iCAM framework for image appearance, differences, and quality

Journal of Electronic Imaging 13(1), 126–138 (January 2004).

iCAM framework for image appearance,differences, and quality

Mark D. FairchildGarrett M. Johnson

Rochester Institute of TechnologyChester F. Carlson Center for Imaging Science

Munsell Color Science Laboratory54 Lomb Memorial Drive

Rochester, New York 14623-5604E-mail: [email protected]

ereuronlpgeoloeln aei-ithingpeofhefor

forys-d to

emsby

for

ente-act,in-rized

IEor-e-IEtan-

-m-ys-for

ing, to

c-

re-of

e-ndro-osslor

003

Abstract. Traditional color appearance modeling has recently ma-tured to the point that available, internationally recommended mod-els such as CIECAM02 are capable of making a wide range of pre-dictions, to within the observer variability in color matching and colorscaling of stimuli, in somewhat simplified viewing conditions. It isproposed that the next significant advances in the field of color ap-pearance modeling and image quality metrics will not come fromevolutionary revisions of colorimetric color appearance modelsalone. Instead, a more revolutionary approach will be required tomake appearance and difference predictions for more complexstimuli in a wider array of viewing conditions. Such an approach canbe considered image appearance modeling, since it extends theconcepts of color appearance modeling to stimuli and viewing envi-ronments that are spatially and temporally at the level of complexityof real natural and man-made scenes, and extends traditional imagequality metrics into the color appearance domain. Thus, two previ-ously parallel and evolving research areas are combined in a newway as an attempt to instigate a significant advance. We review theconcepts of image appearance modeling, present iCAM as one ex-ample of such a model, and provide a number of examples of theuse of iCAM in image reproduction and image qualityevaluation. © 2004 SPIE and IS&T. [DOI: 10.1117/1.1635368]

1 Introduction

The fundamental theme of this research can be considimage measurement, and the application of those measments to image rendering and image quality evaluatiConsideration of the history of image measurement heset the context for the formulation and application of imaappearance models, a somewhat natural evolution of cappearance, spatial vision, and temporal vision modwhen they are considered in a holistic sense, rather thaindividual research fields. Early imaging systems werether not scientifically measured at all, or measured wsystems designed to specify the variables of the imagsystem itself. For example, densitometers were develofor measuring photographic materials with the intentspecifying the amounts of dye or silver produced in tfilm. In printing, similar measurements would be made

Paper RTX-13 received Jan. 2, 2003; revised manuscript received Jul. 17, 2accepted for publication Aug. 8, 2003.1017-9909/2004/$15.00 © 2004 SPIE and IS&T.

126 / Journal of Electronic Imaging / January 2004 / Vol. 13(1)

de-.s

rss

d

the inks as well as measures of the dot area coveragehalftone systems. In electronic systems like television, stem measurements such as signal voltages were usecolorimetrically quantify the imaging system.1 It should benoted that vision-based measurements of imaging systfor image quality do have a long history, as illustratedthe example of Schade’s pioneering work.2 As imaging sys-tems evolved in complexity and openness, the needdevice-independent image measures became clear.

1.1 Image Colorimetry

Electronic imaging systems, specifically the developmof color television, prompted the first application of devicindependent color measurements of images. Wright, in fpoints out that color television could not have beenvented without colorimetry.3 Device-independent colomeasurements are based on the internationally standardCIE system of colorimetry first developed in 1931. Ccolorimetry specifies a color stimulus with numbers proptional to the stimulation of the human visual system, indpendent of how the color stimulus was produced. The Csystem was used very successfully in the design and sdardization of color television systems~including recentdigital television systems!.

Application of CIE colorimetry to imaging systems became much more prevalent with the advent of digital iaging systems and, in particular, the use of computer stems to generate and proof content ultimately destinedother media, such as print. As color-capable digital imagsystems~from scanners and cameras, through displaysvarious hardcopy output technologies! became commer-cially available in the last two decades, it was quickly reognized that device-dependent color coordinates~such asmonitor RGB and printer CMYK! could not be used tospecify and reproduce color images with accuracy and pcision. An additional factor was the open-systems naturedigital imaging in which the input, display, and output dvices might be produced by different manufacturers, aone source could not control color through the entire pcess. The use of CIE colorimetry to specify images acrvarious devices promised to solve some of the new co

;

msndgede

terea

des,inthi

vevi

cklloang,ofob

stlyedto,ea

ntly

llyt ormlorrel-ar, a

nee

nductedlo

ings.fo

etialuanetionsres-

taingle

nt inin

theon-

gend

eenge

d-on,

is

o-lo-ei-

ap-elyededis-ereof

tohehe

li-in

en-for-stic-h-spa-lity

s inin

horof

i.n-

pa-ics,e-arlyofre-

iCAM framework for image appearance . . .

reproduction problems created by open, digital systeThe flexibility of digital systems also made it possible apractical to perform colorimetric transformations on imadata in attempts to match the colors across disparatevices and media.

Research on imaging device calibration and characization has spanned the range from fundamental color msurement techniques to the specification of a variety ofvices including CRT, LCD, and projection displayscanners and digital cameras, and various film recordand print media. Some of the concepts and results ofresearch have been summarized by Berns.4 Such capabili-ties are a fundamental requirement for research and deopment in color and image appearance. Research on decharacterization and calibration provides a means to tamore fundamental problems in device-independent coimaging. For example, conceptual research on designimplementation of device-independent color imagin5

gamut mapping algorithms to deal with the reproductiondesired colors that fall outside the range that can betained with a given imaging device,6 and computer graphicrendering of high-quality spectral images that significanimprove the potential for accurate color in renderscenes.7 This type of research built on, and contributedresearch on the development and testing of color appance models for cross-media image reproduction.

1.2 Color Difference Equations

Color difference research has culminated with the recepublished CIEDE2000 color difference formula.8 A colordifference equation allows for the mapping of physicameasured stimuli into perceived differences. At the hearsuch color difference equations lies some form of unifocolor space. The CIE initially recommended two such cospaces in 1976, CIELAB and CIELUV. Both spaces weinitially described as interim color spaces, with the knowedge that they were far from complete. More than 25 yelater, these spaces are still the CIE recommendationsthough CIELUV has fallen out of favor.

With a truly uniform color space, color differences cathen be taken to be a simple measure of distance betwtwo colors in the space, such as CIEDEab* . The CIE rec-ognized the nonuniformity of the CIELAB color space, aformulated more advanced color difference equations sas CIE DE94 and CIEDE2000. These more complicaequations are very capable of predicting perceived codifferences of simple color patches.

1.3 Image Difference

The CIE color difference formulas were developed ussimple color patches in controlled viewing conditionThere is no reason to believe that they are adequatepredicting color difference for spatially complex imagstimuli. The S-CIELAB model was designed as a spapreprocessor to the standard CIE color difference eqtions, to account for complex color stimuli such as halftopatterns.9 Spatial preprocessing uses separable convolukernels to approximate the contrast sensitivity functio~CSF! of the human visual system. The CSF serves tomove information that is imperceptible to the visual sytem. For instance, when viewing halftone dots at a cerdistance, the dots tend to blur and integrate into a sin

.

-

--

-

gs

l-ceerd

-

r-

f

sl-

n

h

r

r

-

n

-

color. A pixel-by-pixel color difference calculation betweea continuous image and a halftone image would resulvery large errors, while the perceived difference mightfact be small. The spatial preprocessing would blurhalftone image, so that it more closely resembles the ctinuous tone image.

S-CIELAB represents the first incarnation of an imadifference model based on the CIELAB color space acolor difference equations. Recently, this model has brefined and extended into a modular framework for imacolor difference calculations.10 This framework refines theCSF equations from the S-CIELAB model, and adds moules for spatial frequency adaptation, spatial localizatiand local and global contrast detection. This frameworkdiscussed in more detail below.

1.4 Color Appearance

Unfortunately, fundamental CIE colorimetry does not prvide a complete solution for image specification. CIE corimetry is only strictly applicable to situations in which thoriginal and reproduction are viewed in identical condtions. By their very nature, the images produced or ctured by various digital systems are examined in widdisparate viewing conditions, from the original capturscene, to a computer display in a dim room, to printmedia under a variety of light sources, to projection dplays in dark rooms. Thus color appearance models wdeveloped to extend CIE colorimetry to the predictioncolor appearance~not just color matches! across changes inmedia and viewing conditions~not just within a single con-dition!. Color appearance modeling research applieddigital imaging systems was very active throughout t1990s, culminating with the recommendation of tCIECAM97s model in 199711 and its revision, CIECAM02,in 2002.12 Details on the evolution, formulation, and appcation of color appearance models can be foundFairchild.13 The development of these models was alsoabled by visual experiments performed to test the permance of published color-appearance models in realiimage reproduction situations.14 Such research on color appearance modeling in imaging applications naturally higlighted the areas that are not adequately addressed fortially complex image appearance and image quaproblems.

1.5 Image Appearance and Image Quality

Color appearance models account for many changeviewing conditions, but are mainly focused on changesthe color of the illumination~white point!, the illuminationlevel ~luminance!, and surround relative luminance. Sucmodels do not directly incorporate any of the spatialtemporal properties of human vision and the perceptionimages. They essentially treat each pixel of an image~andeach frame of a video! as completely independent stimulA review of some current work in the area provides cotext.

Visual adaptation to scenes and images is not only stially localized according to some low-pass characteristbut also temporally localized in a similar manner. To prdict the appearance of digital video sequences, particulthose of high-dynamic range, the temporal propertieslight and chromatic adaptation must be considered. To p

Journal of Electronic Imaging / January 2004 / Vol. 13(1) / 127

,cy

tial

antud

lexmalhe

tionthaanasseyede

anpti-msge

odear-

ofas

ferrchod

eoine

eopreeothl-

ith

rix

a

rialythe

retice-noare

ffd actsthe

thely

.ofo

thernothee

ein-it isictNDta.by

fi-si-ns-eatm-fil-of

rich-nooralere-is

ob-

ul inr-re-

ofatial

tced

ofionlityof

nce

-of

Fairchild and Johnson

dict the quality~or image differences! of video sequencestemporal filtering to remove imperceptible high-frequentemporal modulations~imperceptible flicker! must be addedto the spatial filtering that removes imperceptible spaartifacts~e.g., noise or compression artifacts!.

It is easy to illustrate that adaptation has a significtemporal low-pass characteristic. For example, if one sdenly turns on the lights in a darkened room~as when firstawakening in the morning!, the increased illumination leveis at first dazzling to the visual system, essentially overposing it. After a short period of time, the visual systeadapts to the new, higher level of illumination and normvisual perception becomes possible. The same is true wgoing from high levels of illumination to low levels~imag-ine driving into a tunnel in the daytime!. Fairchild andReniff15 and Rinner and Gegenfurtner16 have made detailedmeasurements of the time course of chromatic adaptaThese results suggest temporal integration functionscould be used in models of moving image appearance,also illustrate one of the mechanisms for spatially low-padaptation stimuli due to the influence of ever-presentmovements. Such adaptation stimuli are used in the modescribed in this work.

There has been significant research on video qualityvideo quality metrics, often aimed at the creation and omization of encoding/compression/decoding algorithsuch as MPEG2 and MPEG4. By analogy, the still-imavisible differences predictor of Daly17 is quite applicable tothe prediction of the visibility of artifacts introduced intstill images by JPEG image compression. The Daly mowas designed to predict the probability of detecting antifact ~i.e., is the artifact above the visual threshold!. TheiCAM work reviewed here and elsewhere18,19 has had adifferent objective with respect to image quality. Insteadfocusing on threshold differences in quality, the focus hbeen on the prediction of image quality scales~e.g., scalesof sharpness, contrast, graininess! for images with changeswell above threshold. Such suprathreshold image difences are a different domain of image quality reseabased on image appearance that separate the iCAM mfrom previous image quality models.

Likewise, a similar situation exists in the area of vidquality metrics. Metrics have been published to examthe probability of detection of artifacts in video~i.e., thresh-old metrics!, but there appears to be no models of vidimage appearance designed for rendering video anddicting the magnitudes of perceived differences in vidsequences. The latter is one of the ultimate goals ofdevelopment of iCAM. Two well-known video image quaity models, the Sarnoff just noticeable difference~JND!model and the NASA digital video quality~DVQ! model,are briefly described next to contrast their capabilities wthe proposed extensions to the iCAM model.

The Sarnoff JND model is the basis of the JNDmetsoftware package~see www.jndmetrix.com! and relatedvideo quality hardware. The model is briefly described intechnical report published by Sarnoff,19 and is more fullydisclosed in other publications.20 It is based on the multi-scale model of spatial vision published by Lubin,21,22 withsome extensions for color processing and temporal vation. The Lubin model is similar in nature to the Damodel mentioned before, in that it is designed to predict


-

-

n

.td

l

d

l

-,el

-

e

-

probability of detection of artifacts in images. These athreshold changes in images often referred to as just noable differences, or JNDs. The Sarnoff JND model hasmechanisms of chromatic and luminance adaptation, asincluded in the iCAM model. The input to the Sarnomodel must first be normalized, which can be considerevery rudimentary form of adaptation. The temporal aspeof the Sarnoff model are also not aimed at predictingappearance of video sequences, but rather at predictingdetectability of temporal artifacts. As such, the model onuses two frames~four fields! in its temporal processingThus, while it is capable of predicting the perceptibilityrelatively high-frequency temporal variation in the vide~flicker!, it cannot predict the visibility of low-frequencyvariations that would require an appearance-oriented, rathan JND-oriented, model. The Sarnoff model also isdesigned for rendering video. This is not a criticism of tmodel formulation, but an illustration of how the objectivof the Sarnoff JND model is significantly different fromthat of the iCAM model. While it is well accepted in thvision science literature that JND predictions are not learly related to suprathreshold appearance differences,certainly possible to use a JND model to try to predsuprathreshold image differences, and the Sarnoff Jmodel has been applied with some success to such da

A similar model, the DVQ metric has been publishedWatson, Hu, and McGowan23,24of NASA. The DVQ metricis similar in concept to the Sarnoff JND model, but signicantly different in implementation. Its spatial decompotion is based on the coefficients of a discrete cosine traformation ~DCT!, making it amenable to hardwarimplementation, and likely making it particularly gooddetecting artifacts introduced by DCT-based video copression algorithms. It also has a more robust temporalter that should be capable of predicting a wider arraytemporal artifacts. Like the Sarnoff model, the DVQ metis aimed at predicting the probability of detection of thresold image differences. The DVQ model also includesexplicit appearance processing through spatial or tempadaptation, or correlates of appearance attributes, and thfore also cannot be used for video rendering. Again, thisnot a shortcoming, but rather a property of the designjectives for the DVQ model.

While color appearance modeling has been successffacilitating device-independent color imaging and is incoporated into modern color management systems, theremains significant room for improvement and extensioncapabilities. To address these issues with respect to spproperties of vision and image perception~localized adap-tation and spatial filtering! and image quality, the concepof image appearance models has been recently introduand implemented.25,26 These models combine attributescolor appearance models with attributes of spatial vismodels that have been previously used for image quametrics in an attempt to further extend the capabilitiescolor-appearance models. Historically, color-appearamodels largely ignored spatial vision~e.g., CIECAM97s!,while spatial vision models for image quality largely ignored color.17,21 One notable exception, and the themethis special issue of theJournal of Electronic Imaging, hasbeen the Retinex model27–30 and its variousderivatives.31–33 The spatial ATD model34 and the

todequata-oseal

hain

ringuctly

in

e-ndfob

ta-

ceasnean

tiaforessdees

gemerr toheeasncetedain

lat, itncas

s tlitydsteer.

form-

an

M.ed-ivea

soto

luteondas-

ith

re-a

n-

-fica-sedand

-asson,cal

enpri-

nts,

orrmledallticlica-

g-


S-CIELAB model9 also address some of these issuesvarious extents. While the Retinex model was neversigned as a complete model of image appearance andity, its spatially variable mechanisms of chromatic adaption and color constancy serve some of the same purpin image rendering, and provide some of the criticgroundwork for image-appearance modeling.

The goal in developing an image appearance modelbeen to bring these research areas together to create a smodel applicable to image appearance, image rendeand image quality specifications and evaluations. One smodel for still images, referred to as iCAM, has recenbeen published by Fairchild and Johnson25 and is detailedin this work. This model was built on previous researchuniform color spaces,35 the importance of imagesurround,36 algorithms for image difference and imagquality measurement,19,37 insights into observers eye movements while performing various visual imaging tasks aadaptation to natural scenes,38,39 and an earlier model ospatial and color vision applied to color appearance prlems and high-dynamic-range~HDR! imaging.40 The struc-ture of the iCAM model and examples of its implemention for image appearance are presented next.

1.6 Color and Image Appearance Models

A model capable of predicting perceived color differenbetween complex image stimuli is a useful tool, but hsome limitations. Just as a color appearance model isessary to fully describe the appearance of color stimuli,image appearance model is necessary to describe spacomplex color stimuli. Color appearance models allowthe description of attributes such as lightness, brightncolorfulness, chroma, and hue. Image appearance moextend on this to also predict such attributes as sharpngraininess, contrast, and resolution.

A uniform color space also lies in the heart of an imaappearance model. The modular image difference frawork allows for great flexibility in the choice of colospaces. Examples are the CIELAB color space, similaS-CIELAB, the CIECAM02 color appearance model, or tIPT color space.12,35 Thus the modular image differencframework can be implemented within the iCAM model,described in this work, to create a full image appearaand image difference model. It could also be implemenin other color spaces if desired. This is one of the mbenefits of its modularity.

Models of image appearance can be used to formumultidimensional models of image quality. For exampleis possible to take weighted sums of various appearaattributes to determine a metric of overall image quality,described by Keelen41 and Engledrum.42 Essentially, thesemodels can augment or replace human observationweight image attributes with overall appearances of quaFor instance, a model of quality might involve weightesums of tonal balance, contrast, and sharpness. A firsttoward this type of model is illustrated in more detail lat

2 iCAM Framework

Figure 1 presents a flow chart of the general frameworkthe iCAM image appearance model as applied to still iages originally presented by Fairchild and Johnson.25 Fu-ture updates to the model, along with example images

-l-

s

sgle,

h

-

c-

lly

,lss,

-

e

e

o.

p

d

source code can be found at www.cis.rit.edu/mcsl/iCAFor input, the model requires colorimetrically characterizdata for the image~or scene! and surround in absolute luminance units. The image is specified in terms of relatCIE XYZ tristimulus values. The adapting stimulus islow-pass filtered version of the CIE XYZ image that is altagged with absolute luminance information necessarypredict the degree of chromatic adaptation. The absoluminances (Y) of the image data are also used as a seclow-pass image to control various luminance-dependantpects of the model intended to predict the Hunt effect~in-crease in perceived colorfulness with luminance! and theStevens effect~increase in perceived image contrast wluminance!. Last, a low-pass, luminance (Y) image of sig-nificantly greater spatial extent is used to control the pdiction of image contrast that is well established to befunction of the relative luminance of the surrounding coditions ~Bartleson and Breneman equations!. Refer toFairchild13 for a full discussion of the various image appearance effects mentioned earlier and detailed specitions of the data required. The specific low-pass filters ufor the adapting images depend on viewing distanceapplication. Additionally, in some image rendering circumstances, it might be desirable to have different low-padapting images for luminance and chromatic informatito avoid desaturation of the rendered images due to lochromatic adaptation~decrease in visual sensitivity to thcolor of the stimulus!. This is one example of applicatiodependence. Local chromatic adaptation might be approate for image difference or image quality measuremebut inappropriate for image rendering situations.

The first stage of processing in iCAM is to account fchromatic adaptation. The chromatic adaptation transfoembedded in the recently published CIECAM02 mode12

has been adopted in iCAM, since it was well researchand established to have excellent performance withavailable visual data. It is also a relatively simple chromaadaptation model, amenable to image processing apptions. The chromatic adaptation model, given in Eqs.~1!–~6!, is a linear von Kries normalization of RGB image sinals to the RGB adaptation signals derived from

F RGBG5MCAT02F X

YZG , ~1!

MCAT025F 0.7328 0.4296 20.1624

20.7036 1.6975 0.0061

0.0030 0.0136 0.9834G , ~2!

D5FF12S 1

3.6DexpS 2LA242

92 D G , ~3!

Rc5F S 100D

RWD1~12D !GR, ~4!

Gc5F S 100D

GWD1~12D !GG, ~5!



130 / Journal of Ele

Fig. 1 Flowchart of the iCAM image appearance model. Inputs to the model are CIE tristimulus valuesXYZ for the stimulus image or scene and a low-pass version used as an adapting stimulus andabsolute luminance information for the low-pass image and surround. Adapted signals are computedusing the linear chromatic adaptation transform from CIECAM02, and are then converted into anopponend space IPT using the luminance information to modulate a compressive nonlinearity. Therectangular IPT coordinates are then converted to cylindrical correlates of lightness J, chroma C, andhue h. The lightness and chroma correlates can then be scaled by a function of the absolute lumi-nance information to provide correlates of brightness Q and colorfulness M. If desired, a saturationcorrelate can be computed as the ratio of chroma to lightness (or colorfulness to brightness).

ionar1-

iesf-n

in

-,ticlorserng

in-putIt

Bc5F S 100D

BWD1~12D !GB, ~6!

with the low-pass adaptation image at each pixel locat(RWGWBW). The RGB signals are computed using a linetransformation from XYZ to RGB, derived by CIE TC8-0in the formulation of CIECAM02. This matrix transformation has come to be called theMCAT02 matrix, where CATstands for chromatic adaptation transform. The von Krnormalization is further modulated with a degree-oadaptation factorD that can vary from 0.0 for no adaptatioto 1.0 for complete chromatic adaptation. Equation~3! isprovided in the CIECAM02 formulation, and is used

ctronic Imaging / January 2004 / Vol. 13(1)

iCAM for computation ofD as a function of adapting luminanceLA for various viewing conditions. Alternativelythe D factor can be established manually. The chromaadaptation model is used to compute corresponding cofor CIE Illuminant D65, which are then used in the latstages of the iCAM model. This is accomplished by takithe adapted signals for the viewing condition RCGCBC andthen inverting Eqs.~1!–~6! for an illuminant D65 adaptingwhite point, and withD51.0. It should be noted that, whilethe adaptation transformation is identical to thatCIECAM02, the iCAM model is already significantly different, since it uses spatially modulated image data as inrather than single color stimuli and adaptation points.

n,telu-

apcewathe

Ban

ing-s omao-chan-

foren

cee is2.

dace

onrealentsion

cal-

healkeyrd-

ly-ge-.M

andlso

ngesu-

isa totootedisw-anen--

nofis

ce,r-re-sedr-the

otrv-

fiedane-

m-l


also differs completely in the remainder of the formulatioalthough using CIECAM02 equations where appropriaOne example of this is the modulation of the absoluteminance image and surround luminance image using theFLfunction from CIECAM02 given in Eq.~7!. This function,slowly varying with luminance, has been

FL50.2F 1

~5LA11!G4

~5LA!

10.1H 12F 1

~5LA11!G4J 2

~5LA!1/3, ~7!

established to predict a variety of luminance-dependentpearance effects in CIECAM02 and earlier models. Sinthe function has been established and understood, italso adopted for the early stages of iCAM. However,manner in which theFL factor is used in CIECAM02 andiCAM are quite different.

The next stage of the model is to convert from RGsignals ~roughly analogous to cone signals in the humvisual system! to opponent-color signals~light-dark, red-green, and yellow-blue; analogous to higher level encodin the human visual system! that are necessary for constructing a uniform perceptual color space and correlatevarious appearance attributes. In choosing this transfortion, simplicity, accuracy, and applicability to image prcessing were the main considerations. The color spacesen was the IPT space previously published by EbnerFairchild.35 The IPT space was derived specifically for image processing applications to have a relatively simplemulation, and specifically, to have a hue-angle componwith good prediction of constant perceived hue~importantin gamut mapping applications!. More recent work on per-ceived hue has validated the applicability of the IPT spaThe transformation from RGB to the IPT opponent spacfar simpler than the transformations used in CIECAM0The process, expressed in Eqs.~9!–~12!, involves a lineartransformation to a different cone response space~a differ-ent RGB!, application of power function nonlinearities, anthen a final linear transformation to the IPT opponent sp~I is light-dark, P is red-green, and T is yellow-blue!.

F LMSG5F 0.4002 0.7075 20.0807

20.2280 1.1500 0.0612

0.0 0.0 0.9184G FXD65

YD65

ZD65

G , ~8!

L85L0.43; L>0~9!

L852uLu0.43; L<0,

M 85M0.43; M>0~10!

M 852uM u0.43; M<0,

S85S0.43; S>0~11!

S852uSu0.43; S<0,

.

-

s

f-

o-d

-t

.

F IPTG5F 0.4000 0.4000 0.2000

4.4550 24.8510 0.3960

0.8056 0.3572 21.1628G F L8

M 8S8

G . ~12!

The power function nonlinearities in the IPT transformatiare a critical aspect of the iCAM model. First, they anecessary to predict response compression that is previn most human sensory systems. This response compreshelps to convert from signals that are linear in physimetrics ~e.g., luminance! to signals that are linear in perceptual dimensions~e.g., lightness!. The CIECAM02model uses a hyperbolic nonlinearity for this purpose. Tbehavior is that of a power function over the practicranges of luminance levels encountered. Second, and acomponent of iCAM, the exponents are modulated accoing to the luminance of the image~low-pass filtered! andthe surround. This is essentially accomplished by multiping the base exponent in the IPT formulation by the imawise computedFL factors with appropriate normalizationThese modulations of the IPT exponents allow the iCAmodel to be used for predictions of the Hunt, Stevens,Bartleson/Breneman effects mentioned before. They ahappen to enable the tone mapping of high-dynamic-raimages into low-dynamic range display systems in a vially meaningful way.

For image difference and image quality predictions, italso necessary to apply spatial filtering to the image dateliminate any image variations at spatial frequencieshigh to be perceived. For example, the dots in a prinhalftone image are not visible if the viewing distancesufficiently large. This computation is dependent on vieing distance and is based on filters derived from humcontrast sensitivity functions. Since the human contrast ssitivity functions vary for luminance~bandpass with sensitivity to high frequencies! and chromatic~low pass! infor-mation, it is appropriate to apply these filters in aopponent space. Thus in image quality applicationsiCAM, spatial filters are applied in the IPT space. Since itappropriate to apply spatial filters in a linear signal spathey are applied in a linear version of IPT prior to convesion into the nonlinear version of IPT for appearance pdictions. Johnson and Fairchild have recently discussome of the important considerations for this type of filteing in image difference applications, and have specifiedfilters used based on available visual data.18,43 Since thespatial filtering effectively blurs the image data, it is ndesirable for image rendering applications in which obseers might view the images more closely than the speciviewing distance. The result would be a blurrier image ththe original. The contrast sensitivity functions used to dfine spatial filters currently used for image difference coputations are given in Eq.~13! for the luminance I channeand Eq.~14! for the chromatic P and T channels,10 respec-tively.

csflum~ f !5a• f c• exp~2b• f !, ~13!

csfchrom~ f !5a1• exp~2b1• f c1!1a2• exp~2b2• f c2!. ~14!


-

he

rc

to

en

nnd

esgotfoce

le-d.agela

ise

encicsasa

so

thero-

the

is-ionrmon

romb-

onge

e-

ceces/e-foronra-m.ing,

bytialen-eral

balithins

pa-rs,nh-ralisheith

passpa-ndm-thd in

,

the

esanden-


Equations~13! and~14! were derived from fits to a collection of experimental data.10 The parametersa, b, andc inEq. ~13! are set to 75, 0.2, and 0.8, respectively, for tluminance CSF, applied to the I channel. In Eqs.~13! and~14!, spatial frequencyf is defined in terms of cycles pedegree of visual angle~CPD!. For the red-green chromatiCSF, applied to the P dimension, the parameters (a1 , b1 ,c1 , a2 , b2 , c2) in Eq. ~14! are set to~109.14,20.00038,3.424, 93.60,20.00367, 2.168!. For the blue-yellow chro-matic CSF, applied to the T dimension, they are set~7.033, 0.000004, 4.258, 40.69,20.10391, 1.6487!.

It is only appropriate to apply these spatial filters whthe goal is to compute perceived image differences~andultimately image quality!. This is an important distinctionbetween spatially localized adaptation~good for renderingand image quality metrics! and spatial filtering~good forimage quality metrics, bad for rendering!. In image qualityapplications, the spatial filtering is typically broken dowinto multiple channels for various spatial frequencies aorientations. For example, Daly,17 Lubin,22 and Pattanaiket al.40 describe such models. More recent results suggthat while such multiscale and multiorientation filterinmight be critical for some threshold metrics, it is often nnecessary for data derived from complex images andsuprathreshold predictions of perceived image differen~one of the main goals of iCAM!.10,44,45Thus, to preservethe simplicity and ease of use of the iCAM model, singscale spatial filtering with anisotropic filters was adopte

Once the IPT coordinates are computed for the imdata, a simple coordinate transformation from rectanguto cylindrical coordinates is applied to obtain image-wpredictors of lightness (J), chroma (C), and hue angle(h), as shown in Eqs.~15!, ~16!, and ~17!. Differences inthese dimensions can be used to compute image differstatistics and those used to derive image quality metrThe overall Euclidean difference in IPT is referred toD Im ~Eq. 20! for image difference, to distinguish it fromtraditional color difference metricDE, which includes nospatial filtering. In some instances, correlates of the ablute appearance attributes of brightness (Q) and colorful-ness (M ) are required. These are obtained by scalingrelative attributes of lightness and chroma with the apppriate function ofFL ~based on CIECAM02!, derived fromthe image-wise luminance map, as shown in Eqs.~18! and~19!.

J5I , ~15!

C5AP21T2, ~16!

h5tan21S P

T D , ~17!

Q5A4 FLJ, ~18!

M5A4 FLC, ~19!

D Im5~DI 21DP21DT2!1/2. ~20!


t

rs

r

e.

-

For image rendering applications, it is necessary to takecomputed appearance correlates~JCh! and then renderthem to the viewing conditions of a given display. The dplay viewing conditions set the parameters for the inversof the IPT model and the chromatic adaptation transfo~all for an assumed spatially uniform display adaptatitypical of low-dynamic-range output media!. This inversionallows the appearance of original scenes or images fdisparate viewing conditions to be rendered for the oserver viewing a given display. One important applicatiof such rendering is the display of high-dynamic-ran~HDR! image data on typical displays.

3 Modular Image Difference Model

A framework for a color image difference metric has rcently been described.10,43 In this work, the modular imagedifference metric is incorporated into the iCAM appearanmodel to address both image appearance and differenquality within a single model. The image difference framwork was designed to be modular in nature, to allowflexibility and adaptation. The framework itself is basedthe S-CIELAB spatial extension to the CIELAB colospace. S-CIELAB merges traditional color difference equtions with spatial properties of the human visual systeThis was accomplished as a spatial filtering preprocessbefore a pixel-by-pixel color difference calculation.9

The modular framework further extends this ideaadding several processing steps, in addition to the spafiltering. These processing steps are contained in indepdent modules, so they can be tested and refined. Sevmodules have been defined,45 and include spatial filtering,adaptation, and localization, as well as local and glocontrast detection. Figure 2 shows a general flowchart wseveral distinct modules. These modules and their origare described briefly in the following.

3.1 Spatial Filtering

The behavior of the human visual system in regards to stially complex stimuli has been well studied over the yeadating back to the seminal work of Campbell and Robso46

and Mullen.47 Summaries of current knowledge and tecniques for quantifying spatial vision can be found in sevebooks.48–50 The contrast sensitivity function describes thbehavior in relation to spatial frequency. Essentially, tCSF is described in a postretinal opponent color space, wa bandpass nature for the luminance channel and low-nature for the chrominance channels. S-CIELAB uses serable convolution kernels to approximate the CSF, amodulate image details that are imperceptible. More coplicated contrast sensitivity functions, that include bomodulation and frequency enhancement, were discussedetail by Johnson and Fairchild.43 Other models with simi-lar features include the previously mentioned Lubin21

Daly,17 MOM,40 S-CIELAB,9 and spatial ATD34 models.Other relevant discussions and models can be found inwork of Li,51 Taylor et al.,52,53 and Brill’s extension of theLubin/Sarnoff model.54

3.2 Spatial Frequency Adaptation

The contrast sensitivity function in this framework servto modulate spatial frequencies that are not perceptible,enhance certain frequencies that are most perceptible. G

ulipa

ty tvi

re-nd

ti-pahatio

inod

d in

ibedofndi-ges

e to-ac-

om-asage

rnela-h-e-ell

astg inandn-d bytorvesthe

or-

ordefulngisble

s, soean,icsap-erallold

-ec-es-for

ende

n-


erally, CSFs are measured using simple grating stimwith care taken to avoid spatial frequency adaptation. Stial frequency adaptation essentially decreases sensitivicertain frequencies based on information present in thesual field. An early and classic description of spatial fquency adaptation was published by Blakemore aCampbell.55 It should be noted that a multiscale, or mulchannel, spatial vision model is not required to predict stial frequency adaptation. Instead, all that is required is tthe CSF functions be allowed to change shape as a funcof adaptation~clearly indicating multiscale mechanismsthe human visual system not necessary for practical meling!.

Fig. 2 Flowchart of a modular image difference metric.

,-o-

-tn

-

Since spatial frequency adaptation cannot be avoidereal world viewing conditions~as it often is in the contrivedpsychophysical stimuli used to study spatial vision!, severalmodels of spatial frequency adaptation have been descrfor practical applications.43 These models alter the naturethe CSF based on either assumptions of the viewing cotions, or based on the information contained in the imathemselves.

3.3 Spatial Localization

The bandpass and low-pass contrast sensitivity servmodulate high-frequency information, including highfrequency edges. The human visual system is generallyknowledged to be very adept at detecting edges. To accmodate this behavior, a module of spatial localization hbeen developed. This module can be as simple as an improcessing edge-enhancing kernel, although that kemust change as a function of viewing distance. Alterntively, the CSF can be modified to boost certain higfrequency information. The formulation and utility of edgdetection algorithms in vision applications has been wdescribed by Marr.56

3.4 Local Contrast Detection

This module serves to detect local and global contrchanges between images. The utility of such processinreal visual systems has been described by TolhurstHeeger.57 The current implementation is based on the nolinear mask-based local contrast enhancement describeMoroney.58 Essentially, a low-pass image mask is usedgenerate a series of tone-reproduction curves. These cuare based on the global contrast of the image, as well asrelationship between a single pixel and its local neighbhood.

3.5 Color Difference Map

The output of the modular framework is a map of coldifferences" Im, corresponding to the perceived magnituof error at each pixel location. This map can be very usefor determining specific causes of error, or for detectisystematic errors in a color imaging system. Often, ituseful to reduce the error map into a more manageadataset. This can be accomplished using image statisticlong as care is taken. Such statistics can be image mmax, median, or standard deviation. Different statistmight be more valuable than others, depending on theplication, as perhaps the mean error better describes ovdifference, while the max might better describe threshdifferences.

4 Image Appearance Applications „Rendering …

Figure 3 illustrates implementation of the iCAM framework required to complete an image rendering process nessary for HDR image tone mapping. The componentssential in this process are the inversion of the IPT modela single set of spatially constant viewing conditions~thedisplay!, and the establishment of spatial filters for thadapting stimuli used for local luminance adaptation amodulation of the IPT exponential nonlinearity. While thderivation of optimal model settings for HDR image re



134 / Journal of Ele

Fig. 3 Implementation of iCAM for tone mapping of HDR images.

een

re-h a

co

nssis

ndterlennc

PTaccona-al

cem

lorheelle

eratr-el’s

i-ercelymoreor-

reof.

k-s-in

dering is still underway, quite satisfactory results have bobtained using the settings outlined in Fig. 3.

The iCAM model has been successfully applied to pdiction of a variety of color appearance phenomena, succhromatic adaptation~corresponding colors!, color appear-ance scales, constant hue perceptions, simultaneoustrast, crispening, spreading, and image rendering.25

Since iCAM uses the same chromatic adaptation traform as CIECAM02, it performs identically for situationin which only a change in state of chromatic adaptationpresent~i.e., change in white point only!. CIE TC8-01 hasworked very hard to arrive at this adaptation transform, ait is clear that no other model currently exists with betperformance, although there are several with equivaperformance. Thus the chromatic adaptation performaof iCAM is as good as possible at this juncture.

The appearance scales of iCAM are identical to the Iscales for the reference viewing conditions. The IPT sphas the best available performance for constant huetours, and thus this feature is retained in iCAM. This feture makes accurate implementation of gamut mappinggorithms far easier in iCAM than in other appearanspaces. In addition, the predictions of lightness and chro

ctronic Imaging / January 2004 / Vol. 13(1)

s

n-

-

te

e-

-

a

in iCAM are very good and comparable with the best coappearance models in typical viewing conditions. Tbrightness and colorfulness scales will also perform as was any other model for typical conditions. In more extremviewing conditions, the performance of iCAM and othmodels will begin to deviate. It is in these conditions ththe potential strengths of iCAM will become evident. Futher visual data must be collected to evaluate the modrelative performance in such situations.

The color difference performance of iCAM will be simlar to that of CIELAB, since the space is very similar undthe reference viewing conditions. Thus, color differencomputations will be similar to those already commonused, and the space can be easily extended to have aaccurate difference equation following the successful fmat of the CIE94 equations.~Following the CIEDE2000equations in iCAM is not recommended, since they aextremely complex and fitted to particular discrepanciesthe CIELAB space, such as poor constant hue contours!

Simultaneous contrast~or induction! causes a stimulusto shift in appearance away from the color of the bacground in terms of opponent dimensions. Figure 4 illutrates a stimulus that exhibits simultaneous contrast

Fig. 4 (a) Original stimulus and (b) iCAM lightness J image illustrating the prediction of simultaneouscontrast.


Fig. 5 (a) Original stimulus and (b) iCAM chroma C image illustrating the prediction of chroma crisp-ening. Original image from www.hpl.hp.com/personal/Nathan–Moroney/.

eed

i-

er-en

ig-nge-of

neae 6

oto

s-re-

ns-are-

ion

po-ghts ofs-orehating

oftheic

is tot toe.ro-err

at

uldm-

isthe

lightness~the gray square is physically identical on all thrbackgrounds! and its prediction by iCAM, as representeby the iCAM lightness predictor. This prediction is faciltated by the local adaptation features of iCAM.

Crispening is the phenomenon whereby the color diffences between two stimuli are perceptually larger whviewed on a background that is similar to the stimuli. Fure 5 illustrates a stimulus that exhibits chroma crispeniand its prediction by the iCAM chroma predictor. This prdiction is also facilitated by the local adaptation featuresiCAM.

Spreading is a spatial color appearance phenomenowhich the apparent hue of spatially complex image arappears to fill various spatially coherent regions. Figurprovides an example of spreading in which the red huethe annular region spreads significantly from the linesthe full annulus. The iCAM prediction of spreading is illutrated through reproduction of the hue prediction. The pdiction of spreading in iCAM is facilitated by spatial filtering of the stimulus image.

One of the most interesting and promising applicatioof iCAM is to the rendering of HDR images to lowdynamic-range display systems. HDR image dataquickly becoming more prevalent. Historically, HDR images were obtained through computer graphics simulatcomputed with global illumination algorithms~e.g., ray

,

ins

f

-

s

tracing or radiosity algorithms! or through the calibrationand registration of images obtained through multiple exsures. Real scenes, especially those with visible lisources, often have luminance ranges of up to six ordermagnitude. More recently, industrial digital imaging sytems have become commercially available that can measily capture HDR image data. It is also apparent tconsumer digital cameras will soon be capable of capturgreater dynamic ranges. Unfortunately, display and usesuch data are difficult and will remain so, since evenhighest quality displays are generally limited in dynamrange to about two orders of magnitude. One approachinteractively view the image and select areas of interesbe viewed optimally within the display dynamic rangThis is only applicable to computer displays and not apppriate for pictorial imaging and printed output. Anothlimitation is the need for capability to work with greatethan 24-bit~and often floating point! image data. It is de-sirable to render HDR pictorial images onto a display thcan be viewed directly~no interactive manipulation! by theobserver, and appear similar to what the observer woperceive if the original scene was viewed. For printed iages, it is not just desirable, but necessary. Pattanaiket al.40

review several such HDR rendering algorithms, and itworth noting that several papers were presented on

Fig. 6 (a) Original stimulus and (b) iCAM hue h image illustrating the prediction of spreading.


ndtheg,re-n

vi-en

thab

l

arheof

geixth

epthofin

tes-ionadyith

esgeoues-ntof

lly

therful-ite

in

to-er

rre-n a

ce-ce,earcanari-at-is

.r-for-is

ious-

erfor

tionof

aaledr-

m-

sultsnalility

ge

ngcera-e-s an


topic at SIGGRAPH 2002,59–61 illustrating continued inter-est in the topic.

Since iCAM includes spatially localized adaptation aspatially localized contrast control, it can be applied toproblem of HDR image rendering. This is not surprisinsince the fundamental problem in HDR rendering is toproduce the appearance of an HDR image or scene olow-dynamic-range display. Since the encoding in oursual system is of a rather low dynamic range, this is esstially a replication of the image appearance processinggoes on in the human observer and is being modelediCAM. Figure 7 illustrates application of the iCAM modeto HDR images obtained from Debevec~see www.debev-ec.org!. The images in the left column of Fig. 7 are linerenderings of the original HDR data normalized to tmaximum presented, simply to illustrate how the rangethe original data exceeds a typical 24-bit~8-bits per RGBchannel! image display. For example, the memorial imadata ~top row! have a dynamic range covering about sorders of magnitude, since the sun was behind one ofstained-glass windows. The middle column of images rresents a typical image processing solution to renderingdata. One might consider a logarithmic transformationthe data, but that would do little to change the renderingthe first column. Instead, the middle column was generainteractively by finding the optimum power function tranformation~also sometimes referred to as gamma correctnote that the linear images in the first column are alregamma corrected!. For these images, transformations wexponents, or gammas, of approximately 1/6~as opposed to1/1.8 to 1/2.2 for typical displays! were required to makethe image data in the shadow areas visible. While thpower function transformations do make more of the imadata visible, they required user interaction, tend to washthe images in a way not consistent with the visual imprsion of the scenes, and introduce potentially severe quazation artifacts in shadow regions. The rightmost column

Fig. 7 Three HDR images from www.debevec.org. The leftmost col-umn illustrates linear rendering of the image data, the middle col-umn illustrates manually optimized power function transformations,and the rightmost column represents the automated output of theiCAM model implemented for HDR rendering (see Fig. 3).


a

-ty

e-e

d

;

e

t

i-

images shows the output of the iCAM model with spatialocalized adaptation and contrast control~as shown in Fig.3!. These images both render the dynamic range ofscene to make shadow areas visible and retain the coloness of the scene. The resulting iCAM images are quacceptable as reproductions of the HDR scenes~equivalentto the result of dodging and burning historically donephotographic printing!. It is also noteworthy that theiCAM-rendered images were all computed with an aumated algorithm mimicking human perception with no usinteraction.

5 Image Quality Applications „DifferencePerceptibility …

A slightly different implementation of iCAM is required foimage quality applications to produce image maps repsenting the magnitude of perceived differences betweepair of images. In these applications, viewing-distandependent spatial filtering is applied in a linear IPT spaand then differences are computed in the normal nonlinIPT space. Euclidean summations of these differencesbe used as an overall image difference map, and then vous summary statistics can be used to predict differenttributes of image difference and quality. This processoutlined in Fig. 8 and detailed in Johnson and Fairchild45

Image quality metrics can be derived from image diffeence metrics that are based on normal color differencemulas applied to properly spatially filtered images. Thapproach has been used to successfully predict vartypes of image quality data.10 Figure 9 illustrates the prediction of perceived sharpness62 and contrast63 differencesin images through a single summary statistic~mean imagedifference!. This performance is equivalent to, or bettthan, that obtained using other color spaces optimizedthe task.10

The contrast results in Fig. 9~a! were obtained by askingobservers to scale perceived image contrast for a collecof images of various content, subjected to a varietytransformations.63 The resulting interval scale~averagedata! is plotted as perceived contrast in Fig. 9~a!, and themodel prediction of image difference from the original~ar-bitrarily selected! is compared with it. Ideally, the datwould follow a V shape with two line segments of equabsolute slope on either side of the origin. The perceivcontrast data are well predicted by the iCAM image diffeence.

The perceived sharpness results in Fig. 9~b! were ob-tained in a similar manner using a significantly larger nuber of image manipulations and content.62 Observers weresimply asked to scale perceived sharpness, and the rewere converted to an interval scale, again with the origiimage as an arbitrary zero point. There is greater variabin these data, but it can be seen in Fig. 9~b! that the resultsare again well predicted by a fairly simple mean imadifference metric.

6 Conclusions and Future Directions

Advances in imaging and computing technologies, alowith increased knowledge of the function and performanof the human visual system, have allowed for the integtion of models of color, spatial, and temporal vision to crate a new type of color appearance model, referred to a


Fig. 8 Implementation of iCAM for image difference and image quality metrics.

inndhisge

itsal-in-ro-

ceteocaland

scedu/nd

,

lo-

ent

ingfor

ns

0

r

f-

d

e

n

image appearance model. Such models show promisevariety of applications, ranging from image difference aimage quality metrics to the rendering of image data. Twork describes the framework of one example of an ima

Fig. 9 iCAM image differences as a function of (a) perceived imagecontrast and (b) perceived image sharpness for a variety of imagetransformations. (Note that desired predictions are V-shaped datadistributions, since the perceptual differences are signed and thecalculated differences are unsigned.)

aappearance model, referred to as iCAM, and illustratesapplicability to HDR image tone mapping and image quity metrics. Recently, initial efforts have been made tocorporate psychophysical data on the time course of chmatic adaptation15 to extend the model to video appearanand quality applications.64 Future efforts will be directed acompletion of the spatio-temporal filters required for viddifference metrics, the collection of more psychophysidata on image and video appearance and differences,the formulation of specific iCAM algorithms for variouapplications. The iCAM model is not proprietary. Sourcode and updates are freely available at www.cis.rit.emcsl/iCAM for those interested in evaluating the model apotentially suggesting improvements.

References1. R. W. G. Hunt,The Reproduction of Colour, 5th ed., Fountain Press

England~1995!.2. O. H. Schade, ‘‘Optical and photoelectric analog of the eye,’’J. Opt.

Soc. Am.46, 721–739~1956!.3. W. D. Wright, ‘‘50 years of the 1931 CIE standard observer for co

rimetry,’’ AIC Color 81, Paper A3~1981!.4. R. S. Berns, ‘‘A generic approach to color modeling,’’Color Res.

Appl. 22, 318–325~1997!.5. M. D. Fairchild, ‘‘Some hidden requirements for device-independ

color imaging,’’SID Intl. Symp.pp. 865–868~1994!.6. G. J. Braun and M. D. Fairchild, ‘‘General-purpose gamut-mapp

algorithms: Evaluation of contrast-preserving rescaling functionscolor gamut mapping,’’J. Imaging Sci. Technol.44, 343–350~2000!.

7. G. M. Johnson and M. D. Fairchild, ‘‘Full-spectral color calculatioin realistic image synthesis,’’IEEE Comput. Graphics Appl.19~4!,47–53~1999!.

8. M. R. Luo, G. Cui, and B. Rigg, ‘‘The development of the CIE 200colour difference formula,’’Color Res. Appl.26, 340–350~2001!.

9. X. M. Zhang and B. A. Wandell, ‘‘A spatial extension to CIELAB fodigital color image reproduction,’’Proc. SID Symp.pp. 731–734~1996!.

10. G. M. Johnson and M. D. Fairchild, ‘‘Darwinism of color image diference models,’’Proc. of IS&T/SID 9th Color Imag. Conf., pp. 108–112 ~2001!.

11. CIE, ‘‘The CIE 1997 interim colour appearance model~simple ver-sion!,’’ CIECAM97s, CIE Pub. 131~1998!.

12. N. Moroney, M. D. Fairchild, R. W. G. Hunt, C. J. Li, M. R. Luo, anT. Newman, ‘‘The CIECAM02 color appearance model,’’IS&T/SID10th Color Imag. Conf.,pp. 23–27~2002!.

13. M. D. Fairchild,Color Appearance Models, Addison-Wesley, Read-ing, MA ~1998!.

14. K. M. Braun and M. D. Fairchild, ‘‘Testing five color appearancmodels for changes in viewing conditions,’’Color Res. Appl.22, 165–174 ~1997!.

15. M. D. Fairchild and L. Reniff, ‘‘Time-course of chromatic adaptatio


ap-

s-

of

c-m

D-In-

lysis

gni-

o

n

gi-

’’

ret-ob-

le

of

re-

lor

ent

m-

rv-

ics

erg,is-

-

n

tial

es,

to

-

-

e

gesual

m-

,’’

inand

st-ls in

,’’

h

of

ain

ved

g,’’


for color-appearance judgements,’’J. Opt. Soc. Am. A12, 824–833~1995!.

16. O. Rinner and K. R. Gegenfurtner, ‘‘Time course of chromatic adtation for color appearance discrimination,’’Vision Res.40, 1813–1826 ~2000!.

17. S. Daly, ‘‘The visible differences predictor: An algorithm for the asessment of image fidelity,’’ inDigital Images and Human Vision, A.Watson, Ed., pp. 179–206, MIT Press Cambridge, MA~1993!.

18. G. M. Johnson and M. D. Fairchild, ‘‘A top down descriptionS-CIELAB and CIEDE2000,’’Color Res. Appl.~in press!.

19. Sarnoff Corporation, ‘‘JND: A human vision system model for objetive picture quality measurement,’’ Sarnoff Technical Report frowww.jndmetrix.com~2001!.

20. ATIS, ‘‘Objective perceptual video quality measurement using a JNbased full reference technique,’’ Alliance for Telecommunicationsdustry Solutions Technical Report T1.TR.PP.75-2001~2001!.

21. J. Lubin, ‘‘The use of psychophysical data and models in the anaof display system performance,’’ inDigital Images and Human Vision,A. Watson, Ed., pp. 163–178, MIT Press, Cambridge, MA~1993!.

22. J. Lubin, ‘‘A visual discrimination model for imaging system desiand evaluation,’’ inVision Models for Target Detection and Recogntion, E. Peli, Ed., pp. 245–283, World Scientific, Singapore~1995!.

23. A. B. Watson, ‘‘Toward a perceptual video quality metric,’’Proc.SPIE3299, 139–147~1998!.

24. A. B. Watson, J. Hu, and J. F. McGowan, ‘‘DVQ: A digital videquality metric based on human vision,’’J. Electron. Imaging10,20–29~2001!.

25. M. D. Fairchild and G. M. Johnson, ‘‘Meet iCAM: A next-generatiocolor appearance model,’’IS&T/SID 10th Color Imag. Conf., pp.33–38~2002!.

26. M. D. Fairchild, ‘‘Image quality measurement and modeling for dital photography,’’International Congress on Imaging Science ’02, pp.318–319~2002!.

27. E. H. Land, ‘‘Recent advances in retinex theory,’’Vision Res.26,7–21 ~1986!.

28. E. H. Land and J. J. McCann, ‘‘Lightness and the retinex theory,J.Opt. Soc. Am.61, 1–11~1971!.

29. E. H. Land, ‘‘The retinex,’’Am. Sci.52, 247–264~1964!.30. J. J. McCann, S. McKee, and T. Taylor, ‘‘Quantitative studies in

inex theory: A comparison between theoretical predictions andserver responses to ‘color Mondrian’ experiments,’’Vision Res.16,445–458~1976!.

31. B. Funt, F. Ciurea, and J. J. McCann, ‘‘Retinex in Matlab,’’Proc. ofIS&T/SID 8th Color Imag. Conf.,pp. 112–121~2000!.

32. K. Barnard and B. Funt, ‘‘Analysis and improvement of multi-scaretinex,’’ Proc. 5th IS&T/SID Color Imag. Conf., pp. 221–226~1997!.

33. D. H. Brainard and B. A. Wandell, ‘‘Analysis the retinex theorycolor vision,’’ J. Opt. Soc. Am. A3, 1651–1661~1986!.

34. E. M. Granger, ‘‘Uniform color space as a function of spatial fquency,’’Proc. SPIE1913, 449–457~1993!.

35. F. Ebner and M. D. Fairchild, ‘‘Development and testing of a cospace~IPT! with improved hue uniformity,’’ IS&T/SID 6th ColorImag. Conf., pp. 8–13~1998!.

36. M. D. Fairchild, ‘‘Considering the surround in device-independcolor imaging,’’Color Res. Appl.20, 352–363~1995!.

37. M. D. Fairchild, ‘‘Modeling color appearance, spatial vision, and iage quality,’’ Color Image Science: Exploiting Digital Media, pp.357–370, Wiley and Sons, New York~2002!.

38. J. S. Babcock, J. B. Pelz, and M. D. Fairchild, ‘‘Eye tracking obseers during color image evaluation tasks,’’Proc. SPIE5007, 218–230~2003!.

39. M. A. Webster and J. D. Mollon, ‘‘Adaptation and the color statist


of natural images,’’Vision Res.37, 3283–3298~1997!.40. S. N. Pattanaik, J. A. Ferwerda, M. D. Fairchild, and D. P. Greenb

‘‘A multiscale model of adaptation and spatial vision for image dplay,’’ Proc. SIGGRAPH 98, pp. 287–298~1998!.

41. B. W. Keelan,Handbook of Image Quality: Characterization and Prediction, Marcel Dekker, New York~2002!.

42. P. G. Engledrum, ‘‘Extending image quality models,’’Proc IS&TPICS Conf., 65–69~2002!.

43. G. M. Johnson and M. D. Fairchild, ‘‘On contrast sensitivity in aimage difference model,’’Proc. IS&T PICS Conf., 18–23~2001!.

44. A. B. Watson and C. V. Ramirez, ‘‘A standard observer for spavision,’’ Investigative Ophthalmology Vis. Sci.41, S713~2000!.

45. G. M. Johnson and M. D. Fairchild, ‘‘Measuring images: Differencquality, and appearance,’’Proc. SPIE5007, 51–60~2003!.

46. F. W. Campbell and J. G. Robson, ‘‘Application of Fourier analysisthe visibility of gratings,’’Am. J. Physiol.197, 551–566~1968!.

47. K. T. Mullen, ‘‘The contrast sensitivity of human color vision to redgreen and blue-yellow chromatic gratings,’’Am. J. Physiol.359, 381–400 ~1985!.

48. R. L. DeValois and K. K. DeValois,Spatial Vision, Oxford UniversityPress, Oxford~1988!.

49. Visual Science and Engineering: Models and Applications, D. H.Kelly, Ed., Marcel Dekker, New York~1994!.

50. B. A. Wandell,Foundations of Vision, Sinauer Associates Inc., Sunderland, MA~1995!.

51. B. Li, G. W. Meyer, and R. V. Klassen, ‘‘A comparison of two imagquality models,’’Proc. SPIE3299, 98–109~1998!.

52. C. C. Taylor, Z. Pizlo, J. P. Allebach, and C. A. Bouman, ‘‘Imaquality assessment with a Gabor pysamid model of the human visystem,’’Proc. SPIE3016, 58–69~1997!.

53. C. C. Taylor, Z. Pizlo, and J. P. Allebach, ‘‘Perceptually relevant iage fidelity,’’ Proc. SPIE3299, 110–118~1998!.

54. M. H. Brill, ‘‘Color management: New roles for color transformsIS&T/SID 5th Color Imaging Conf., pp. 78–82~1997!.

55. C. Blakemore and F. W. Campbell, ‘‘On the existence of neuronsthe human visual system selectively sensitive to the orientationsize of retinal images,’’Am. J. Physiol.203, 237–260~1969!.

56. D. Marr,Vision, Freeman, New York~1982!.57. D. J. Tolhurst and D. J. Heeger, ‘‘Comparison of contra

normalization and threshold models of the responses of simple celcat striate cortex,’’Visual Neurosci.14, 293–309~1997!.

58. N. Moroney, ‘‘Local color correction using non-linear maskingProc. IS&T/SID 8th Color Imaging Conf., pp. 108–111~2000!.

59. R. Fattal, D. Lischinski, and M. Werman, ‘‘Gradient domain higdynamic range compression,’’Proc. SIGGRAPH 2002, pp. 249–256~2002!.

60. F. Durand and J. Dorsey, ‘‘Fast bilateral filtering for the displayhigh-dynamic-range images,’’Proc. SIGGRAPH 2002, pp. 257–266~2002!.

61. E. Reinhard, M. Stark, P. Shirley, and J. Ferwerda, ‘‘Gradient domhigh dynamic range compression,’’Proc. SIGGRAPH 2002, pp. 267–276 ~2002!.

62. G. M. Johnson and M. D. Fairchild, ‘‘Sharpness rules,’’Proc. IS&T/SID 8th Color Imaging Conf., pp. 24–30~2000!.

63. A. J. Calabria and M. D. Fairchild, ‘‘Compare and contrast: Perceicontrast of color images,’’Proc. IS&T/SID 10th Color Imaging Conf.,17–22~2002!.

64. M. D. Fairchild and G. M. Johnson, ‘‘Image appearance modelinProc. SPIE5007, 149–160~2003!.

Biographies and photographs of the authors not available.

iCAM framework for image appearance, differences, and quality

Documents

Transcript of iCAM framework for image appearance, differences, and quality