The LabPQR Color Space Giordano B. Beretta Print Production Automation Lab Hewlett-Packard Laboratories Palo Alto, California 8 April 2010 G. Beretta (HP Labs) LabPQR Overview 8 April 2010 1 / 35
Sep 06, 2014
The LabPQR Color Space
Giordano B. Beretta
Print Production Automation LabHewlett-Packard Laboratories
Palo Alto, California
8 April 2010
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 1 / 35
Color matching
Colors are assessed by matching them with reference colors on asmall-field bipartite screen
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 2 / 35
Color-matching functions
Given a monochromatic stimulus Qλ of wavelength λ, it can be writtenas
Qλ = RλR + GλG + BλBwhere Rλ, Gλ, and Bλ are the spectral tristimulus values of Qλ
Assume an equal-energy stimulus E whose mono-chromaticconstituents are Eλ (equal-energy means Eλ ≡ 1)The equation for a color match involving a mono-chromatic constituentEλ of E is
Eλ = r(λ)R + g(λ)G + b(λ)Bwhere r(λ), g(λ), and b(λ), are the spectral tristimulus values of Eλ
Definition (color-matching functions)
The sets of such values r(λ), g(λ), and b(λ) are called color-matchingfunctions (CMF)
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 3 / 35
Color-matching functions
-0.5
0.0
0.5
1.0
1.5
2.0
2.5
3.0
b(λ)
g(λ)
r(λ)
700 600 500 400
nm
Stiles-Burch (1955;1959)
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 4 / 35
CIE 1931 standard colorimetric observer
We want to build an instrument delivering results valid for the group ofnormal trichromats (95% of population); since
R = k∫
Pλr(λ)dλ
G = k∫
Pλg(λ)dλ
B = k∫
Pλb(λ)dλ
an ideal observer can be defined by specifying values for thecolor-matching functions
Definition (CIE 1931 standard colorimetric observer)
The Commission Internationale de l’Éclairage (CIE) has recommendedsuch tables containing x(λ), y(λ), z(λ) for λ ∈ [360nm,830nm] in 1nmsteps
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 5 / 35
Illumination
The spectral power distribution of the light reflected to the eye by anobject is the product, at each wavelength, of the object’s spectralreflectance value by the spectral power distribution of the light source
500 700600400 500 700600400 500 700600400
500 700600400500 700600400 500 700600400
Incident SPD Reflected SPDReflectance curvex =
CWF
DeluxeCWF
Complexion
Complexion
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 6 / 35
Mathematical interpretation
R = k∫
Pλ · r(λ)dλ
means that the red color coordinate is obtained by integrating the SPDusing the red CMF for the measure, where
Pλ = E(λ) · S(λ)
is the product of the SPD of an illuminant E with the object spectrum S.Usually we are interested in the coordinates of various objects under afixed illuminant for a standard observer, so we reorder to
R = k∫
r(λ)E(λ) · S(λ)dλ
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 7 / 35
Discretization
In practice, the CMF are given as a table with 1nm steps, andinstruments measure at steps of 1,4,10,20nm etc., so in reality this isa summation [for red R]:
R = k∫
r(λ)E(λ)S(λ)dλ ≈ k∑
r(λi)E(λi)S(λi)∆λ
The integration resp. summation is over the visible range [380,780]nm,but in practice it is often over [380,730]nm for n = 36 samples
Instead of doing color science with measure theory, we can do itwith simple linear algebraIn 1991 H. Joel Trussell has made available a comprehensiveMatLab library and several key papers for color scientistsSince then, spectral color science is mostly done with linearalgebra
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 8 / 35
Formalism
We use the vector-space notationWLOG, let k = 1R = (R>E)S, G = (G>E)S, B = (B>E)SInstead of doing this for each of R,G,B or X ,Y ,Z , using linearalgebra we can write it as a single equation by combining the CMFin an n × 3 matrix A with the CMFs data in the columns:
Υ = (A>E)S
Sometimes we are interested in the color of a fixed object underdifferent illuminants, then we write
Υ = A>(ES) = A>η
η corresponds to the Pλ from earlier
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 9 / 35
Metameric stimuli
Consider two color stimuli
Q1 = R1R + G1G + B1BQ2 = R2R + G2G + B2B
Definition (metameric stimuli)If Q1 and Q2 have different spectral radiant power distributions, butR1 = R2 and G1 = G2 and B1 = B2, the two stimuli are calledmetameric stimuli
FactColor reproduction works because of metamerism
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 10 / 35
Fundamental and residue
How can we reconcile metamerism and color reproductiontechnology?In 1953 Günter Wyszecki pointed out that the SPD of stimuliconsists of a fundamental color-stimulus function η(λ) intrinsicallyassociated with the tristimulus values, and a metameric blackfunction κ(λ) called the residueκ(λ) is orthogonal to the space of the CMF
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 11 / 35
Matrix R theory
How does this translate to the discrete case?In 1982 Jozef Cohen with William Kappauf developed the matrix RtheoryUse an orthogonal projector to decompose stimuli in fundamentaland residueThe fundamental is a linear combination of the CMF AThe metameric black is the difference between the stimulus andthe fundamentalFor a set of metamers η1(λ), η2(λ), . . . , ηm(λ):
A>η1 = A>η2 = · · · = A>ηm = Υ
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 12 / 35
Development of matrix R
R is defined as the symmetric n × n matrix
Definition (matrix R)
R := A(A>A)−1A>
Matrix R is an orthogonal projectionA(A>A)−1 =: Mf , so R = Mf A> (remember: Υ = A>η)Because A has 3 independent columns, R has rank 3It decomposes the stimulus spectrum into fundamental η(λ) andthe metameric black κ:
η = Rηi
κ = ηi − η = ηi − Rηi = (I − R)ηi
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 13 / 35
Corollaries
Metameric black has tristimulus value zero
A>κ = [0,0,0]>
η = Rηi means that any group of metamers has a commonfundamental η, but different residues κInversely, a stimulus spectrum can be expressed as
ηi = η + κ = Rηi + (I − R)ηi
i.e., the stimulus spectrum can be reconstructed if thefundamental metamer and metameric black are knownWhy is this useful?
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 14 / 35
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 15 / 35
Spectral color reproduction
Sometimes colorimetry is insufficientSpectral printer modelsMapping from one device to anotherFluorescent inks and/or mediaPhysical media modelsInk-media interactionsSecurity printingMore than 3 colorant hues (e.g., CMYKOGV)Multiple illuminants (metamerism index minimization)Mapping K generation between two different CMYK printersScanner and camera characterization. . .
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 16 / 35
Reducing the data
Storing a multidimensional vector for each pixel is expensiveCan we project on a lower-dimensional vector space?Yes, because the spectra are relatively smoothPopular technique: principal component analysisDue to the usually smooth spectra, the dimension can be quitelow: between 5 and 8
We have known how to deal with this for decades, it just requireslinearly more processing
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 17 / 35
The hard problem
We would like to use an ICC type workflow also for spectralimagingColorimetric workflow:
profile connection space 3-hue printerimage
The killer is the LUT used in the PCS:bands in bands out levels per band size [bytes]
3 6 17 30K
6 6 17 145M
9 6 17 700G
31 6 17 8 · 1027G
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 18 / 35
Interim Connection Space
Proposal by Mitchell Rosen et al. at RITIntroduce a lower-dimensional Interim Connection Space ICS
PCS to ICS
multi-hue printerscene
ICS to counts via low-dim. LUT
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 19 / 35
Choosing the basis vectors
Can we deviate from the usual PCA method of choosing thelargest eigenvectors and build on some other useful basis?When defining the basis vectors for XYZ, the new basis waschosen so that one vector coincides with luminous efficiency V (λ) compatibility of colorimetry with photometry1995 proposal by Bernhard Hill et al. at RWTH Aachen:incorporate three colorimetric dimensions compatibility of spectral technology with colorimetry
X Y Z V 4 ............. V 16
S1 S2 S3 .............. S64
L* a* b*
L* a* b* V* 4 ............ V* 16
L bit abit b bit V 4,bit .......... V 16,bit
nonlinear transform
spectral scan values
conventionalthree channel
display or printersystem interface
S1 S2 S3 .............. S16
multichanneldisplay orprinting
smoothing inverse
basis functions
quantization
spectral reconstruction
multispectral values
nonlinear representationencoding
decoding
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 20 / 35
LabPQR Approach
Mitchell Rosen et al. at RIT1 Calculate operator similar to matrix R using regression analysis
on a specific printer (unconstrained), or matrix R directly(constrained)
2 Calculate residue using principal component analysis3 Calculate tristimulus values XYZ4 Calculate PQR from residue (3 largest EV)5 Calculate LabPQR from XYZPQR using CIE equations
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 21 / 35
LabPQR notation
Reconstructed spectrum (LabPQR transform): P = TNc + VNpT : colorimetric transformationNc : tristimulus vector ΥV : basis vectors in PQRNp: residual
Constrained: Tck = A(A>A)−1 = MfRemember: matrix R = Mf A>
Unconstrained: Tu = RN>c (NcN>c )−1 via least squares analysisover a number of tristimulus vectors for spectra R = ηi
Calculation of V : first 3 eigenvectors in metameric black κ viaprincipal component analysis
Conventional notation:
η = Rηi (= Mf Υ)
κ = ηi − η = ηi −Rηi = (I −R)ηi
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 22 / 35
LabPQR gamut
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 23 / 35
Using LabPQR
The diagram in the previous slide indicates how the algorithm isverifiedNote in particular the meaning of gamut mapping in PQR
The usage is to print a color chart and measure it spectrallyThe resulting table from device coordinates to spectra is then
1 converted to LabPQR2 inverted
The inverted table is used to interpolate LabPQR values to obtainthe device coordinates to reproduce a requested spectrum
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 24 / 35
Canon i9900 dye-based inks
G
K
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 25 / 35
Caveats
Green and black dyes tend to have an increasing reflectance inthe far redPaper brighteners act in the blue rangeRIT work: [400,700]nm for n = 31 samplesMost real world data: [380,730]nm for n = 36 samplesVisible range: [380,780]nm
The range has a strong effect on the principal components
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 26 / 35
PQR
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 27 / 35
Quality metric
objective function = minimize (CIEDE2000 + k ·∆PQR)
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 28 / 35
Accuracy of Matrix R vs. unconstrained
What price in loss of accuracy do we pay for compatibilityconventional metamerism theory?
Constrained model depends only on CMFUnconstrained model additionally depends on device
Based on simulations (no LUT),the constrained model is more accurate in generalfor a single fixed printer, the unconstrained method allows the useof less principal components: LabPQ
Short spectral range [400,700]nm caused problems with green ink
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 29 / 35
Summary
1 Conventional ICC workflow is based on colorimetry2 A spectral workflow can can solve many more problems
proof printingfluorescencemetamerism. . .
3 LabPQR is low-dimensional and compatible with colorimetry
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 30 / 35
Bibliography I
Henry R. Kang.Computational Color Technology.SPIE, Bellingham, 2006.
Bernhard Hill.The history of multispectral imaging at Aachen University ofTechnology.Web Document, May 2002.http://www.ite.rwth-aachen.de/Inhalt/Documents/Hill/AachenMultispecHistory.pdf.
Thomas Keusen and Werner Praefcke.Multispectral color system with an encoding format compatible withthe conventional tristimulus model.In IS&T/SID Third Color Imaging Conference, pages 112–114,1995.
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 31 / 35
Bibliography II
Mitchell R. Rosen and Ohta Noboru.Spectral color processing using an interim connection space.In IS&T/SID Eleventh Color Imaging Conference, pages 187–192,2003.
Maxim W. Derhak and Mitchell R. Rosen.Spectral colorimetry using LabPQR —- an interim connectionspace.In IS&T/SID Twelfth Color Imaging Conference, pages 246–250,2004.
Mitchell R. Rosen and Maxim W. Derhak.Spectral gamuts and spectral gamut mapping.In Mitchell R. Rosen, Francisco H. Imai, and Shoji Tominaga,editors, Spectral Imaging: Eighth International Symposium onMultispectral Color Science, volume 6062, pages60620K–1–60620K–11, 2006.
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 32 / 35
Bibliography III
Maxim W. Derhak and Mitchell R. Rosen.Spectral colorimetry using LabPQR: An interim connection space.Journal of Imaging Science and Technology, 50(1):53—63, 2006.
Shohei Tsutsumi, Mitchell R. Rosen, and Roy S. Berns.Spectral gamut mapping using LabPQR.Journal of Imaging Science and Technology, 51(6):473—485,2007.
Shohei Tsutsumi, Mitchell R. Rosen, and Roy S. Berns.Spectral color reproduction using an interim connectionspace-based lookup table.Journal of Imaging Science, 52(4):040201–040201–13, 2008.
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 33 / 35
Bibliography IV
Shohei Tsutsumi, Mitchell R. Rosen, and Roy S. Berns.Spectral color management using interim connection spacesbased on spectral decomposition.Color Research & Application, 33(4):282–299, August 2008.
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 34 / 35
Discussion
http://www.hpl.hp.com/personal/Giordano_Beretta/
G. Beretta (HP Labs) LabPQR Overview 8 April 2010 35 / 35