2024年5月11日发(作者：360杀毒清理大师)

IEEETRANSACTIONSONIMAGEPROCESSING,VOL.6,NO.3,MARCH1997451

PropertiesandPerformance

ofaCenter/SurroundRetinex

,Zia-urRahman,Member,IEEE,l

Abstract—ThelastversionofLand’sretinexmodelforhuman

vision’slightnessandcolorconstancyhasbeenimplementedand

usresearchhases-

tablishedthemathematicalfoundationsofLand’sretinexbuthas

notsubjectedhislightnesstheorytoextensiveimageprocessing

soughttodefineapracticalimplementation

oftheretinexwithoutparticularconcernforitsvalidityasa

describe

thetrade-offbetweenrenditionanddynamicrangecompression

r,unlike

previousresults,wefindthattheplacementofthelogarithmic

functionisimportantandproducesbestresultswhenplacedafter

likepreviousresults,wefindbest

renditionfora“canonical”gain/offsetappliedaftertheretinex

sfunctionalformsfortheretinexsurroundare

evaluated,andaGaussianformfoundtoperformbetterthanthe

thatviolatethegray

worldassumptions(implicittothisretinex)areinvestigatedto

provideinsightintocaseswherethisretinexfailstoproducea

goodrendition.

(a)

I.I

NTRODUCTION

O

FTHEMANYvisualtasksaccomplishedsogracefully

byhumanvision,oneofthemostfundamentaland

approachableformachinevisionapplicationsislightnessand

completelysatisfactorydefinitionis

lacking,lightnessandcolorconstancyrefertotheresilience

ofperceivedcolorandlightnesstospatialandspectralil-

stheoriesforthishavebeen

proposedandhaveacommonmathematicalfoundation[1].

ThelastversionofLand’sretinex[2]hascapturedouratten-

tionbecauseoftheeaseofimplementationandmanipulation

ofkeyvariables,andbecauseitdoesnothave“unnatural”

se,thesimplicity

ofthecomputationwasappealingandinitialexperiments

rsionoftheretinexhas

beenthesubjectofpreviousdigitalsimulationsthatwere

limitedbecauseoflengthycomputertimeinvolvedandwas

implementedinanalogverylarge-scaleintegratedcircuits

(VLSI)toachievereal-timecomputation[3],[4].Evidence

thatthisretinexversionisanoptimalsolutiontothelightness

problemhascomefromexperimentsposingLand’sMondrian

target,randomlyarrangedtwo-dimensional(2-D)graypatches,

ManuscriptreceivedJune26,1995;revisedMay24,kofZ.

RahmanwassupportedbyNASALangleyResearchCenterunderContract

ociateeditorcoordinatingthereviewofthismanuscript

Gabbouj.

larewithNASALangleyResearchCenter,

Hampton,VA23681-0001USA(e-mail:@).

iswithScienceandTechnologyCorporation,Hampton,VA

23666USA.

PublisherItemIdentifierS1057-7149(97)00428-4.

(b)

lformofthecenter/surroundretinexoperator.(a)3-Drepre-

sentation(distortedtovisualizesurround).(b)Cross-sectiontoillustratewide

weaksurround.

asaprobleminlinearoptimizationandalearningproblemfor

backpropagatedartificialneuralnetworks[5],[6].

Theutilityofalightness–colorconstancyalgorithmfor

machinevisionisthesimultaneousaccomplishmentof:

1)dynamicrangecompression;

2)colorindependencefromthespectraldistributionofthe

sceneilluminant;

3)colorandlightnessrendition.

Land’scenter/surroundretinexdemonstrablyachievesthe

firsttwo,althoughLandemphasizedprimarilythecolorcon-

-knowndifficultiesarise,though,for

1057–7149/97$10.00©1997IEEE

Authorized licensed use limited to: Nanjing Southeast University. Downloaded on May 26,2010 at 00:51:12 UTC from IEEE Xplore. Restrictions apply.

452IEEETRANSACTIONSONIMAGEPROCESSING,VOL.6,NO.3,MARCH1997

trationofretinexcolorconstancyanddynamicrangecompression(priortooptimizingrendition)foraGaussiansurroundwithsmall

spaceconstant(15pixels).

colorandlightnessrendition[1],[3],[6].Theseconsistof

i)lightnessandcolor“halo”artifactsthatareespecially

prominentwherelargeuniformregionsabuttoformahigh

contrastedgewith“graying”inthelargeuniformzonesinan

image,andii)globalviolationsofthegrayworldassumption

(e.g.,anall-redscene)whichresultinaglobal“graying

out”y,theretinex(perhapslikehuman

vision)functionsbestforhighlydiversescenesandpoorest

analogoustosystemsof

simultaneousequationswhereauniquesolutionexistsifand

onlyifthereareenoughindependentequations.

Thegeneralformofthecenter/surroundretinex(Fig.1)is

similartothedifference-of-Gaussian(DOG)functionwidely

usedinnaturalvisionsciencetomodelboththereceptive

fiy

extensionsrequiredarei)togreatlyenlargeandweakenthe

surroundGaussian(asdeterminedbyitsspaceandamplitude

constants),andii)toincludealogarithmicfunctiontomake

,arithmetic

division).WehavechosenaGaussiansurroundformwhereas

Landoptedfora

function[2]andMooreetal.[3]useda

illbecomparedinSection

atically,thistakestheform

thcolor

spectralband,“

isthesurroundfunction,and

where

(2)

Authorized licensed use limited to: Nanjing Southeast University. Downloaded on May 26,2010 at 00:51:12 UTC from IEEE Xplore. Restrictions apply.