CB B 8M M MR

hue

Figure 66. Maximum deviation quantification used in the visualization tool.

This quantification also

clearly shows IPT's improved hue uniforrnity in the blue region,

Berns'

especially with
Flung and (CL) data.

3.2.2.3.5 Munsell Value 5 data

Munsell renotation data is plotted to show that the chroma response of the space is not

terribly different than that of CIELAB. No quantification has been done of the eccentricities

of the constant Munsell chroma loci. Figure 67 shows the Munsell value 5 data plotted in IPT

space on the left and in CIELAB on the right. The CIELAB plot appears to be slightly more

uniform in the low chroma circles, but neither appears to be perfect, and neither appears to be

grossly bad. This is a possible area of improvement of the color space.

However, one would

need to be careful to only use the non-extrapolated data to tune the chroma response of the

space. Other chroma data sets would be interesting to find as well to corroborate the Munsell

data base.

152
 1
150
0 .75

100

... 'a.'.s .Vl.s

-
* * ;

i.i=V>Vin::::
:
'25-
-1-0.75-

O'.'&Q :.%$--l'<l-
0 . 5 0.75 1 -150 ~iqcr -.5.&.;-.n
?$?:; '.'loo 150

-
0 , 5 [
-100

0.75

-150

-1

Figure 67. Munsell Value 5 data IPT CIELAB

plotted in
(left) and
(right)

3.2.2.3.6 Neutral Tightness Response

The CIELAB lightness function the comparison metric for

was used as
judging appropriate

L*
lightness response. CIELAB is known to be very representative of the lightness response

L*
of humans under average surround conditions. The CIELAB function was mapped as

closely as possible while

maintaining a simple power function. The final response exponent,

0.43 is very close to, but not

exactly the same as the RLAB response for average surround,

which is 1/(2.3), or 0.43478. If desired, for coherence, the exponent could be changed to

1/(2.3) with
virtually no effect on the response of the color space. Figure 68 shows the IPT

L*
lightness response as a function of CIELAB Note the major difference in the response is

at the dark end of the curve. The difference is due to the fact that the slope of a power

function (power between 0 tends toward infinity as the value goes to 0.

and 1)

153
 L*
.CIELAB

-Ix 100

OTj-OOCMCM^-OOCD'^^t
OOT;'!,'oj<NCOCO'd:

i-
CM CT IO CO I"- OS

L*
CIELAB

L*
Figure 68. IPT lightness correlate as a function of CIELAB

3. 2. 2. 3. 7 Chromatic Tightness Response

The chromatic lightness response is modeled on data derived from Fairchild and Pirrotta. The

Wyszecki data set referenced in that paper was not investigated, but might be used in further

refinements of the model. The observed lightness response of chromatic colors was measured

by having observers make lightness matches between chromatic colors and neutral colors. The

chromatic lightness response is a measure of how well the color space predicts the lightness of

chromatic colors. To show this, a comparison is done between the lightness response (in the

color space of
interest) of the chromatic colors to the lightness response of the neutral colors

that were judged as matches. Thus, in a color space that had perfect chromatic lightness

response, the metric lightness of chromatic and neutral colors (which were judged by

observers as
having equal
lightness) would be the same. This is best explained
along with an

L*
illustration. In figure 69 the CIELAB of a neutral color that matched a chromatic stimulus

154
is plotted as a function of the CIELAB L*
of the chromatic stimulus. If CIELAB had a

perfect chromatic lightness response, the data would fall on a straight line with a slope of 1.

L*
Neutral

L*
CIELAB
30 40 50 60 70 80 90

L"
Figure 69. Chromatic lightness response of CIELAB. x axis is of chromatic stimulus, y axis is
L*
of neutral stimulus that matched in lightness (mean observer judgement). RMS error between
L*
observed lightness and of chromatic color is 7.12.

L*
The RMS error between the observed lightness of
gray stimulus and the CIELAB of

matching lightness is 7.12. When a simple model (which is a function of

L*) is used to predict

observed lightness, the RMS error can be reduced.

L*
Neutral

predicted

30 40 50 60 70 80 90
L*ew= L'
Figure 70. CIELAB matching L* response as a function of predicted + 0.098 C\b.

RMS error between observed lightness and predicted is 5.97.

155
 L*
If a linear regression is applied to the model L*new= + k C*ab, the constant k is found to be

0.098. The plot of this simple prediction function is shown in figure 70. Note that the match

points are closer to the straight line. The RMS error for this simple prediction is 5.97. By

adding enhancements to the predicted Lightness model, the RMS error can be reduced to 4.2.

The form of this most elaborate model is denoted L**, and is a function of lightness, chroma,

and hue of the chromatic color. Equation 1 1 shows the entire lightness prediction model and

its parameters.

L** =
L*+f2(T*)-f\(h)C

.. (h -90
f\(h ) -
0.116 Sin + 0.085

2.5-0.025- T
/2(L*) =

Equation 12. Final lightness predictor for chromatic colors from Fairchild and Pirrotta68. RMS
error between observed lightness and L**(L*,C*ab,h) is 4.2.

156
The same data is shown plotted in IPT in figure 71. Although the RMS error is
only slighdy

smaller than the RMS error in CIELAB space, notice that the layout of the points falls along a

path that is straighter than the CIELAB data points. This allows the creation of a more simple

prediction model for lightness of chromatic colors.

Specifically, the RMS error can be reduced

to below the error of the best CIELAB function dependent

predictor without the
being on

hue.

Neutral I

0.9
0.8
0.7

0.6

0.5
0.4 m*

0.3

IPT I
0.3 0.4 0.5 0.6 0.7 0.8 0.9

L* L*
Figure 71. Chromatic lightness response of IPT. x axis is of chromatic stimulus, y axis is
of neutral stimulus that matched in lightness (mean observer judgement). ). RMS error between
L*
observed lightness and of chromatic color is 6.73.

157
 model Inew= I + k done find k
Using the C*IPT, a regression was to the value of =
0.202. The

RMS error of this predictor is 4.35, which is

nearly the same value as the CIELAB predictor

that is a function of L*,C*ab, and h (RMS error was 4.2). The equivalent of
2(L*) was used to

enhance the prediction, which is shown in its full form in equation 12.

Ipnd + /(/) 0.2024-

=/
C*,PT
/(/) =
2.5 -2.5-7

C*IPT=Jp2+T2

Equation 13. Final lightness predictor for IPT color space for chromatic colors from Fairchild
and Pirrotta68. RMS error between observed lightness and predicted is 3.5.

Figure 72 shows the plot of the final model of a lightness predictor for chromatic colors in

IPT. Note that the RMS error is lower, and the model is simpler because it has no hue angle

dependency.

Neutral I

predicted
0.3 0.40.5 0.60.7 0.80.9

Figure 72. Plot of final lightness predictor for IPT color space for chromatic colors. RMS i
between observed lightness and predicted is 3.5.

158
 3.2.2.3.8 MacAdam (PGN) Discrimination Ellipses and Spectral Locus

Although the intention of a uniform color space is to accurately model appearance attributes

(not differences), it is how differences are

color nonetheless
interesting to see small color

mapped in the color space. Therefore, both discrimination (at or near color difference

threshold)and suprathreshold color difference data are investigated. Figure 73 shows the

MacAdam ellipses plotted on IPT space (left), and CIELAB space (right). Along with the

MacAdam ellipses, the spectral locus, monochromatic equienergy signals, are plotted

(assuming D65 adaptation).

300

0 200 300

-3001

CIELAB the spectral

Figure 73. MacAdam ellipses plotted in IPT (left) and (right) along with

locus.

IPT for the blue region, models the ellipses with less
Note that the color space, especially

difference between major and minor axes. The spectral locus for both of the color spaces

159
approaches chroma of 0 as the stimuli reach the edge of the spectral sensitivity of the color

matching functions.

3.2.2.3.9 Suprathreshold Color Difference Ellipses

Supra threshold color difference ellipses are shown plotted in IPT and CIELAB in figure 74.

These color ellipses model a color difference on the order of 1 delta E*ab unit. Because the

unit difference of the reference was near gray, the color difference was
definitely above

threshold, but is still considered a small color difference. Notice that these plots exhibit a

similar behavior in the blue region as did the MacAdam ellipses. Most notably, the eccentricity

of the ellipses in the blue region is reduced in IPT space.

75
0.4

50

i"\
0.2 i
25
*?0
0 o
^ i*\ or . i 4*$$
'
"O (n **
-0.4
-i?.-2>-S 0.2 0.4 75 25 50 75
...*/
*
-5'O-'--^0
-^
?'

.
--C.

: jiq:2
-50

-0.4
-75

Figure 74. RIT-DuPont IPT CIELAB (right).

visual color-difference ellipses plotted in
(left) and

160
 3.2.2.3. 10 OSA Color System (L =
lightness plane
medium gray 0) constant

Constant lightness plane from the OSA color system is shown in figure 75. Data are plotted in

IPT on the left and CIELAB on the right. This data set is included in addition to Munsell V5

data because the scaling was done on a rectangular

grid, thus all colors in the plane are equally

spaced. These data were scaled

using colored paint chips. The color difference between any

two colors is more than 20 times just noticeable.

? ?
* *
? i

? ? *
1-
-#
- -
? '-! -?-

1*

-0 0 5
*

Figure 75. OSA color system medium gray constant lightness plane plotted in IPT (left) and

CIELAB (right).

Note that in CIELAB, the vertical loci data are curved dramatically, whereas in IPT, though

the data are not exactly evenly spaced, the uniformity is much improved. It appears that the

curvature is slighdy reversed in IPT with respect to CIELAB. This may suggest that the color

blue hue locus for This may

slightly overcompensating for the
space is constant object colors.

point to a fundamental difference in perception between object colors and self luminous (CRT

displayed) colors.

161
3.3 Experiments E2. Verification Experiments

The constant hue data sets and the IPT color space were tested to measure the difference in

uniformity when entire hue leaves were viewed. This task is fundamentally different from the

task that was performed in deriving either of the constant hue data sets in that there was a

"same"

large amount of color context (the entire CRT gamut at the hue) to judge uniformity

within. Both experiments A and B were designed to be identical in procedure. The only

difference between the experiments was that the first one compared CIELAB, hue correction

based on
Hung and Berns, and hue correction based on Ebner and Fairchild, and the second

experiment compared the IPT color space to the 2 constant hue data sets.

3. 3. 1 Comparing CIELAB to Hung & Berns and Ebner data sets

3.3.1.1 Introduction

A paired-comparison experiment was conducted to test the hue uniformity difference between

two uniform perceived hue data sets and CIELAB color space. A colorimetrically

characterized and calibrated CRT display was used to present the stimuli.
Viewing conditions

were set to dark surround, near D65 white point, with a maximum luminance of 70cd/m2.

Berns' Braun70

Flung and constant lightness data set was used

by to derive a two dimensional

table
lookup scheme that enabled forward and inverse transformation between CIELAB color

space and the

Flung and Berns uniform hue data space. Ebner and Fairchild's data set of

constant perceived hue was used

by the author to derive a three dimensional table to enable

transformation between CIELAB color space and the respective uniform hue data space.

162
3.3.1 .2
Experimental

Fifteen hue angles were sampled in lightness the respective color

uniformly and chroma within

spaces, then transformed to CIELAB for display on the calibrated CRT display. Out of gamut

colors were converted to CIELAB coordinates 50,0,0. For each hue angle sampled, the three

color spaces were compared. For each hue angle, three presentations were required to

compare each color space to the other (N*(N-l)/2). For each observation, two repeats of the

data set were presented to the user to account for screen non-uniformity. Each observation

session required 90 judgements (15 hues X 3 Images X 2 repeats). The list of pairs of hue

comparisons was
pseudo-randomly shuffled for each observation session. The user interface

used to present the stimuli is shown in the figure 76.

io.5">|r

JVhite^^. White

high chroma high chroma

S
Black^^
Gray out of gamut
Blacjtx^
region

C left ) ( riant ]
( back j

Gray background
( uuil ) <-13.5"

Figure 76. User interface for verification experiment.

163
The task was as follows:

'You will be shown pairs of images. For each pair of images shown, pick the image that has the best hue

uniformity. The maximum chroma color for each image is the same color. Compare the colors in each image

"

separately, don 't compare colors between images.

Only same base hue images were compared to each other, never were a mix of hues presented

to the observer.

Thirty observations of the entire data set were made. Nine observers took part in the

experiment. All observers had experience with color, and were familiar with the terminology,

judgement71
and the concept of hue uniformity.
Using Thurstone's law of comparative (case

V), fifteen interval scales were derived, one for each hue sampled, that compared relative hue

uniformity of the three color spaces. Confidence limits were calculated based on the number

of observations of the entire data set. The confidence limits were calculated
using the

1.96-V2
following equation: j=
,
where N =
30, the number of observations (not number of

observers). This formula is based on the assumption that each scale value is a mean of a

normal distribution where the units of the scale are V2 times the standard deviation of the

distribution. Thus, a 95% confidence interval is calculated from 1.96 times the standard

error of the sample mean, which is 1.96 times ==

,
where S is arbitrarily assigned a value of 1,
4N

as is done in Thurstone's case V. This seems inconsistent with the calculation of the 95%

confidence limit by a factor of sqrt(2).

164
3.3.1.3 Results

Figure 77 shows the results from the experiment. Higher bars mean that the color space was

judged to be more uniform.

1.5

0.5

]J- MJ 4U -17 :
=e_ I io 'r4 T51 \i)g
1
>
-0.5
llCIELAB

0) H&B
-1
u QE&F
V)

-1.5

-2

-2.5

-3

-3.5

base CIELAB hue angle

Figure 77. Scale values for judged uniformity between CIELAB and the 2 constant hue data
sets. H&B are
Hung and Berns data, E&F are Ebner and Fairchild's data.

The most prominent feature of this figure is the significant difference in judged uniformity at

the 264 and 288 degrees reference hue angles (these represent the blue region).
Flung and

Berns and Ebner and Fairchild data sets were judged to be not
significantly different in the

blue colors. Most of the other reference hue angles seem to have very close scale values,

165
which leads one to believe that they are equally uniform at those hue angles. There is an

the reference hues 24, 48, and 72, where the Ebner and Fairchild
interesting trend observed at

hue surfaces are judged to be less uniform than both CIELAB and the Hung and Berns hue

correction. This trend was also observed in the gamut mapping experiment conducted
by

Braun70, where the red color of the parrots head was mapped to a color that was too magenta

when
using correction based on either constant hue data set. The results from Braun's

experiment led him to make a new hue correction LUT that only affected the blue region of

CIELAB space. This solved an immediate problem, but is much less general than a functional

color space approach to the constant hue problem. The root of the judged non-uniformity of

is not well understood. While the Hung and Berns results are not, from this result, judged to

be statistically less uniform, the shape of the E&F curve (at 48 degrees) and the shape of the

red stimulus of H&B have the same trend (see figure 48, left plot ). This bend in both loci of

the red stimulus may be more apparent in spot colors, and there may be preference influences.

Regardless of the cause, it is a systematic and repeatable effect.

3.3.2 Comparing IPT color space to Hung <& Berns and Ebner data sets
The experiment described in E2, part A was repeated
using hue planes from the new IPT

color space in place of the CIELAB color space. The identical user interface, script, data

preparation, and data analysis was performed. The number of observers for this experiment

was 10, and 30 observations were made over the entire data set. Confidence limits were

calculated the same

way as in experiment E2 A. which resulted in 95% values that were the

166
same (since the number of observations was the same). Figure 78 shows the results for the

mean
uniformity judgements.

base CIELAB hue angle

Figure 78. Scale values for judged uniformity between IPT and the 2 constant hue data sets.

Fairchild's data.
H&B are Hung and Berns data, E&F are Ebner and

The first characteristic to notice about this plot is that the extent of the scale values is of the

figure blue region is discounted. The next

thing to notice is
same order as that of 77, if the

for IPT color space are greater than either of the constant hue
that the scale values uniformly

significandy different
statistically).
data set scale values (although not all

167
Clearly, the IPT color space is judged either more uniform (when the confidence limits are

beyond the mean of the other two space, such as with reference hue 48), or no less uniform

(e.g. reference hue 144) than the constant hue data sets. Experiment E2 A. has shown that for

the reference hue of 48 degrees, CIELAB was judged more uniform than either of the

constant hue data sets. The magnitude of uniformity difference between CIELAB and the

next judgement was about 0.5 units, which is contrasted to roughly 0.75 units in figure 78 (for

reference hue angle 48). From this we can conclude that the IPT space is roughly as uniform

as CIELAB at the reference hue of 48 degrees, and more uniform than CIELAB everywhere

else. Additionally, the IPT color space is judged more uniform than the constant hue data sets

from which it was derived.

The last statement is a bit confusing. That a derived color space is more uniform than data

sets that were derived to be perceptually uniform seems to be wrong. But one must consider

that in making lookup tables that model the constant hue color space, and transform between

CIELAB and the constant hue data, some error will be introduced. Even when the data used

to create the transformations are very accurately modeled in the LUT transformations (which

it is in both cases), there will be uncertainty in accuracy of the transformation in other parts of

the space. It seems that it is always better to have an analytical description of a transformation

than to use lookup table based approaches because of the above issues (noise and accuracy

between sample points). Additionally, perceptually derived data bases are always calculated as

a central
tendency of a dispersion of individual results. The average response should reflect

the nature of the tendency, but there are many ways to calculate a central tendency, none of

168
which will be
"accurate"
all the time (accurate is in quotes because its judgement relies on

individuars responses which are not the same as the mean in almost all cases).

169
4 Conclusions

This work began with an in depth literature search that describes aspects of color such as color

order systems, color harmony, color names, color meaning, and preferred color reproduction

of pictorial images. From the literature search, a set of experiments was defined that sought to

find differences in gamut

mapping decisions based on color context. Several contexts were

defined that represented a taxonomy of graphical

(non-pictorial) image types. Experiments

were conducted that sought to determine the influence of image context (in business graphics

type images), and task description on gamut

mapping decisions. Additional color
meaning

experiments were conducted to

try to find dimensions of
meaning of color in the context of

business graphics.

A general color gamut

mapping model was derived from the above work that has many

applications. A paper
describing this and the gamut
mapping experiment was published in

Color Research and Application.

Based on the above work and other's

experiments, it was decided that gamut mapping in

CIELAB was not

sufficiently accurate with hue, and there were not sufficient data in the

community to generate a better color model.

Therefore, an experiment was conducted that

found 306 points that spanned the gamut of a CRT display that found 15 surfaces of constant

hue in color space.

170
Based on the above experimental results, a general tool for visualization of color space

transformations in real time was developed. The techniques used for visualization and the

controls that altered the parameters of color models were described.

A new model of color space named IPT was formulated from the use of the visualization tool.

Many popular color appearance data sets were used in the development of the new color

space. The IPT color space is a simple transformation from CIEXYZ D65, average surround

conditions (reference viewing conditions). The model consists of a 3x3 matrix transform, a

nonlinear power function, and another 3x3 matrix transform. This model is a generalization of

the CIELAB color space that takes advantage of a transformation to cone fundamental

tristimulus space before signal compression, then a linear transform to create the opponent

and lightness channels.

The constant hue data sets were judged against the CIELAB color space, and the IPT color

space in 2 subsequent experiments. The CIELAB color space was judged much worse in blue

hue data sets, the IPT color space was judged

slightly better
and in red than the constant while

superior in hue uniformity to both constant hue data sets at all hue angles tested.

is culmination of all the above work, either direcdy, or indirectly. The

The IPT color space the

has many appearance attributes that are superior to the CIELAB color space
IPT color space

it most in the blue region, and in the dependence on hue

which was compared against, notably

angle for lightness prediction of chromatic colors.

171
There are several areas where the IPT color space can be utilized. The color space is both easy

to implement for Image Processing Transformations (IPT's), and is a much better predictor of

color appearance attributes than is CIELAB, and it is much more uniform in hue than any of

CIELAB, CIELUV, or CIECAM97s.

5 Future Work

Additional work needs to be done in two areas. Firstly, further verification must be done to

test whether the IPT color space is the preferred color space to use for gamut
mapping

applications. This can

only be done by testing the space under realistic gamut
mapping

conditions. This remaining work

may conclude that there are other similar, but better color

spaces to gamut map within, and much of this

may depend heavily on observer preference as

well as psychometric uniformity. The second area of work is involved with refinement of the

color space to test and create more

perceptually uniform representation of human color vision.

Specifically, the relationship of rate of change of perceived hue as a function of hue angle

should be tested to give better data with which to model hue uniformity. These goals are

complementary, but need to be approached from different directions.

172
6 Appendices

6.1 Appendix A: Data plots from experiment PI.

orange

green

Dyellow
Oviolet
Dblue

gray
Dyel green

Dpurple
red

Dcyan

Figure 79. Mean scale values for colors and descriptive terms for abstract representative image
type.

173
 1 i 1 .
=
i r^-^ (t
g ^^=S==;
1-| _^
1

'
1 "
UXUUTIUIVU i i
J- "

1 1 1
^^^==
1 1
1
i
' 1
1 h

=^ 1

H
1
l~j 3 1 1
1 i r i

cornplicatecP
BBBBBBBJ 1
=t
'
1
~"

[
1 ] 1

OBBBSHS9H 1
B orange
B green
' 1
1 1
1= ? yellow
' B violet
|
H ? blue

1 1
' Bgray
|
1 ? yel green

1 1 ? purple
,_

' Bred
t :

i
1
| 1
? cyan
1 i : | i

BBIB1BBJW 1
MU^ 1

1 i

^g
1-
|l'

1 1
1

==i 1

1 '

""

1 1=
1
t
1 |
1
1

1 1 1
!

Figure 80. Mean scale values for colors and descriptive terms for data representative image
type.

174
6.2 Appendix B: Data from experiment P2.

Table 5. Luminance and chromaticities of setup.

viewing

X
y Yfcd/m2)
D65 white point .313 .329 1
Luminance of Halon under .320 .333 90.6
source

Luminance of Paper under .320 .333 80.3

source

Background gray matte under .307 .321 13.3

source

CRT Workspace background .306 .322

CRT Window background .305 .322

Root color .306 .322

CRT calibrated white (RGB) 255 240 220

CRT Center calibrated white .305 .321 79.1
lights on

CRT Left calibrated white lights .301 .316

on

CRT Right calibrated white lights .307 .322

on

CRT calibrated white lights off .305 .314 75.0

Flare at center with rear 2 banks .320 .340 3.96
on

175
 Table 6. Colors chosen for gamut mapping color match.

L* b*
h degrees
a*

C*ab color name

65.15 49.51 67.93 84.06 53.91 orange

63.19 60.67 -55.49 82.22 -42.44 violet

56.72 74.36 1.86 74.38 1.43 maroon

20.56 48.12 -71 .30 86.02 -55.98 dark blue

87.73 23.26 -8.80 24.87 -20.72 reddish pink

75.63 -25.54 -36.77 44.77 -124.78

sky blue
42.65 70.86 -60.76 93.34 -40.61 purple

89.17 -48.15 -16.86 51.02 -160.69 bright cyan

85.64 -89.15 79.44 119.41 138.29 bright green

57.51 71.26 67.01 97.82 43.23 red

88.65 -57.56 23.25 62.08 158.00 light green

55.20 -7.30 -43.57 44.18 -99.51 medium blue
87.73 16.31 7.29 17.86 24.09 skin tone
97.11 -21.92 92.24 94.81 103.36 bright yellow

64.02 90.43 -54.42 105.54 -31 .03 bright violet

32.64 78.94 -106.64 132.68 -53.49
deep blue
60.14 8.97 -60.98 61.64 -81.62 medium blue
60.86 66.92 28.69 72.82 23.20 light red
86.59 -67.48 18.19 69.89 164.91 light green
48.60 30.06 -76.90 82.57 -68.64 blue
76.74 -30.65 -31.15 43.70 -134.53 light blue
58.85 80.44 -28.84 85.46 -19.72 magenta
79.64 14.69 79.05 80.40 79.47 tannish orange
57.61 43.68 -64.74 78.11 -55.99 light purple
88.55 -46.26 83.74 95.67 118.91 pea green

176
6.3 Appendix C: Three parameter files corresponding to Figure 57
"premat"

Notice that for each of the parameter files, only the second row of the entry named

is different. Compare these differences to the change in color

matching function and

corresponding shape of the color space in figure 57.

Parameter file for top figure:

premat

0.380400 0.707600 -0.088000

-0.215700 1.165300 0.049800

0.000000 0.000000 1.000000
oppmat

1.000000 0.000000 0.000000

0.000000 1.000000 0.000000
0.000000 0.000000 1.000000
rotmat

1.000000 0.000000 0.000000

0.000000 1.000000 0.000000
0.000000 -0.000000 1.000000

postmat

0.000000 1.000000 0.000000

3.500000 -3.500000 -0.000000

0.000000 1.006356 -1.006356

nlinvals

43.000000 10.000000 43.000000 10.000000 43.000000

10.000000
Opponentargs

3.500000 -3.500000 0.000000 1.006356

Ulargs

2 110 4

Parameter file for middle figure:

premat

0.380400 0.707600 -0.088000

-0.166667 1.100000 0.066667

0.000000 0.000000 1.000000

oppmat

1.000000 0.000000 0.000000

0.000000 1.000000 0.000000

0.000000 0.000000 1.000000

rotmat

1.000000 0.000000 0.000000

177
0.000000 1.000000 0.000000
0.000000 -0.000000 1.000000
postmat

0.000000 1.000000 0.000000

3.500000 -3.500000 -0.000000

0.000000 1.006356 -1.006356

nlinvals

10.000000 43.000000 10.000000 43.000000 10.000000 43.000000

Opponentargs
3.500000 -3.500000 0.000000 1.006356
Ulargs
2 110 4

Parameter file for bottom figure:

premat

0.380400 0.707600 -0.088000

-0.083333 0.933333 0.150000

0.000000 0.000000 1.000000
oppmat

1.000000 0.000000 0.000000

0.000000 1.000000 0.000000
0.000000 0.000000 1.000000
rotmat

1.000000 0.000000 0.000000

0.000000 1.000000 0.000000
0.000000 -0.000000 1.000000
postmat

0.000000 1.000000 0.000000

3.500000 -3.500000 -0.000000

0.000000 1.006356 -1.006356

nlinvals

10.000000 43.000000 10.000000 43.000000 10.000000 43.000000

Opponentargs
3.500000 -3.500000 0.000000 1.006356
Ulargs
2 110 4

178
 7 References

1
Ebner, F., and
Fairchild, M, Gamut Mapping From Below: Finding minimum perceptual distances for colors outside the
gamut volume, Color Research and
Application, VOL, NO, PP. (1997)
2
Ebner, R, and Fairchild, M, Finding constant hue surfaces in color space, Proceedings o/SPIE, Color Imaging: Device-Independent
Color, Color Hardcopj, and Graphic Arts III, 3300, 107-117, (1998)
3
Munsell, A. FL, A Color Notation, Munsell Color Co. 12th ed. 1971
4
Munsell, A. H., A Grammar of Color, Van Nostrand Reinhold Co., 1969
5
Birren, F., Application of the Ostwald Color System to the Design of Consumer Goods,/. Opt. Soc. Am. 34 No. 7 396-399
(1944)
6
Zeishold, H., Philosophy of the Ostwald Color System,/. Opt. Soc. Am., 34, No. 7, 355-360 (1944)
7
Foss, C. and Nickerson, D., Analysis of the Ostwald Color System,/. Opt. Soc. Am., 34, No. 7,361-381
(1944)
8
Stevens, J.C., and Stevens, S.S., Brightness Function: Effects of
Adaptation, ,/. Opt. Soc. Am., 53, No. 3,375-385 (1963)
5
Hard, A. and Sivik, L., NCS-Natural Color System: A Swedish Standard for Color Notation, Color Res. and Appl 6 No. 3
(1981)
10
Hard, A. and Sivik, L., Some Reflections on
Studying Colour Combinations, Color Res. and Appl, 19, No. 4, (1994)
11
Hard, A. Sivik, L., Outiines
and of a
Theory of Colors in Combination, Man-Environment Systems, 9:217-228, (1979)
12
Hard, A. and Sivik, L., Distinctness of Border: An Alternative Concept for a Uniform Color Space, Color Res. and Appl, 11,
No. 2, (1986)

13
Moon, P. and Spencer, D., Geometric Formulation of Classical Color Harmony, ,/. Opt. Soc. Am., 34, No. 1, 46-59 ,(1944)

14
Moon, P. and Spencer, D., Area in Color Harmony, / Opt. Soc. Am., 34, No. 2, 93-103, (1944)
15
Pope, A., Notes on the Problem of Color Harmony and the
Geometry of Color Space,/ Opt. Soc. Am., 34, No. 12, 759-765,
(1944)
16
Moon, P. and Spencer, D., Aesthetic Measure Applied to Color Harmony, / Opt. Soc. Am., 34, No. 4, 234-242, (1944)
17
Granger, G.W., Area Balance in Color Harmony: An Experimental Study, Science, YI, 59-61, Jan. (1953)
18
Nayatani, Y. et. al., An Appraisal of Two-Color
Harmony by Paired Comparison Method, Acta Chromatica, 2, No. 1 (1969)
19
Berlin and Kay, Basic Color Terms, U. of California Press, (1969)
20
Kelly, K. and Judd, B., The ISCC-NBS Method of
Designating Colors and A Dictionary of Color Names, National Bureau of
Standards Circular 553, (1955)
21
Birren, F., Color and Human Response, Van Nostrand Reinhold, (1978)
22 Studies of Color
Sivik, L., Meaning, Man-Environment Systems, 5, no. 3, 155-160, (1975)
23
Taft, C, The Stability and Variability of Color Meaning Associations Across Time and Cultures, Colour 89, Proceedings of the 6th
AIC Congress, (1989)

24 Color dimensions: A study

Sivik, L., Meaning and perceptual color of color samples, Goteborg Psychological Reports, 4, No. 1,
(1974)
25
Taft, C. and Sivik, L., Cross National Comparisons of Color Meaning, Goteborg Psychological Reports, 22, No. 3, (1992)

179
26
Sivik, L., Color Meaning and perceptual color dimensions: A study of Exterior Colors, Goteborg Psychological Reports, 4, No. 1 1,

(1974)
27
Kobayashi, S., The Aim and Method of the Color Image Scale, Color Research and Application, 6, No.2, (1981)
28
Buck, G.B., and Froelich, H.C., Color Characteristics of Human Complexions, Illuminating Engineering, Jan (1948)
29
Bartleson, C.J., Memory Colors of Familiar Objects, Journal of Opt. Soc. ofAmerica, 50, No. 1, (1960)
30
Newhall, S.M., Burnham, R.W., and ClarkJ.R, Comparison of Successive with Simultaneous Color Matching, Journal of Opt.
Soc. of America, 47, No. 1, (1957)

31
Bartleson, C.J., Some Observations on the Reproduction of Flesh Colors, Photographic Science and Engineering, 3, No. 3, (1959)
32 Green-Grass
Bartleson, C.J. and Bray, C.P., On the Preferred Reproduction of Flesh, Blue-Sky, and Colors, Photographic
Science and Engineering, 6, No. 1, (1962)

33 Blue Green Caucasian Skin in Colour

Hunt,R.W.G, Pitt, IT., Winter, L.M., The Preferred Reproduction of Sky, Grass, and

Photography, The Journal of Photographic Science, 22, (1974)

34
Asada, T., Sakaue, K, Haneishi, H., and Miyake, Y., Preferred Skin Color Reproduction of Hardcopy, WWW page

http://www.icsd6.tj.chiba-u.ac.jp/ICSD6/papers/asada/skin.html, (1994)
35 Natural Scenes Transformed in
Fedorovskaya, E., Blommaert, F., de Ridder, H., Perceptual Quality of Color Images of
CIELUV Color Space, Proceedingsfrom the First IS&T and SID 's Color Imaging Conference, (1 993)

36
Jones, L., Psychophysics and Photography, Journal of Opt. Soc. ofAmerica, 34, No. 2, (1944)
37
MacAdam, D.L., Quality of Color Reproduction, Journal of the SMPTE, 56, (1951)
38
Roufs, J.A.J Perceptual Image Quality: Concept and Measurement, Philips J. Res., 47, (1992)
.,

39
RS. Gentile, E. Walowit, and J.P. Allebach, A Comparison of Techniques for Color Gamut Mismatch Compensation, /. of
y,16, No. 4, Oct. 1990
40 M.C. Stone, W.B. Cowan, J.C. Beatty, Color Gamut Mapping and the Printing of Digital Color Images, ACM Transactions on
Graphics, 7, No. 4, Oct. 1988
41 J. Meyer, B. Barth, Color Gamut Matching for Hard Copy, SID 89 Digestpp 86-89, 1989
42
Osgood, Suci, and Tannenbaum, The Measurement of Meaning, University of Illinois Press, 1978
43
Ibid, page 73. Relative importance of sematic space dimensions.

44
Same as ref. 24, page 10.

45
Stone and J. Beatty, Color Gamut Mapping and the Printing of Digital Color Images, ACM Transactions on Graphics, 7, No. 4,
249-292 (1988).

46
W. Wallace and M. Stone, Gamut Computer Generated Imagery, SPIE Vol 1460 Image Reproduction
Mapping Handling and

Systems Integration, 20-28, (1991).

47
R. S. Gentile, E. Walowit, and J. P. Allebach, A Comparison of Color Gamut Mismatch Compensation,/, of Imaging Technology,
16, No. 5, 176-181, (1990).
48 M. Wollski, J. P. Allebach, C. A. Bouman, Gamut Mapping, Squeezing the Most out of Your Color System, IS&T and SID 's
2nd Color Imaging Conference: Color Science, Systems, and Applications, 89-92, (1994).
49
T. Hoshino and R Berns, Color gamut mapping techniques for color hard copy images, SPIE Vol 1909, 152-165, (1993).

50
Montag, E.D. and Fairchild, M.D., Simulated Color Gamut Mapping Using Simple Rendered Images, Proc. SPIE 2658, San
Jose, 316-325(1996).

180
 51 N. Katoh and M. Ito, Gamut Mapping for Computer Generated Images (II), IS&T and SID's 4th Color Imaging Conference:
ColorScience, Systems, and Applications,
126-129, (1996).
52 William H. Press, Saul A.Teukolsky, William T. Vetterling and Brian P. Numerical Recipes in
Flannery, C, Second Edition,
Cambridge University Press, ISBN 0-521-43108-5, pp.408-412, 1992.
53 A P. Hung and R.
Berns, Determination of Constant Hue Loci for CRT Gamut Their Predictions Color
a and
Using
Appearance Spaces, Color Res. Appl, 20, No. 5, 285-295, (1995).
54
Newhall, Nickerson, Judd, Final Report of the O.S.A. Subcommittee on the Spacing of the Munsell Colors, ,/. Opt. Soc. Am.,
33, No. 7,385-418(1943).
55 RW.G. Hunt, Measuring Color, Second Edition, Ellis Horwood, ISBN 0-13-567686-X, pp.142, (1991).

56 R W .G.
Hunt, "Colour Terminology", Col. Res. <&Appl, 3, No. 2, pp. 79-87, 1978.

57
B. Berlin and P. Kay, "Basic Color Terms", pp. 17-23, University of California Press, Berkeley, 1969.
58 R. S. Berns, R J. Motta, and M. E. Gorzynski, "CRT colorimetry, part I: theory and practice", Col Res. e>Appl., 18, pp. 299-

314, 1993.

59
Timothy Masters, Practical Neural Network Recipes in C++, 1993 Academic Press

60
Timothy Masters, Signal and Image Processing with Neural Networks, 1994 Wiley
61
Timothy Masters, Advanced Algorithms for Neural Networks: A C++ Sourcebook, 1995 Wiley
62 R Rolleston, "Using Shepard's Interpolation to Build Color Transformation Tables", IS&T and SID's 2nd
Color Imaging
Conference: Color Science, Systems and Applications pp. 74-77 (1994).
,

83 S. L. "Model for
Guth, color vision and light adaptation", J. Opt. Soc. Am. A, 8, No. 6, pp. 976-993 (1991)
64 Color- Appearance
M.D. Fairchild, RS.
Berns,
and "Image Specification Through Extension of CIELAB", Col Res. &AppL,
18, No. 3, pp. 178-189, 1993.
65 G. Wyszecki and W.S. Stiles, "Color Science: Concepts and Methods, Quantitative Data and Formulai", Second Edition,
Wiley Press, ISBN 0-471-02106-7, (1982).
66 S.M. OSA. Subcommittee
Newhall, D. Nickerson, D.B.Judd, "Final Report of the on the
Spacing of the Munsell Colors",
/ Opt. Soc. Am., 33, No 7, pp. 385-418, (1943).
<"
M.D. Fairchild, "Refinement of the RLAB Color Space", Col. Res. &Appl, 21, No. 5, pp. 338-346, 1996.

68 M.D. Chromatic Object Colors Col Res.

Fairchild, E. Pirrotta, "Predicting the Lightness of
Using CIELAB", <& Appl, 16,
No. 6, pp. 385-393, 1991.
69 M. Melgosa, E. Hita, A.J. Poza, D.H. Alman, R.S. Berns, "Suprathreshold Color-Difference Ellipsoids for Surface Colors",
Col. Res. &Appl, 22, No. 3, pp. 148-155, 1997.

70 Personal communication with Gus Braun, and an abstract submission by Gus Braun, the author, and Mark Fairchild for

CIC'98, "Color Gamut mapping in a Hue-Linearized CIELAB Color Space"-

71 G.A.Gescheider, "Psychophysics, Method, Theory, and Application", Second Edition, Lawrence Erlbaum Associates, ISBN
0-89859-375-1, pp. 152 (1985).

181