• No results found

Evaluation of the Quality Indicators in Dehazed Images: Color, Contrast, Naturalness, and Visual Pleasingness

N/A
N/A
Protected

Academic year: 2022

Share "Evaluation of the Quality Indicators in Dehazed Images: Color, Contrast, Naturalness, and Visual Pleasingness"

Copied!
12
0
0

Laster.... (Se fulltekst nå)

Fulltekst

(1)

Contents lists available atScienceDirect

Heliyon

journal homepage:www.cell.com/heliyon

Research article

Evaluation of the quality indicators in dehazed images: Color, contrast, naturalness, and visual pleasingness

Laksmita Rahadianti

a,∗

, Aruni Yasmin Azizah

a

, Hilda Deborah

b

aFacultyofComputerScience,UniversitasIndonesia,Indonesia

bDepartmentofComputerScience,NorwegianUniversityofScienceandTechnology(NTNU),Norway

A R T I C L E I NF O A B S T R A C T

Keywords:

Dehazing Imagerestoration Hazyimages Imagequality Psychovisualexperiment

Hazyimagessufferfromlowqualityduetoblurring,veilingeffects,andlowcontrast.Toimprovetheirvisibility, dehazingmethods attempttorestorethem totheircorrespondingclear scenes,often byfocusingmoreon obtaininganaccurateestimatebasedonaknowngroundtruth.Theperceptualqualityofdehazedimages, whichcanbedescribedbymeansofobjectiveandsubjectivequalityassessments,isoftennotconsidered.This paperprovidesaqualityassessmentofdehazedimages,focusingonaspects,e.g.,color,imagestructure,and naturalness.Fourimagedehazingmethodsareconsidered,i.e.,ContrastLimitedAdaptedHistogramEqualization (CLAHE),Dark Channel Priorand Refinement(DCP-R),Perception InspiredDeep Dehazing Network with Refinement(PDR-Net)andConditionalGenerativeAdversarialNetwork(CGAN)Pix2pix.Thedehazingresults arethenputthroughobjectiveandsubjectiveassessments,foracomprehensiveevaluationonimagequality.

Overall,Pix2pixshowsthebestresults objectively,excellingin therecoveryof colorandimagestructure.

AlthoughitisoutperformedbyDCP-Rintermsofnaturalness,oursubjectiveassessmentshowsthatPix2pix isalsomostpreferredbyhumanobservers.

1. Introduction

Inadigitalimageacquisition,animagesensorcaptureslightrays reflected byobjects. Inclearconditions, these lightrays areableto travelunhinderedintheenvironment.However,inconditionssuchas fogandhaze,micro-particlesinthesurroundingmediainterferewith thepropagatinglightbychangingitsdirectionandintensity.Thistype ofmediaisoftencalledscatteringorparticipatingmediabecauseitplays anactiveroleinimagecapture.Animagecapturedinscatteringmedia, e.g.,hazyimages,willappearwithblurringeffects,lowcontrast,and low visibility.Theappearanceof hazyimagesmakesitverydifficult forobserverstoobtainusefulinformationof asceneandtheobjects init,makinginchallengingforimageunderstandingbybothhuman andcomputervision.Thus,itisoftennecessarytodehazeahazyimage toitscorrespondingclearscene [1]priortoanyfurtherprocessingor analysis.

Image dehazingmethodsseektorestorea hazyimagetoits esti- matedclearcounterpart,which canbe achievedby,e.g.,contrast or visibilityenhancementtechniques [2, 3].Otherapproachesstudythe physicalprocessofscatteringmodelandtheninvertitwiththehelpof

*

Correspondingauthor.

E-mailaddress:laksmita@cs.ui.ac.id(L. Rahadianti).

additionalconstraints,e.g.,statisticalorspatialpriors [4,5].Deepnet- workshavealsobeenusedtomodelthetranslationbetweenhazyand clearimagedomains [6,7].

Thispaperis builton apreliminarystudyof thequalitydehazed images by Azizah, et al. [8]. In this paper, we evaluate theimage qualityofdehazedimagesusingfourdifferentdehazingmethods.Con- trastLimitedAdaptiveHistogramEqualization(CLAHE)[9]isanimage enhancementtechnique,asopposedtoarestorationtechnique,toim- provequalityandvisibility.Thedarkchannelprior(DCP) [5]usesthe physicalscatteringmodelandattemptstoreverseittoobtaintheorig- inalclearimage.Then,twodeeplearningbaseddehazingmethodsare alsoconsidered. Perception-inspiredSingle Image Dehazing Network withRefinement(PDR-Net)[10]isthestate-of-the-artgenerativemodel thatusesaperception-basedapproachtotranslatehazyimagestotheir clearcounterparts.Pix2pix [11]isaConditionalGenerativeAdversar- ialNetwork(CGAN)thathasbeenusedextensivelyforimagetoimage translation.Inourcase,wewilluseittotranslatehazyimagestotheir clearcounterparts.

Dehazingresultsareoftenevaluatedbyhowclosetheyaretotheir knowngroundtruth.However,hazyimagescapturedinnaturalandreal

https://doi.org/10.1016/j.heliyon.2021.e08038

Received8June2021;Receivedinrevisedform13August2021;Accepted16September2021

2405-8440/©2021TheAuthor(s). PublishedbyElsevierLtd. ThisisanopenaccessarticleundertheCCBY-NC-NDlicense(http://creativecommons.org/li- censes/by-nc-nd/4.0/).

(2)

Fig. 1.Image formation model in scattering media.

settingsoftendonotcomewithgroundtruthclearimages,resultingin thelackofstandardizedanduniversaldatasets.Additionally,themajor- ityofexistingworksevaluatedehazedimagesbyvariationsofobjective qualitymetrics,makingbenchmarkingofthequalityofdehazedimages verydifficulttodo [12].Inourwork,weaddressthechallengebyfram- ingtheevaluationofdehazingmethodsfromadifferentstandpoint.In additiontomeasuringaccuracywithvariousobjectivequalitymetrics, wealsocarryoutasubjectiveimagequalityassessment(IQA)basedon visualpleasingness.Wethenstudythecorrelationofbothobjectiveand subjectiveassessments,providingnotonlyamorecomprehensiveIQA ofdehazedimages,butalsoanalysesoftheirperceptualattributes.Re- gardingthedataset,aimingtoencompassdifferenttypesofhazyscenes, weuseacompilationofvariousdehazingdatasetsandalsogeneratea simulateddataset.Wehopethispaperwillsparkdiscussionandfurther researchon theconceptofqualityinimagerestorationtaskssuchas dehazing.

2. Onimagedehazing

Todehazehazyimagestotheirclearform,itisnecessarytostudy thephysical process thatcaptures a hazyimagefrom a clearscene.

This sectionprovidetheimageformationmodelinscatteringmedia, thestate-of-the-artimagedehazingmethods,andtheIQAapproaches fordehazedimages.

2.1. Hazyimages

Digitalimagesareformedwhenlightraysarecapturedonadigital sensorgridofacamera.Theselightraysusuallycomefromaprimary lightsource,andarereflectedbyasceneintothelensofacamera.The lightraysthatarecapturedatdifferentpixellocationsbuildupanim- agerepresentationofthescene.Thisprocessisquitestraightforwardin clearmedia,whichallowsthelightraystopassthroughunhindered.

However,thisprocessisnotpossibleinscatteringmediaenvironments asmicro-particlescontainedinitmayalterthedirectionofthetravel- ing raysandscattertheminmanydirections.Thiswillalsoresultina reducedlightintensitybecausethemediawillabsorbordiverttheorig- inallight.Themediaactivelyinterfereswithimageformationresulting in abelowstandardrepresentationofthereal scene.Thealteredin- tensitieswillthuscompromisethevisualfeaturesnecessaryforimage understanding[13].

Theimageformationmodelusedinthispaperisbasedonthegen- eralatmosphericscatteringmodel [14],seeillustrationinFig.1.This modelassumesthatthemicro-particlesinthemediaareverysmall,so thescatteringcanbeconsideredhomogeneousandhasrelativelylow density. These assumptionsholdin most naturalhazyconditions. In these naturalconditions,theappearance ofobjectsintheimagewill differbasedontheirdistancefromthecamera.Themicro-particlesin themedia coupledwiththeamountof distancetraveled, directlyaf- fecttheamountoflightthatisabletopenetratethemedia [15].The fractionoflightthatisabletopenetratethescatteringmediadecayex- ponentiallyproportionaltothedistancefromthecamera𝑑,andcanbe definedas:

𝑡=𝑒𝛽⋅𝑑 (1)

where𝛽isthescatteringcoefficientofthemedia,anddisthedistance tothecamera.Inahomogeneousscatteringmedia,thecapturedhazy image𝐼ofanoriginalclearscene𝐽canbedefinedusingthetransmis- sion𝑡asfollows,where𝐴istheairlight:

𝐼=𝐽𝑡+𝐴(1 −𝑡) (2)

The term airlight 𝐴 refers to the color of ambient light in the scene [4]. Thisconsistsoftheaccumulation ofscatteringeffectsand carriesnoinformationabouttheoriginalsceneitself.Airlightisscat- teredin the scene,creating aveiling effectthat subtlyobscures the entireimage[16].Theexactappearanceandhueoftheveilingeffect willdependhighlyonthescatteringmediaintheenvironment.Forim- agescaptured inenvironmentssuchasfogorhaze,theairlightoften presentsapalewhitishcolor.Asafinalresult,theimagescapturedin theseenvironmentswillnotbeabletoportraythesceneperfectly,due toobscureddetails,lossof intensityandlow contrast.Inthispaper, theseimageswillbereferredtoashazyimages.

2.2. Imagedehazing

Imagedehazingisanimagerestorationtask,mappinghazyimages totheirclearcounterpartswiththehighestpossibleaccuracy.However, there aremany cases in which a precise ground truth is not avail- able.Thus,werestatetheimagedehazingprocesstoworktowardsthe achievablegoalofobtainingavisuallypleasingimagewithbetterqual- ity.Theendresultshouldbejustgoodenoughsothatobserversareable tounderstand thescenethroughits visualfeatures.A visuallypleas- ingimagedoesnotnecessarilyneedtobeanexactmatchtoaknown standard,thusitcanbeachievableusingimageenhancementmethods.

Thesemethodshavetheadvantageofnotrequiringaknowncleartar- get,astheiraimistosolelyreducethehazyeffectssuchasdescribed inSection2.1.Thesemethodsattempttoimprovethevisibilityofthe imagethroughmethodssuchascolorcorrection [3,17],contrastcor- rection [2,18],contrastbalancing [9,19],andothers.

Itis alsopossible tomodel thephysical transformationbetween hazyandclearimages.Themodelcan thenbe usedtorestorehazy imagestotheirclearformusingsingleimagedehazingtechniques.Sin- gleimagedehazingconsidersthephysicalscatteringmodeldescribed inSection2.1asthe processinwhich ascenebecomes hazy.Then, singleimagedehazingfollowstheimageformationmodelofhazyim- ages in Eq. (2) and uses its inverse to extract the original scene𝐽 fromthehazeimage𝐼followingEq. (3).Thisequationinvolvesmany unknowns.Thus,singleimagedehazingmethodsmustestimatetheun- knownvariablesinphases,i.e.,estimatingtheambientlightorairlight 𝐴,̂ predictingtransmissionmap̂𝑡,thenrecoveringtheoriginalscene ̂𝐽 (Fig.2).

̂𝐽=𝐼𝐴

𝑡 +𝐴 (3)

Unfortunately,thetransmission𝑡orthedistance𝑑 arerarelyknown.

Furthermore,thescatteringcoefficientofthemedia𝛽canbedifferent foreveryscene,hinderingaone-fits-allestimateforallimages.Tosolve thisissue,itisnecessarytoemployadditionalconstraintstothemodel.

Onepossibleapproachistousestatisticalpriors,i.e.,theDarkChannel

(3)

Fig. 2.General step-by-step process of single image dehazing.

Fig. 3.An example of a clear scene (a) and its depth map (b) from the NYU Depth Dataset [21], and the (c) synthesized hazy image generated based on Eq. (2).

Prior(DCP) [5]orAdaptiveDarkandBrightChannelPrior [20].These priorobservationsof naturalhazyimagesformpre-determinedinfor- mationtohelpestimatingvariablesinvolvedinanimageformation.

Theproblemofimagedehazingcanbemodeled asanimagetoim- agetranslationproblem,e.g.,fromthesetofhazyimagestotheirclear counterparts. There havebeen variousdeep networks that proposed thispurpose,usingvariousstructuresofconvolutions [6,22],attention cues [23]orevenhumanperceptioncues [10].Inrecentyears,these deepnetworkbasedapproacheshavebeenpopularforotherimageto imagetranslationproblemsaswell.Imagetranslationcanbeusefulfor variousapplications,e.g.,semanticsegmentation [24],domainadapta- tion [25],orimagetodepthestimation [26].

2.3. Imagequalityassessment

Imagequalityassessment(IQA)triestomeasurethequalityofdig- ital images in depicting the original scene, andin turn, in relaying information.Itcanbedividedintosubjectiveandobjectiveapproaches.

ObjectiveIQAinvolvesimagequalitymetricstocomputeanddepict imagequalityquantitatively.Thesemetricscanbedividedintofull-,re- duced,orno-referencemetrics.Full-reference(FR)metricsneedaground truthimage,suchthatthedifferenceorerrorfromthetargetimagecan be computed [27].Reduced-reference (RR) metricscan be used when onlyincompletereferencesareavailable,andno-reference(NR)metrics whentherearenoneatall.NRmetricsarealsoreferredtoasblindimage qualitymetrics[28].SubjectiveIQAgivesthetasktohumanobservers, toevaluatethequalityofasetofimagesbasedonacertaincriteriaby, typically,indicatingtheirpreferences.Thehumanvisualsystemiscon- sideredthemostreliablebiologicalimagecapturedeviceabletojudge visualimagequalityinaconsistentmanner [29].

ThemainissuefortheIQAofdehazedimagesisthelackofstan- dardizedevaluationprotocol.Firstofall,hazyimagesthemselvesare difficulttocomeby,andthereisalackofdatasetsthatcanbe used forqualitybenchmarking.Furthermore,itisverydifficulttoobtaina referencesetofhazyimagesandtheirexactclearcounterpart.Natural hazeoccursinoutdoorenvironmentswheremanyfactorsareinvolved inimagecapture.Mostofthesefactors areuncontrollableforexperi- mentaldatacaptureandthesceneisneverreliableenoughtocapture

inboth hazyandclearconditions.The nextbest optionis capturing imagesusingsynthetic haze,butitisstilldifficulttoobtainthe ex- actsameconditiontocapturethesceneinbothconditions.Thiswork wouldalsohighlighttheneedtoconsiderIQAfordehazedimages.Nat- uralatmosphericscattering [14]occursoutdoors evenintheclearest ofconditions,thushumanobserversexpectsomehazeinanynatural image.Meanwhile,objectivemetricswouldseeanyremaininghazeas anindicationoflowquality.Thus,itisunrealistictorelyonaccuracy alone,sinceitisimportanttoconsidervariousfactorsofimagequality dependingonthecontextoraimofthetask.

3.Materialsandmethods

This section will describe the image dehazing experiments con- ductedforatestsetof50images,followedbyathoroughexplanation oftheusedIQAmethods.

3.1. Dataset

Inourexperiments,acombineddatasetofhazyimageswithground truthclearimagesfrom3differentsourceswereused.Theentiredataset iscomprisedof600hazy-clearimagepairs,whichthendividedintoa trainingsetof550imagestotrainthePDR-Net(Section3.2.3),anda testsetof50images.OnlythetestsetwasusedtoperformIQAonthe dehazingresults.

3.1.1. Synthetichazyimages

TheNewYorkUniversity(NYU)depthdatasetiscomprisedof1449 pairsofRGB-Depth(RGBD)data,i.e.,indoorscenesandtheircorre- spondingdepthmaps [21],seeanexampleinFig.3aand3b.Theyare capturedinvariouscommercialandresidentialbuildings,withdepth mapsobtainedusingtheMicrosoftKinect.

Forourexperiment,wegeneratedsynthetichazyimagesfrom210 randomlyselectedRGBDpairsfromtheNYUdatasetfollowingEq. (2), seeexampleinFig.3c.First,thepixel-wisedepthmap𝑑wereusedto createadensetransmissionmap𝑡usingseveralscatteringcoefficient values𝛽∈ {0.1,0.2,0.3,0.4}.Then,𝑡wasusedtosynthesizehazyimages 𝐼usingdifferentairlightvalues𝐴∈ [0.7,1][26,30].

(4)

3.1.2. Hazyseriesdataset

FourdatasetsareavailablefromTheNewTrendsInImageRestora- tion And Enhancement Workshop And Challenges (NTIRE) chal- lenge [31].TheIndoorHaze(I-Haze)dataset [32]consistsofhazy-clear imagepairscapturedinindoorconditions,whiletheOutdoorHaze(O- Haze)dataset [33]wascapturedinoutdoorconditions.Inbothdatasets, hazewasgeneratedusingtwoprofessionalfogmachines(LSM1500PRI) which was dispersed using fanstoensure homogeneity. The Dense- Haze [34]andNon-Homogeneous Haze(NH-Haze) [35]datasetsare bothextensionsoftheO-Hazedataset.Thehazefromthefogmachines was releasedforalongerdurationtocreateadenserhazeeffect for theDense-Hazedataset.ForNH-Haze,thegeneratedhazewasnotdis- persed evenlythrough the scene ensuringa non-homogeneous haze condition.Atotalsetof180hazyimagesandtheirclearcounterparts wascollectedfromallfourofthesedatasets.

3.1.3. REalisticsingle-imageDEhazingdataset

Li,etal. [12]createdtheREalisticSingle-ImageDEhazing(RESIDE) datasetwiththeintentiontoprovideastandardizedlarge-scaledataset of hazy-clearimagepairstofacilitate thebenchmarkingofdehazing methods.RESIDEconsistsofrealisticallysimulatedhazyimagesandis organizedinto3subsets,i.e.,theIndoorTrainingSet(ITS),Synthetic Outdoor TrainingSet (SOTS),andtheHybridSubjective Testing Set (HSTS).Amongallthreesubsets,SOTSisthemostsuitableforourneeds asithasmanyimagesof urbanoutdoorhazyscenes.Forourexperi- ment,210outdoorimagepairsfromtheSOTSsubsetwererandomly selected.

3.2. Dehazingmethods

Fourmethodsareusedinourexperiments.Theyareanimageen- hancementapproach,i.e.,CLAHE [9],asingleimagedehazingmodel usingdarkchannelprior(DCP) [5],andtwodeeplearningbasedde- hazingmodels.Thedeeplearningbasedmodelsselectedencompassa generativemodelPDR-Net [10]andanadversarialmodelPix2pix [11].

3.2.1. Contrastlimitedadaptivehistogramequalization(CLAHE) Oneoftheeffectsthatoccurinhazyimagesislowcontrastandloss ofdetail.Thisiswhyhazyimagesaresodifficulttoperceive,because contrast carries significant information about image structure [36].

Thus,toimprovevisibility,wecanattempttoimproveorcorrectthe contrast.Contrastcorrectioncanbedonebybalancingthedistribution ofintensityvaluesinanimage.Thehistogramofalowcontrastimage willbeconcentratedinanarrowrangeofvalues,makingdifferentiating intensityvaluesdifficult.

Wecanimproveimagecontrastbymappingtheintensityvaluesto awiderrangethroughhistogramequalization.Theincreaseofdiffer- encebetweenindividualintensitieswillamplifyedgesandimagestruc- tures,herebyimprovingvisibility.Therehavebeenvariousproposals towardshistogramequalization,andinthispaper,weselectthewell- knownandeasilyimplementedCLAHE [9].CLAHEseparatesanimage intodifferentcontextualregions,andperformsanadaptivehistogram equalizationoneach regionseparately. Additionally,italsoclipsthe intensitydistributionatacertainlimittopreventover-saturationinho- mogeneous areas.Our experiments involvecolorimages, sowe first transformedtheRGBcolorimagetotheCIELABcolorspace[37]and CLAHEwasperformedonthebrightnesschannel𝐿[9].

3.2.2. Darkchannelpriorandrefinement(DCP-R)

The imageformation model in Eq. (2) shows thetransformation fromaclearimage𝐽toahazyimage𝐼,whichwewouldliketoreverse.

However,thisequation involvesmany unknownvariables,makingit difficulttosolve.Additionalconstraintsareneededtoallowreversing theequationtoestimate𝐼 from𝐽. Inthis paper,weselecttheDark ChannelPrior(DCP) [5]whichiswidelyacknowledgedandusedasa basisofmanydehazingworks.

DCPisastatisticalpriorbasedontheobservationofnaturaloutdoor images.Itstipulatesthatforlocalpatchesofnon-skyareasinnatural images,thereisalwaysonecolorchannelthathasaverylowvalueor min(𝐼𝑠)≈ 0,𝐼𝑠∈ [𝑅,𝐺,𝐵].Thesedarkpixelsoccurduetonaturaldark objects,lighting,anddistinctlycoloredobjects.Insuchcases,thereis oftenonechannelwithlowintensity,whileotherchannelscontainthe colorinformation,thuscreatingtheaforementioneddarkchannel.

Darkchanneltypicallyoccursinnaturalclearimages.Inhazyim- ages,adarkchannelbecomessaturatedwiththeeffectofhaze,which canbedenotedasaproportionofairlightintensity𝐴fromtheimage formationmodelinEq. (2).Basedon theconceptoftransmission in Eq. (1),itisknownthatairlightwillaggregate in thedarkchannel, withhigherquantityindicatingfurtherdistance [5,38].Consequently, thedistancevaluewillalsogrowproportionallywiththevalueofDCP, allowingittobeusedasanindicatorofrelativedepth.He,etal. [5]

proposedamethodtoestimatetransmissionbasedonDCPasdescribed in Eq. (4). Assuming anRGB hazyimage 𝐼 in theRGB colorspace (𝑠∈ [𝑅,𝐺,𝐵]),thetransmission ̂𝑡can be estimated foreverypixel 𝑥 inthehazyimage𝐼 consideringeverypixel𝑦inthelocalareaΩ(𝑥) surrounding𝑥.

̂𝑡𝐷𝐶𝑃(𝑥) = 1 − min

𝑠∈{𝑅,𝐺,𝐵}

(

𝑦∈Ω(𝑥)min (𝐼𝑠(𝑦)

𝐴𝑠

)) (4)

Note that the value of airlight 𝐴 in the RGB color space (𝑠∈ [𝑅,𝐺,𝐵])orits estimationisneededtocompute ̂𝑡.Therearevarious approachesthatcanbeusedtoestimate𝐴,thesimplestofwhichisby usingtheDCPvalueitself.He,etal. [5]setthevalueofairlightwith thepixelvalueatthelocationofthemaximumDCP.Airlight,orambi- entlight,isoftenequatedwiththecolorofthesky,whichcanbefound atamaximumdistance𝑑≈ ∞.SincealargerDCPvaluewouldindicate afurtherdistance,themaximumvalueofDCPshouldindicatethefur- thestpointaswell,atwhichpointcanbeusedasanestimateof ̂𝐴.With theestimatedtransmission̂𝑡basedonEq. (4),theoriginalscene𝐽 can berecoveredforeverypixel𝑥basedontheprinciplesofDCPaccording toEq. (5).Toavoidadivisionbyzero,itisnecessarytolimitthevalues of̂𝑡byaminimumvalueof𝑡0.

̂𝐽𝑠(𝑥) = 𝐼𝑠(𝑥) −𝐴𝑠 max(̂𝑡(𝑥), 𝑡0

)+𝐴𝑠 (5)

Inourexperiments,imagedehazingusingDCPwasperformedbya two-stepapproach.Thefirststepresultsinanapproximatedclearim- ageofthescene.However,sinceDCPassumesconstanttransmission inlocalareas,theseimagessufferfromthelossofdetailsorappearing tobepatchy.Thus,arefinementstepwasaddedusingtheguidedfil- ter [39],aimingtofurtherimproveandsmooththedehazedimages.

Withanedge-preservingsmoothingproperty,theguidedfilterusesa referenceimagetoguidetherefinementprocess,resultinginalinear transformoftheguidanceimage.Finally,thecompletesingleimagede- hazingapproachinthispaperisacombinationofDCPandrefinement usingguidedfilter,furtherreferredtoasDCP-R,anditwasusedtode- hazethe50hazyimagesfromthetestset.

3.2.3. Perception-inspireddehazingnetworkwithrefinement(PDR-Net) Li,etal. [10]proposedPDR-Net,aperception-baseddeeplearning architecturetomodelend-to-endimagetoimagetranslationbetween hazyimagesandtheir clearcounterparts.Similar toDCP-R,ittreats theprocessofdehazingastwosteps,i.e.,hazeremovalandrefinement, eachimplementedastwoseparatesub-networks.PDR-Net employsa perceptuallossin thehazeremovalsub-network,todescribehigher- levelperceptualcues.Meanwhile,therefinementsub-networkemploys amulti-term losstorecover colordistortionand enhancethevisual qualityofthedehazedimage.Throughthesub-networks,PDR-Netde- hazesimagesintwosteps,resultinginnotonlyanaccuratelydehazed image,butalsoavisuallypleasingone.

Inourexperiments,PDR-Netwastrainedusingthetrainingsetof 550hazy-clearimagepairs.Sincedeeparchitecturesoftenbenefitfrom

(5)

moresamples,thetrainingsetwasaugmentedbyflippingeachimage horizontally, creatingatotal of1100 trainingimages.Followingthe proceduredetailed in [10],each sub-networkwastrainedseparately, with700trainingimagestotrainthehazeremovalsub-network,and 400fortherefinementsub-network.Eachsub-networkwastrainedfor 200epochs.AswithourDCP-R,PDR-Netwasalsousedtodehaze50 hazyimagesofthetestset.

3.2.4. Pix2pix

Commongenerative networks mayexperience difficulty in train- ingimagetoimagetranslationsduetothelarge numberofvariables andprobabilisticcomputationstoestimate.Thus,theGenerativeAd- versarialNetworks(GAN) [40]frameworkmaybeusedtoassistwith trainingthesegenerativenetworks.AGANsubjectsagenerativemodel toadiscriminativemodel,trainedtoworkasopponents.Thegenera- toristrainedtogeneraterealisticrenderingsoftheintendedoutputs, whilethediscriminatoraimstodetectthese reconstructionsfromthe groundtruth output [40]. AConditionalGenerativeAdversarialNet- work(cGAN)extendstheconventionalGANframework,i.e.thegener- atoranddiscriminatorareconditionedwithextrainformationtoguide thegenerativeprocess [41].

Pix2pixisacGANmodelproposedforgeneralnon-specificimageto imagetranslationtask [11].Pix2pixusesamodifiedU-Net [24]asits generativemodel,guidedbyaPatchGANdiscriminator [42].Pix2pix was thoroughly examined using various image to image translation tasks, suchasgrayscaletocolorimagesordaytonightimages.The resultsshowitsabilitytoreproduceavisuallysimilaroutputtothetar- getforseeminglyarbitraryimagepairs.Consideringadehazingtask, weareoptimisticthatthediscriminatorcomponentcanimprovethevi- sualpleasingnessoftheresultingimage.Sincethediscriminatorjudges thegeneratedoutputbasedonoverallsimilarity,notonlyonamini- mizederror,webelievethismaymimictheobservationorjudgement ofahumanobserver.Inourexperiments,wealsousedPix2pixtotrain image toimage translationfromhazyimagestotheir clearcounter- parts.FollowingthetrainingandtestingsetupofPDR-Net,Pix2pixwas alsotrainedfor200epochsusingthesameaugmentedhazy-clearim- agepairs,with1100trainingimages.Thefinaltrainedmodelwasthen usedtodehaze50hazyimagesofthetestset.

3.3. IQAofdehazedimages

Thedehazingresultswerethenassessedbymeansofobjectiveand subjectiveIQAs.Theimagequalitymetricsandthedesignofthepsy- chovisualexperimentforthesubjectiveIQAaredetailedinthissection.

3.3.1. Objectivemetrics

Inthis study,fourIQA metricswereselectedtoevaluatevarious quality aspectsofan image.Three metricsarefull-referencemetrics thatassignaquantitativevaluetotheaccuracyofdehazingcompared tothegroundtruth,basedonthequalitycuesfocusedoninthispaper, namelycolor,imagestructure,andpixelintensity.

First, we evaluate color using Δ𝐸𝑎𝑏, a color difference function defined for the CIELAB color space [37], where perceptual non- uniformitiesareaccountedfor.Ahazyimage𝐼anditsdehazedversion 𝐽 willbeconvertedtotheCIELABcolorspacebeforethecomputation oftheircolordifferenceasinEq. (6).ImagesintheCIELABcolorspace havethreecomponentsorchannels,i.e.,𝐿(lightness),𝑎(green-red chromaticity),and 𝑏 (blue-yellow chromaticity).Therangeof color differencevaluesfromΔ𝐸𝑎𝑏is[0,100],whichwerescaleto[0,1],with alowervalueindicatingbetterquality.

Δ𝐸𝑎𝑏(𝐼, 𝐽) =

(𝐿𝐽𝐿𝐼)2+ (𝑎𝐽𝑎𝐼)2+ (𝑏𝐽𝑏𝐼)2 (6) Toevaluatetheaccuracyofpixelwiseimageintensities,weusethe rootmeansquareerror(RMSE),showninEq. (7).RMSEcomputesthe differencesinimageintensityorbrightnesslevels[43].Assumingpixel

valuesintherangeof[0,1],theRMSEoftheentireimageisaveraged overthetotalnumberofpixelsintheimage(𝑛).Thus,theRMSEwill alsorangebetween0and1,withlowervalueindicatingabetteresti- mationoftheclearimage.

RMSE(𝐼, 𝐽) =

√√

√√1 𝑛

𝑛 𝑖=1

(𝐼𝑖𝐽𝑖)2 (7)

Theaccuracyofthepixelintensitiesthemselves,doesnotensurea cleardistinctstructureinanimage.Thus,wealsousedthestructural similarityindexmeasure(SSIM)[44]toevaluatetheimagestructure recovered.SSIMisusedtodefinetheperceptualdifferenceoftwoim- agesbymeansoftheirluminance(𝑙),contrast(𝑐),andstructural(𝑠) components.TheformulaofSSIMisprovidedinEq. (8),with𝜇, 𝜎𝐼2, and𝜎𝐼𝐽 asthemeasureofaverage,variance,andcovariancebetween 𝐼and𝐽,respectively.Otherparametersintheformulaare𝑐1= (𝑘1𝐿)2, 𝑐2= (𝑘2𝐿)2,𝑐3=𝑐2∕2,𝐿= 2𝑏− 1,𝑘1= 0.01,𝑘2= 0.03,(𝛼,𝛽,𝛾)asadapt- ableweights,and𝑏asthenumberofbitsperpixelinanimagewhich willtypicallybe8.

SSIM(𝐼, 𝐽) = [𝑙(𝐼, 𝐽)𝛼𝑐(𝐼, 𝐽)𝛽𝑠(𝐼, 𝐽)𝛾],where 𝑙(𝐼, 𝐽) = 2𝜇𝐼𝜇𝐽+𝑐1

𝜇2𝐼+𝜇2𝐽+𝑐1

, 𝑐(𝐼, 𝐽) = 2𝜎𝐼𝜎𝐽+𝑐2

𝜎𝐼2+𝜎𝐽2+𝑐2

, 𝑠(𝐼, 𝐽) = 𝜎𝐼𝐽+𝑐3

𝜎𝑖𝜎𝐽+𝑐3

. (8)

Therangeof values that theSSIMcan take on is [−1,1], witha largervalueindicatingbetterquality.Meanwhile,withalloftheother metricsusedinthisstudy,betterqualityisindicatedwithalowervalue.

Furthermore,SSIMisalsotheonlymetricthatcanberepresentedwith anegativenumber.Thus,theSSIMisconvertedintoaninvertedSSIM whichwillbedenoted𝑆𝑆𝐼𝑀−1withamodifiedrangeof[0,1],witha lowervalueindicatingbetterqualitybasedonEq. (9).

SSIM−1(𝐼, 𝐽) = 1 −

(SSIM(𝐼, 𝐽) + 1 2

)

(9) BelongingtothecategoryofFRmetrics,Δ𝐸𝑎𝑏,RMSE,andSSIM−1 allrequiretwoimagesasinput,i.e.,groundtruthorreferenceimage andthetargetof evaluation.Notethattheimagedimensionofboth imageshavetobeexactlythesame.Thesethreemetricsmeasurethe accuracyofthedehazedimagecomparedtotheknowngroundtruth.

Wehavealreadysuggestedthatoften,accurateimagesdonotnec- essarilyhavebettervisualquality.Frequently,humanperceptionwill considerabstractconceptssuchasnaturalness. Thus,wealsouseda metricthatwasproposedtomeasurenaturalness,TheNaturalnessIm- ageQualityEvaluator(NIQE)[45].NIQEisanNRmetricthatcomputes thescoreofanimage againstamodelof naturalsceneimages.This modelisrepresentedasmultidimensionalGaussiandistributions.The rangeofvaluesthatNIQEcantakeonis[0,∞),withalowervaluein- dicatingbetterimagequality.

3.3.2. Psychovisualexperiment

BasedontheobjectiveevaluationresultinSection3.3.1,wecould naivelyinferthatadehazedimagewithagoodobjectiveresultindicates thatitisvisuallypleasingtothehumaneye.Thesameprincipleshould intuitivelyapplytotheoppositecase.However,subjectiveandobjective measuresdonotalwaysalign [44].Thus,subjectiveIQAbymeansof apsychovisualexperimentwasalsocarriedoutforthetestsetof50 images.

Dehazingresultsofthefourdehazingmethodsweresentoutina surveyinwhichobserverswereaskedtoidentifytheirpreferredresult.

ThedisplayofthepsychovisualexperimentcanbeseeninFig.4.Each imageisshownonanindividualslide,withaneutralgraybackground andtheslide numberwritten atthetop-left of eachslide indicating whichimageiscurrentlydisplayed.Oneachslide,fiveimagesaredis- played,i.e.theoriginalhazyimageinthefirstrowandfourdehazed imagesonthesecondrow,eachobtainedusingCLAHE,DCP-R,Pix2pix, andPDR-Net,respectively.Observerswerethenaskedtorankthefour dehazedimagesfromtheimagethatis mostvisuallypleasingtothe

(6)

Fig. 4.Image arrangement for the psychovisual experiment.

least.Consideringobserverfatigue,eachobserverwasonlyshown25 randomlyselected imagesfromtheentiretestsetof 50images. Ob- serverswerefreetouseanycriteriatheydeemimportanttodetermine theirpreference.

Responses from observers were then populated using the Mean OpinionScore(MOS) [46].TheMOSofimage𝐼isdeterminedbythe scores𝑆assignedtoitby𝑁observers,whichcanbecomputedbased onEq. (10).

MOS(𝐼) = 1 𝑁

𝑁 𝑛=1

𝑆𝑛 (10)

Inourexperiment,thescoresweregivenbasedontherankordergiven byhumanobservers.Foreachresponse,ascoreof1wouldbeassigned tothemostpreferredresult,2tothenext,3tothethird,and4tothe leastpreferredresultforthatparticularimage𝐼𝑖.Thus,alowerscore willindicatebetterperceivedquality,whichisconsistentwiththeob- jective metricsin Section3.3.1. Thescore for each image wasthen averagedoverthenumberofitsobservers𝑁.Theaveragescorewas thenfurtherconvertedtoamodifiedrangeof[0,1].Thiswascomputed forall50testsetimages𝐼𝑖,for𝑖∈ [1,50].Attheendoftheexperiment, observerswerealsoaskedtogiveopen-endedcommentstoelaborate furtherabouttheirobservationsaboutthedehazedimages.Thesecom- mentsmaygive anindication aboutthefactorsthatinfluenced their preferenceintermsofcolor,structure,clarity,ornaturalness.

4. Resultsanddiscussion

Thissectionwilldisplay,evaluate,andanalyzetheresultsusingthe fourdehazingmethodsinourexperimentsonthetestsetof50images describedinSection3.2.Arandomlyselectedsubsetoftheresultsare showninFig.5.Fromtheresults,itisvisiblethatallmethodsperformed wellinlightandsimulatedhazesuchasintheO-Haze,Synthetic,and RESIDEimages.However,itismorechallengingtohandleimagesfrom theDense-HazeandNH-Hazedatasets.

Inthecaseoftestimageswithdensehaze,noneofthefourdehaz- ingmethodswereabletorecoverthescenecorrectly,asshowninthe Dense-HazecolumninFig.5.Fig.6showsmoredetailsoftherecovered Denze-Hazeimages,whicharestillshroudedinaveilinghaze.CLAHE wastheleastsuccessful,duetotheover-saturatedhazethatdominates theimageandhenceitshistogram.DCP-Rwasmoresuccessfulinrecov- eringdetailsofthesceneandtheresultsshowsmoresaturatedcolors.

However,thesecolorsarefarfromcolorsinthegroundtruthimage.

PDR-Netproducesadarkerimagewithlesssaturatedcolors.Italsoap- pearstohavealotoftexturesimilar tothatof thetreeobjectinthe groundtruthimage.However,notmuchinformationwassuccessfully restoredasidefromthewhitestructuresattoprightcorneroftheim- age.ThedehazedimageusingPix2pixwasabletoreconstructthearea

Table 1.AverageobjectiveandsubjectiveIQAmetricresults onthetestsetusing4dehazingmethods.Allmetricshavebeen convertedandscaledsothatthelowervaluewillindicatebetter quality.

Objective IQA

Metric Range↓ CLAHE DCP-R PDR-Net Pix2pix RMSE [0,1] 0.4221 0.3361 0.3729 0.2367 𝑆𝑆𝐼𝑀−1 [0,1] 0.2206 0.2425 0.2292 0.1919 Δ𝐸𝑎𝑏 [0,1] 0.2391 0.2291 0.2194 0.1514 NIQE [0,∞) 3.1521 2.8290 4.7216 2.8626 Subjective IQA

Metric Range↓ CLAHE DCP-R PDR-Net Pix2pix MOS [0,1] 0.4461 0.4877 0.8447 0.2220

oftheleaves,butalsoproducedcolorartefactsinthehazeareasofthe image.

Section2.1showsthatthebasichazyimageformationmodelitself isbuiltontheassumptionthatscatteringoccurshomogeneously.Thus, imageswithnon-homogeneoushazeposeachallengetocommonde- hazingmethods.SeeNH-HazecolumninFig.5,whereCLAHE,DCP-R, andPix2pixwereallunabletofullyremovethehaze,andthedehazed imagesstillcontainhazeinareaswherethehazeisthicker.Meanwhile, theresultofPDR-Netwassuccessfullyvoidofhazeeffects,albeitwith lowcolorsaturation.DetailedobservationsoftheresultsonNH-Haze canbeseeninFig.7.

Objectiveevaluationofthedehazingresultsforthefourdehazing methodsispresentedinTable1.Notethatallobjectivemetricshave beenconvertedandscaledsothatalowervalueindicatesabetterre- construction.Overall,Pix2pixisabletoobtainthebestscoreinRMSE, 𝑆𝑆𝐼𝑀−1,andΔ𝐸𝑎𝑏.ThebestNIQEscoreisachievedbyDCP-R,show- ingitssuperiorityintermsofnaturalness.Furthermore,allothermeth- odscomeinsecondindifferentaspects,whereCLAHEhasthesecond best𝑆𝑆𝐼𝑀−1score,DCP-RinRMSE,andPDR-NetinΔ𝐸𝑎𝑏.CLAHEis amethodthatmodifiestheimageveryminimally,explainingthegood 𝑆𝑆𝐼𝑀−1score.DCP-RhasagoodRMSE,indicatingasuperiorpixel intensityrecovery.However,DCP-Rdoessufferfromcolordistortion, explainingalowerΔ𝐸𝑎𝑏 scorecomparedtoPDR-Net.

Thepsychovisualexperimentofthe50testimageswasconducted withatotalof40humanobservers.TheaverageMOSforeachmethod is shown in thebottom rowof Table 1, indicating a preferencefor dehazingbyPix2pixreflectedin itsloweraverageMOS.CLAHEand DCP-Rcomeinsecondandthird,withPDR-Netastheleastpreferred dehazingmethod.Thepsychovisualexperimentalsoprovidedanoppor- tunityforobserverstoaddopen-endedcomments,whichmayfurther provideinsightsonthepossibleperceptualattributesfortheperception ofdehazedimages.Theywillbevaluableforadeeperanalysisofvisual cuesaffectingtheperceivedimagequalityandpotentialimprovements

(7)

Fig. 5.Examplesofdehazingresultsoftestimagesfromvariousdatasets,i.e.,Dense-Haze,NH-Haze,O-Haze,Synthetic,andRESIDEdatasets.Theoriginalhazy imageisshowninthefirstrow,followedbythegroundtruthclearimageinthesecond.ThedehazedimagesobtainedusingCLAHE,DCP-R,PDR-Net,andPix2pix areshowninrow3-6.

Fig. 6.Detailedcomparisonofaresultof(a)CLAHE,(b)DCP-R,(c)PDR-Net,and(d)Pix2pixforanimagewithdensehaze.Seethecorrespondingoriginaland groundtruthimagesinDense-HazecolumninFig.5.

(8)

Fig. 7.Detailed comparison of a result of(a)CLAHE,(b)DCP-R,(c)PDR-Net, and(d)Pix2pix for an image with non-homogeneous haze.

Table 2.Insightsfromtheopen-endedcommentsprovidedbyobserversinthe psychovisualexperiment.

No. Comment

1 ImagesdehazedwithCLAHEstillcontainaveryvisiblehaze,indicatinga lesssuccessfuldehazingprocess.

2 ImagesdehazedwithCLAHEarestillmorevisuallypleasingregardlessof thehaze,becausetheydonotcontainnoiseorartifacts.

3 TheleftoverhazeinCLAHEimagesactuallycontributestothenaturalness oftheimage.

4 ImagesdehazedwithDCP-Raregenerallybright,smoothandsharp.DCP-R seemstoperformverywellwhenthehazeintheoriginalimageislight.

5 ImagesdehazedwithDCP-Roftenhavecolordistortionsmakingitless visuallypleasing.Thisisespeciallyvisibleinoutdoorsceneswithhave unnaturalskycolor.

6 ImagesdehazedwithDCP-Rsometimesappeardullanddarkwithsome scenes.

7 SeveralimagesdehazedwithPDR-Nethavespots,artifacts,ornoise.

Althoughsomeproducecolorsthataresimilartorealobjectsinthescene buttheartifactsmakesPDR-Nettheleastvisuallypleasing.

8 PDR-Netresultsseemtohavelostthesenseofdepthofthescene,duetoa veryhighcontrast.

9 PDR-NetandPix2pixbothareabletoeliminatemorehaze,butarevisibly unnatural,makingobserversuncomfortable.

10 ImagesdehazedwithPix2pixtendtohavesoftercolorswhichmakeitmore visuallypleasing.

11 ImagesdehazedwithPix2pixaremostoftenmorevisuallypleasing comparedtotheothermethods,asidefromsomeunnaturalcolors.

12 Clearobjectsandminimalartifacts/distortionarealargefactorin evaluatingvisuallypleasingimages.ThatiswhyCLAHEandPix2piximages werefrequentlyplacedatthetoprank.

fordehazingmethods.Someinterestingpointsabouttheobservations thatcanbeconcludedfromthesecommentsareshowninTable2.

Asdescribedbefore,theconceptofimagequalityisverycomplex.

Thequantitativemetricsthatarecommonlyusedtodescribeaccuracy ofdehazedimages,arenotalwaysadequatetodenoteitsquality.This isapparentintheobjectiveandsubjectiveresultspresentedinthissec- tion.Fig.8showsthescatterplotofMOSagainstall4objectivemetrics, i.e.RMSE,𝑆𝑆𝐼𝑀−1𝐸𝑎𝑏,andNIQE.Forvisualizationpurposes,each objective metricis scaledtotherange of [0,1]. We thenattemptto infer thecorrelationbetween MOS andeach objective metric, using curvefittingwithapolynomialregressionoforder1.Theplotshows weakcorrelationofMOSwithΔ𝐸𝑎𝑏and𝑆𝑆𝐼𝑀−1,whoselineplotsare almosthorizontal,indicatingnorelation.Meanwhile,theMOSscores

Fig. 8.ScatterplotbetweenthesubjectiveMOSandRMSEinred,𝑆𝑆𝐼𝑀−1in blue,Δ𝐸𝑎𝑏 inyellow,andNIQEingreen.Thecorrelationlineforeachmetricis obtainedthroughcurvefittingwithpolynomialregressionorder1.

withRMSEandNIQEshow astrongercorrelationthroughlinesthat inclinevisibly.

Tomeasurethe correlationquantitatively, wealso computed the Pearson𝑟correlationcoefficientbetweenMOSandeachobjectivemet- ric [47]. The Pearson𝑟 correlationcoefficient is commonly used to quantitativelyrepresent therelationshipbetweenvariables [47].The Pearson𝑟coefficientofMOSwithRMSE,𝑆𝑆𝐼𝑀−1𝐸𝑎𝑏,andNIQEare showninTable3.Fromtheresults,weareabletoconfirmtheinsights weobtainedfromFig.8.All4objectivemetricsarepositivelycorre- latedwithMOS,todifferentdegreesofstrength.Amongthem,NIQEis theobjectivemetricthathasthestrongestcorrelationwithsubjective MOSwithaPearson𝑟coefficientof0.4486.

4.1. Coloranalysis

Thecolorqualityofthedehazedimagescanbeobjectivelymeasured usingthecolordifferencemetricΔ𝐸𝑎𝑏.FromTable1,Pix2pixisclearly

(9)

Fig. 9.Detailedareasofthedehazedimagesusing(a)CLAHE,(b)DCP-R,(c)PDR-Net,and(d)Pix2pixintermsofcolorandcolorrecovery,contrast,andsmoothness.

Table 3. Pearson 𝑟 correlation coefficient between MOSand4objectivemetrics.

Metric paired with MOS

RMSE 𝑆𝑆𝐼𝑀−1 Δ𝐸𝑎𝑏 NIQE Pearson𝑟 0.2284 0.0914 0.1856 0.4486

aheadwithanaverageΔ𝐸𝑎𝑏of0.1514followedbyPDR-Net,DCP-R,and CLAHE,inthatorder.Thiscanalsobeconfirmedbyvisuallyobserving thedehazingresultsshowninFig.9.Thetwodeeplearningmethodsare intheleadinthiscriterion,astheybothlearncomplexmappingsfor pixel-basedregressionsfrominputtooutputimageswithlittleexternal factors.

CLAHEenhancestheimagebyre-distributingtheintensitieswith- out correctingthe colorwhich is apparent Fig. 9a andFig. 10a. In the two first images of Fig. 9d, where Macbeth ColorCheckers are present,Pix2pixshowsitscapabilitytorecoverthecolorsonthecolor checkers.DCP-R andPDR-Netalsodowellrestoring thevividnessof thecolor in the color checkers in the first two images. Forthe se- lectedareasin imagesin columns 3-5of Fig.9b,DCP-R showsthat its recoveryof colors results in higher saturation while maintaining smoothness.However, DCP-R ishighly relianton the airlightof the scene,whichisfrequentlyinaccuratelyestimated.Thiserrormanifests intheunnaturalskycolors,whichwehighlightinthefirsttwoimages ofFig.10bForthesimulateddatasetinvolvingindoorimagesincol- umn3,thesecolordistortionsresultincolorartifactssuchasin the

(10)

Fig. 10.Detailed areas of the dehazed images using(a)CLAHE,(b)DCP-R,(c)PDR-Net, and(d)Pix2pix, specifically the artifacts and color distortion.

rightmostimageofFig.10b.Whilethiscolordistortiondoesnotoc- curintheresultofPDR-Net,artifactsareintroduced,suchasshownin Fig.10c.

Finally,thepsychovisualexperimentresultsalsoshowtheobservers’

generalpreferencetowardsthedehazedimagesobtained byPix2pix.

Basedon theMOS in Table 1, theorder of preferenceafterPix2pix isCLAHE,thenDCP-R,andlastlyPDR-Net.Mostobserverstakenote of thecolor recovery asan major factor todetermine visual pleas- ingness.DCP-Respeciallywassingledoutfrequentlyaboutitscolors, due to thecolor distortions that occur. Manyobservers commented particularlyon the sky areas,that thecolorswere “uncomfortable”,

“unnerving”,and“unnatural”.BasedonMOS,wecanseethatCLAHE isthusmorepreferred,althoughitclearlydoesnotrecovercolorcor- rectly.

4.2. Contrastandclarity

Thecontrast andclarityof dehazedimagescan be inferredfrom RMSEor𝑆𝑆𝐼𝑀−1inTable1.Basedonthosetwometrics,onceagain Pix2pixis superiortotheothermetrics.Interms ofimagestructure, CLAHEisabletokeepitintactsinceitonlymodifiesimageintensities withoutchangingthestructure atall.Thisis shownbytheCLAHE’s

(11)

𝑆𝑆𝐼𝑀−1scorethatissecondbesttoPix2pix.Lastly,wehavePDR-Net followedbyDCP-R.Therecoveryofstructureanddetailinanimageis difficult,particularlyifpixelintensitiesaresaturatedwithhaze.

However,indensehazeimages,suchasinthefirstcolumnofFig.9, theresultsofDCP-Rshow abilitytorecovercontrastanddetail,par- tiallyduetothehelpoftheguidedfilterintherefinementstep,which guarantees asmoothimage.Inthese conditions,CLAHEisonly able toenhancetheedgesthatwerealreadythere,butisincapableofhan- dlingover-saturatedhazeareas.Meanwhile,dehazingbyPDR-Netand Pix2pixisdeeplearningbased,ensuringahighlyaccuratereconstruc- tionasprojectedbyalowerRMSE.Despitebeingthebest-performing methodintermsofRMSEor𝑆𝑆𝐼𝑀−1,Pix2pixalsostrugglestorecover imagesindensehaze.Additionally,sincedeeplearningmethodsusea globalaccuracyovertheentireimage,localartifactsarestillpresent, e.g.,thefirstandthirdimagesofFig.10.

The resultsof the subjectiveevaluationshow thatobservers also gravitatetowardsPix2pixdehazing.Manyobserverstakenoteofedges, visibility,anddistinctionofobjectswhenevaluatingthedehazedim- ages. CLAHEisoften notedforbeingunabletohandlevarioushazy images, especiallyindense haze.Observershave takennotice about theabilityofDCP-RandPix2pixtorecoverobjectsclearlywithdistinct edges.Next,imagesdehazedwithCLAHEoftenstillcontainhazeob- scuringtheobjectsandedgesofthescene.Althoughedgesandcontrast is importantin ascene,some observersmentioned thatthecontrast andedgesinimagesdehazedbyPDR-Netwereextreme.PDR-Netcre- atedunnecessaryedgesandartifactstahtwerepointedoutexplicitlyas reasonstorejectthedehazingresultsofPDR-Net.

4.3. Naturalness

Naturalnessisadifficultconcepttoquantifybecauseitisalargely subjectiveconcept.This involvesmany immeasurablefactorssuchas preference,perception,andfamiliarity.Thispaperattemptstomeasure naturalnessquantitativelyusingNIQE.BasedontheresultsinTable1, DCP-RobtainedthebestNIQEof2.8290,followedbyPix2pix,CLAHE, andlastlyPDR-Net.Amongthetoptwo,DCP-RandPix2pix,thesub- jectiveevaluationshowsapreferencetowardsPix2pixdehazedimages.

CLAHE,ontheotherhand,mayobtainnaturalresultsbutis unableto completethetaskofdehazing.Finally,PDR-Netperformstheworstin reconstructinganaturaldehazedimage.Avisualobservationofsome oftheresultscanbedonethroughFig.10.

PDR-Netdehazedimagesaregeneratedbyadeep-learningnetwork throughconvolutions,allowingahigh-levelabstractionthatdoesnot necessarily matchwith howthe humanvisual system or perception works.Thisprocessiscomputedbasedonaccuracyanderror,bothof which areobjectivemeasuresthatdonot reflectoraccount forcues thatareimportantforhumanperception.PDR-Netwasdesignedtouse amodifiederrorthatconsidersbothaccuracyandvisualperception,but theresultshereshowthatitstillhasthislimitation.Meanwhile,while Pix2pixisalsoadeepgenerativenetwork,itistrainedusinganadver- sarialdiscriminator,notjustbyasimpleerrororlosscomputation.The discriminatorcomponentofPix2pixactsasifitisperceivingtheimage, tryingtoidentifythegeneratedimagefromthegroundtruth.Thus,the generatortriestogenerateanoutputthatcantrickthediscriminator, resultinginavisuallypleasingandperhapsmorenaturalimage.

ImagesdehazedwithPix2pixaremostpreferredbyobserversbased ontheMOSinTable1.However,althoughthecommentsinTable2 aregenerallypositivetowardsthePix2pixresults,theystillpointout Pix2piximagestobeunnatural.Intermsofnaturalness,weinferthat thecolordistortionscreatedbyDCP-Rskewedthesubjectivescoresin favorofPix2pix,eventhoughDCP-Rhassharperedgesandsmoother regions.SomecommentsalsofavoredCLAHEintermsofnaturalness,al- thoughobserversacknowledgethehazewasnotsuccessfullyremoved.

Infact,manycommentedthatthepresenceofhazeactuallyaddstothe naturalnessoftheimage,asitiscommontoseesuchscenesinthereal world.ManyobserversagreethatthedehazedimagesbyPDR-Netare

theleastvisuallypleasing,consistentwiththeNIQEscores.Manycite theroughedges,artifacts,noise,andlossofdepthasthereasonforthis judgement.

5.Conclusion

Inourexperiments,weconductedimagedehazingonauniformset imagesusingfourmethods,i.e.,CLAHE,DCP-R,PDR-NetandPix2pix.

Thetwo deep learningapproaches,PDR-Net and Pix2pixwereboth trainedusingthesamesetoftrainingimages.Inthiswork,weevaluate dehazingresultsnotonlybyhowsimilartheyaretothegroundtruth, butalsobyimagequalityusingobjectiveandsubjectiveIQA.Theobjec- tivemetricswereselectedtomeasurecolordifference(Δ𝐸𝑎𝑏),contrast andimagestructure(RMSE,SSIM),andnaturalness(NIQE).Thesub- jectiveresultismeasuredwiththemeanopinionscores(MOS).Among allfourmethods,Pix2pixissuperiorintermsofcolordifference,con- trastandimagestructure.TheMOSalsoshowsthatPix2pixisthemost preferredbyhumanobservers.AlthoughPix2pixissurpassedbyDCP-R intermsofobjectivenaturalness,DCP-Rdehazingobtainedthesecond toworstMOSduetothecolordistortionthatsometimesoccurs.CLAHE isunabletoremovehazeatallasreflectedinitsbadobjectivescores.

Interestingly,CLAHEobtainsthesecondbestMOSafterPix2pix,indi- catingthatthepresenceofhazeisnotaproblemforhumanobservers.

Finally,PDR-Netdehazingdoesnotexcelatanyoftheobjectivecrite- ria,althoughitis secondplaceforcolorrecovery.PDR-Net dehazing alsofallsinlastplacefornaturalnessandMOS.

Thecontextofthedehazingbecomesveryrelevanttothediscussion.

Incaseswheretheimagesareintendedtobeviewedbyhumans,CLAHE issimpleandeasytoimplement.However,althoughCLAHEisableto enhancethevisibilityofahazyimage,itwouldbemisleadingtoclaim CLAHEasadehazingmethodsincetheresultsareoftenstillhazy.For afullyautomatedmachinebasedapproach,PDR-Netisstraightforward andsucceedstoobtaingoodreconstructionwithminimalhumaninter- vention.Alternatively,DCP-Risabletodehazeanimagemostnaturally, butinordertoavoidthecolordistortions,itisnecessarytoimprove airlightestimation.Itisnecessarytolookintotherelevantqualitycues ofeachdehazingmethod,anddeterminewhichonesaremoreimpor- tantinthecontextof acertaintask.This isparticularlyrelevantfor largercomputervisionapplications, inwhich higher-levelimageun- derstandingisnecessary.Often,theseapplicationsassumeclearimages asinputs,soanyhazyinputsmustbedehazedpriortoprocessing.In thispaper,weprovidedageneralevaluationofimagequality,focusing oncomprehensiveanalysisanddiscussiononthecolor,imagestructure, andnaturalnessofthedehazedimages.

Inclosing,theabsenceofastandardizedhazydatasetisacommon obstacleinhazyimages.Assuch,ourexperimentwasconductedona fairlylimiteddatasetofhazyimages.Thus,thereismuchtobedesired inregardstothegeneralizationabilityofdehazingmethods.Forfuture work,ageneralapproachwouldbeinstrumentalfordehazingandits subsequentapplications.Inlinewiththistarget,theestablishmentofa largestandardizeddatasetwithhazy,clear,anddepthinformationwill beverybeneficialtoscatteringmediaimageunderstandingingeneral, e.g.,forunderwaterenvironments.

Declarations

Authorcontributionstatement

L.Rahadianti, A. Y. Azizah: Conceivedand designedthe experi- ments;Performedtheexperiments;Analyzedandinterpretedthedata;

Wrotethepaper.H.Deborah:Conceivedanddesignedtheexperiments;

Analyzedandinterpretedthedata;Wrotethepaper.

Fundingstatement

This research was funded by Universitas Indonesia through Hi- bahPublikasi TerindeksInternasional(PUTI) Q3grantnumberNKB-

Referanser

RELATERTE DOKUMENTER

http://www.tabnak.ir/pages/?cid=42. As there is a steady, very important stream of illegal smuggling of fuel out of Iran, where the price is among the world’s lowest, the claim

interview that, “Even if problematic, the Pakistani leadership has realised it has an internal problem it needs to control.” 4 While the Afghan government has repeatedly

228 It further claimed that, up till September 2007, “many, if not most, of the acts of suicide terrorism and attacks on the Pakistani Armed Forces since the Pakistan Army's

A styrofoam mannequin was dressed up with the two suits, one at the time, and the two camouflaged targets were then recorded in 6 various natural backgrounds (scenes) in Rhodes in

This research has the following view on the three programmes: Libya had a clandestine nuclear weapons programme, without any ambitions for nuclear power; North Korea focused mainly on

The system can be implemented as follows: A web-service client runs on the user device, collecting sensor data from the device and input data from the user. The client compiles

As part of enhancing the EU’s role in both civilian and military crisis management operations, the EU therefore elaborated on the CMCO concept as an internal measure for

Based on the above-mentioned tensions, a recommendation for further research is to examine whether young people who have participated in the TP influence their parents and peers in