DigitalImageAuthenticationfromThumbnailsEricKeeandHanyFaridDepartmento - PDF document

DigitalImageAuthenticationfromThumbnailsEricKeeandHanyFaridDepartmentofComputerScience,DartmouthCollege,HanoverNH03755,USAABSTRACTWedescribehowtoexploittheformationandstorageofanembeddedimagethumbnailforimageauthentication.Thecreationofathumbnailismodeledwithaseriesoflteringoperations,contrastadjustment,andcompression.Weautomaticallyestimatethesemodelparametersandshowthattheseparametersdiersignicantlybetweencameramanufacturersandphoto-editingsoftware.Wealsodescribehowthissignaturecanbecombinedwithencodinginformationfromtheunderlyingfullresolutionimagetofurtherrenethesignature'sdistinctiveness.Keywords:DigitalImageForensics,DigitalImageAuthentication1.INTRODUCTIONManyorganizationsarestrugglingwiththeissueofphototampering.Forexample,digitalimages,videos,andaudioarenowroutinelyintroducedasevidenceincivil,criminal,andnationalsecuritycases.Insuchcases,theintegrityofdigitalevidenceiscentral.Digitalforensictechniqueshavebeendevelopedtodetect:regionduplication1,2;inconsistenciesincameraresponsefunction3;inconsistenciesinlighting4;andinconsistenciesinsensornoise5{see6forageneralsurvey.However,relativelybenignmodicationssuchascroppingorcontrastadjustmenteithercannotbedetectedbythesetechniques,orrenderthesetechniquesineective.Inthelegalsettingwearesometimesinterestedindeterminingifadigitalimagehasbeenalteredinanywayfromthetimeofrecording,includingsomethingassimpleascropping.Hereweapproachthisproblemofimageauthenticationbyexploitingacamera'scustomizedimageencodingformat.Wedescribehowtoexploittheformationandstorageofanembeddedimagethumbnailforimageauthentica-tion.Specically,wemodelthecreationofathumbnailwithaseriesoflteringoperations,contrastadjustment,andcompression.Wedescribeatechniqueforautomaticallyestimatingthesemodelparameters,andshowthattheseparameters,althoughnotunique,candiersignicantlybetweencameramanufacturersandphoto-editingsoftware,Figure1.Assuch,theseparameterscanbeusedasasignatureforimageauthentication.Wealsode-scribehowthissignaturecanbecombinedwithencodinginformationfromtheunderlyingfullresolutionimage7tofurtherrenethesignature'sdistinctiveness.2.METHODSEmbeddedinanimageheaderisathumbnailversionofthefullresolutiondigitalimage.Wedescribeamodelforthecreationofathumbnail,andthendescribehowtoestimatethesemodelparameters.2.1.ThumbnailModelAthumbnailistypicallyontheorderof160120pixelsinresolution.Givenanimagef(x;y),thethumbnailiscreatedbyaseriesofsixsteps:crop,pre-lter,down-sample,post-lter,contrastandbrightnessadjustment,andJPEGcompression.Iftheoriginalresolutionimageisofadierentaspectratiothanthethumbnail,thentheimageneedstoeitherbepaddedorcroppedaccordingly.Theamountofpadding/croppingisspeciedbyfourparameterscl,cr,ct,cb,whereclandcrcorrespondtothepadding/croppingontheleftandright,andctandcbonthetopandbottom,Figure2.Apositivevaluecorrespondstoapadding(withapixelvalueof0),andanegativevaluecorrespondstoacropping.Denotingthecroppedimageas^f(x;y),thenextfourprocessingstepsarespeciedasfollows:t(x;y)=Df^f(x;y)?h1(x;y)g?h2(x;y)+;(1) Contact:kee@cs.dartmouth.eduandfarid@cs.dartmouth.edu Figure1.ShownfromlefttorightareanoriginalthumbnailgeneratedbyaNikonD200,athumbnailofthesameimagegeneratedbyAdobePhotoshopCS3,andthedierencebetweenthesethumbnails.TheNikonthumbnailis160120pixelsinsize,whilethePhotoshopthumbnailis160107.InadditiontheseJPEGencodedthumbnailsemploydierentcroppingboundariesandquantizationtablesleadingtofurtherdierencesascanbeseenintheright-mostpanel.wheret(x;y)isthethumbnail,h1(
)isthepre-lter,Df
gisthedown-samplingoperator,h2(
)isthepost-lter,?istheconvolutionoperator,andandarethemultiplicativecontrastandadditivebrightnessadjustmentterms,respectively.Thepre-lteristypicallyalow-passlterappliedtoavoidspatialaliasingpriortodown-sampling,andtheoptionalpost-lteristypicallyasharpeninglter.Inthenalstep,thethumbnailisJPEGcompressedwithaspeciedquantizationtable.Inordertosimplifythismodel,aseriesofassumptionsaremade:thepre-lterisassumedtobeacircularlysymmetricGaussian,exp((x2+y2)=2),withwidththepre-lterisunitsumthepost-lteris33pixelsinsizethepost-lterissymmetric(h2(x;y)=h2(x;y)andh2(x;y)=h2(x;y)),yieldingalteroftheform(aba;bcb;aba)thepost-lterisunit-sum,constrainingc=1(4a+4b).Withtheseconstraints,thefullmodelforcreatingathumbnailisspeciedby11processingparameters:2forthesizeofthethumbnail,4forthecropping/padding,1forthepre-lter,2forthepost-lter,2forthecontrastandbrightness.Inaddition,thereare128compressionparameters:theJPEGquantizationtableisspeciedbytwo88tablescorrespondingtothequantizationfortheluminanceandchrominancechannels.Thisyieldsatotalof139modelparameters.Inthefollowingsectionwedescribehowtoestimatethesemodelparameters.2.2.ThumbnailEstimationIntherststepofthumbnailconstruction,arectangularcroppingboundaryrelativetothefullresolutionimageisspecied.Thiscroppingboundaryisdeterminedbyanisotropicallyscalingandtranslatingaboundingboxofthesamesizeasthefullresolutionimagesuchthatthecroppedanddownsampledimagematchestheextractedthumbnail,Figure2.Thecroppingparametersareestimatedbyrstspecifyinganinitialboundarythathasthesameaspectratioasthethumbnail,isscaledtoencompassthemaximumimagedimension,andistranslatedsuchthattheimageiscenteredwithintheboundary,Figure2.Thefullresolutionimagef(x;y)isthencroppedaccordingtothisboundary,paddedwithzeros(wherenecessary),anddownsampledtoyieldaninitialthumbnail:^t0(x;y)=Df^f0(x;y)g;(2)where^f0(x;y)istheinitialcroppedimage,andthedownsamplingrateismax(Nx=nx;Ny=ny),where(Nx;Ny)and(nx;ny)aretheimageandthumbnaildimensions,respectively.Thisinitialthumbnailisthenanisotropicallyscaledandtranslatedtomatchtheextractedthumbnailt(x;y):t(x;y)=^t0(sxx+
x;syy+
y);(3) Weassumethatthetwochrominancechannelsemploythesamequantizationtable. Figure2.Creatingthecropboundaryforthefullresolutionimagef(x;y).Toprow:createaninitialcropboundary^f0(x;y)sothatitsaspectratiomatchesthethumbnailt(x;y);down-sampletocreate^t0(x;y);alignwiththeactualthumbnailt(x;y)byscalingandtranslatingbyM=(sx;sy;
x;
y)toyield^t(x;y).Bottomrow:adjusttheinitialcropboundary^f0(x;y)byMtoyieldthedesiredcropboundary^f(x;y).Thecropboundaryisspeciedbythemarginscl,cr,ct,cb.where(sx;sy)and(
x;
y)arethescalingandtranslationparameters.Theseparametersareestimatedusingacoarse-to-nedierentialregistrationtechnique.8,9Thisregistrationisperformedonagrayscaleversionofthecolorthumbnail.Thescalingandtranslationparametersarethenusedtospecifythecroppingboundaryforthefullresolutionimage.Specically,thecoordinatesofthefourcornersoftheinitialcroppedimage^f0(x;y)arescaledandtranslatedyby(sx;sy)and(
x;
y),yieldingthedesiredrectangularcroppingboundary^f(x;y),Figure2.Thisboundaryisparametrizedbythehorizontalandverticalmarginscl;cr;ct,andcb,speciedasapercentageoftheimagedimensions,Figure2.Notethatwhendown-sampling^f(x;y),thesamplingrateinthehorizontalandverticaldirectionsmustbeindependentlyadjustedsothatthenaldimensionsmatchthatofthethumbnailt(x;y).Next,thepre-lterh1(
),post-lterh2(
),andcontrastandbrightnesstermsandareestimatedbyminimizingthefollowingerrorfunction:E1(h1;h2)=Xx;yht(x;y)Df^f(x;y)?h1(x;y)g?h2(x;y)i2:(4) yBecausethetranslationparameters(
x;
y)arespeciedinthumbnailcoordinates,theymustbescaledbythedownsamplingratebetween^f0(x;y)and^t0(x;y). Notethatthiserrorfunctionisonlyspeciedintermsofthepre-andpost-lter,andnotthecontrastandbrightness.Givenapairoflters,thecontrastandbrightnessareestimatedbyminimizingthefollowingerrorfunction:E2(;)=Xx;yt(x;y)(^th(x;y)+)2;(5)where^th(x;y)=Df^f(x;y)?h1(x;y)g?h2(x;y).Thiserrorfunctionisminimizedusingstandardleast-squaresestimation.ThesummationinE1(
)andE2(
)areperformedoverallthreecolorchannelsofthethumbnail.TheerrorfunctionE1(
)isminimizedusingabrute-forcesearchoverthelterparameters,whereoneachiterationofthissearch,theerrorfunctionE2(
)isminimizedanalyticallytoyieldthebestcontrastandbrightnesstermsforthespeciedlters.TheseparametersarerenedusinganiterativeNelder-Meadminimization,whichisbootstrappedwiththeresultsofthebrute-forceminimization.Inpractice,wehavefoundthattheminimizationofE1(
)isslightlymoreeectivewhenperformedintheFourierdomain.Specically,weminimize:E1(h1;h2)=X!x;!yW(!x;!y)jT(!x;!y)jj^T(!x;!y)j2;(6)whereT(
)istheFouriertransformoftheactualthumbnailt(
),^T(
)istheFouriertransformofourconstructedthumbnail(Df^f?h1g?h2)+,andj
jdenotesthemagnitudeoftheFouriertransform.Thiserrorfunctionisfrequencyweightedwithahighpasslter,W(
),becauseaftertheinitialalignmentofEquation(3)thelow-frequenciesarealreadywellaligned.Lastly,thethumbnaildimensionsand128thumbnailcompressionparametersareextracteddirectlyfromtheJPEGheader.Notethatsincetheoriginalthumbnailt(
)wascompressedwiththeseparameters,wemustalsocompressourestimatesofthethumbnailduringtheparameterestimationstagesdescribedabove.Specically,^t0(
)iscompressedpriortoestimatingthescalingandtranslationparameters,and^th(
)iscompressedpriortoestimatingthecontrastandbrightnessterms.3.RESULTSToverifythatthumbnailparameterscanbereliablyusedforimageauthentication,adatabaseof1514unmodiedimagesweredownloadedfromFlickr.AnimagewasconsideredtobeunmodiediftheimagewastaggedbyFlickras\original",nometadataeldshadbeenremoved,themetadata\modication"and\original"dateswereconsistent,andthemetadata\software"eldwasempty.These1514imagesspanned142camerasofdierentmakeandmodel,AppendixA.Camerasofthesamemakeandmodelsometimesvarytheirimageandthumbnailsizeandquantizationtablez.Becausethesevariationsaecttheoverallthumbnailparametrization,wepartitionedthe1514imagesinto245\cameraclasses"containingimagesfromthesamemakeandmodelandimage/thumbnailsizeandquantizationtable.Forcomputationaleciencythethumbnailparameterswererstestimatedforarepresentativeimagefromeachofthe245cameraclasses.Thecropboundarywasestimatedasdescribedintheprevioussection.Thepre-lter,post-lter,andcontrastandbrightnesswereestimatedusingabrute-forcesearch,Equation(4).Theparameterspaceforthepre-lterwaspartitionedintotherange[0:005;1]in20equalsteps.xTheparameterspaceforthetwocomponentsofthepost-lterwerepartitionedintotherange[0:5;0:5]in11equalsteps.Recallthatthecontrastandbrightnesstermswereestimatedanalyticallywithineachiterationofthebruteforcesearch,Equation(5).ThesebruteforcesolutionswerethenusedtobootstraptheNelder-Meadminimizationfortheremainingimageswithineachcameraclass. zVariationsinimage/thumbnailsizeandquantizationtablecan,forexample,beduetodierencesinrmware.xInordertoappropriatelycomparethewidthofthepre-lterh1(
)acrossimagesofvaryingsize,thesizeofthepre-lterh1(
)isnormalizedrelativetotheimagedimensions.Specically,thepre-lterissampledovertheinterval[1;1]atarate1=150ofthemaximumdimensionoftheoriginalimage. 5 10 15 20 25 30 35 40 45 50 55 0 0.2 0.4 0.6 0.8 1 Equivalence Class SizeProbability 5 10 15 20 25 30 35 40 45 50 55 0 20 40 60 80 100 Equivalence Class SizeCount 5 10 15 20 25 30 35 40 45 50 55 0 0.2 0.4 0.6 0.8 1 Equivalence Class SizeProbability 5 10 15 20 25 30 35 40 45 50 55 0 20 40 60 80 100 Equivalence Class SizeCount Figure3.Shownintheleftcolumnistheprobabilitythatacameraiscontainedinanequivalenceclassofaspeciedsize.Shownintherightcolumnisthenumberofequivalenceclassesofaspeciedsize.Shownalongthetoprowarethedistributionsbasedonallofthethumbnailparameters(40:8%ofthecamerashaveuniqueparameters,i.e.,anequivalenceclasssizeofone).Showninthebottomrowarethedistributionsforonlythethumbnailsizeandquantizationtable(23:7%ofthecamerashaveuniqueparameters).Becausetheabsolutevalueofthecontrastandbrightnesstermsmaydependontheunderlyingimagecontent,theseparametersarecombinedintoasinglebinary-valuedparametercorrespondingtothepresenceorabsenceofcontrast/brightnessadjustment.Specically,acameraissaidtoapplyacontrast/brightnessadjustmentifthecamera'scontrastdeviatesbymorethan0:075fromunity,orifthebrightnessdeviatesbymorethan0:05fromzero(assumingaluminancescaleof[0;1]).Thesethresholdsweredeterminedbyrstcomputingthedierencebetweeneachimage'sthumbnailparameterandthemeanparameterofthecorrespondingcameraclass.Thethresholdwasthentakentobethewidthofthe99%condenceintervalofthedistributionofthesedierences.Wenextevaluatetheconsistencyanddistinctnessofthethumbnailparameterswithinandacrosscameraclasses.Thethumbnailparametersforagivencameraclassaredeterminedbycomputingthemeanofthecropboundary,themeanofthepre-andpost-lters,andthemodeofthecontrast/brightness,acrossallimagesinacameraclass.Thethumbnailsizeandquantizationtableare,bydenition,thesameforallimagesinacameraclass.Inordertoreasonablyequatereal-valuedmodelparameters(cropboundary,pre-lter,post-lter),theseparameterswereconsideredasequivalentiftheirabsolutedierencewasbelowaspeciedthreshold.The 1 2 3 4 5 6 0 0.2 0.4 0.6 0.8 1 Equivalence Class SizeProbability 1 2 3 4 5 6 0 0.2 0.4 0.6 0.8 1 Equivalence Class SizeProbability Figure4.Shownontheleftistheprobabilitythatacameraiscontainedinanequivalenceclassofaspeciedsizebasedonalloftheimageandthumbnailparameters(72:2%ofthecamerashaveuniqueparameters,i.e.,anequivalenceclasssizeofone).Shownontherightisthecumulativeprobabilitydistribution.thresholdforthecropboundarywas0:01,thepre-lter0:2,andthepost-lter0:075.Thesethresholdsweredeterminedinthesamewayasforthecontrastandbrightnessthresholdsdescribedabove.Twothumbnailmodelswereconsideredtobeequivalentiftheirinteger-valuedparameters(contrast/brightness,size,quantizationtable)werethesameandiftheirreal-valuedparameters(cropboundary,pre-lter,post-lter)werewithinthespeciedthresholdsdescribedabove.Cameraswerethengroupedintoequivalenceclassesofidenticalmodels.ShowninthetopleftpanelofFigure3istheprobabilitythatacameraiscontainedinanequivalenceclassofaspeciedsizebasedonusingallofthethumbnailparameters:40:8%ofthecamerasareinanequivalenceclassofsizeone(i.e.,areunique),9:8%areinanequivalenceclassofsizetwo,8:2%areinanequivalenceclassofsizethree,andthelargestequivalenceclassisofsize48,with19:6%ofthecameras.Thereisonlyoneclassofsize48anditcontains42CanonPowershotsand6CanonDigitalIxusofvaryingmodels.ShowninthetoprightpanelofFigure3isthedistributionofequivalenceclasssizes(i.e.,thenumberofequivalenceclassesofeachsize).Forcomparison,showninthebottomrowofFigure3arethedistributionsforonlythethumbnailsizeandquantizationtable:23:7%ofthecamerasareinanequivalenceclassofsizeone,8:25%areinanequivalenceclassofsizetwo,7:4%areinanequivalenceclassofsizethree,andthelargestequivalenceclassisofsize56,with22:9%ofthecameras.Notethattheadditionofthethumbnailprocessingparametersimprovesthedistinctivenessofthesignature.ShowninFigure4istheprobabilitythatacameraiscontainedinanequivalenceclassofaspeciedsizebasedonusingallofthethumbnailparametersandthefullresolutionimagesizeandquantizationtable.Inthiscase,72:2%ofthecamerasareinanequivalenceclassofsizeone(i.e.,areunique),11:4%areinanequivalenceclassofsizetwo,6:1%areinanequivalenceclassofsizethree,andthelargestequivalenceclassisofsize6,with2:5%ofthecameras.Theadditionoftheimageparameterssignicantlyimprovesthedistinctivenessofthesignature.Acloserlookattheseequivalenceclassesrevealssomeinterestingtrends.Thereisoneequivalenceclassofsize6anditcontainssix8-megapixelCanonPowershotcamerasofvaryingmodels.Therearethreeequivalenceclassesofsize5:onecontains8-megapixelCanonPowershotandCanonIxuscamerasandtheremainingclasseseachcontainamixtureof7-megapixelCanonPowershotandCanonIxuscamerasofvaryingmodels.Withonlytwoexceptions,theremainingequivalenceclassesofsizegreaterthan1arepopulatedwithCanoncameraswhichseemtobeparticularlyconsistentintheirthumbnailandimageparameters.SeeAppendixAforacompletebreak-downofthesedistributions. 1 2 3 4 5 6 0 0.2 0.4 0.6 0.8 1 Equivalence Class SizeProbability 1 2 3 4 5 6 0 0.2 0.4 0.6 0.8 1 Equivalence Class SizeProbability Figure5.Shownontheleftistheprobabilitythatacameraiscontainedinanequivalenceclassofaspeciedsizebasedononlytheimagesizeandquantizationtable(40:4%ofthecamerashaveuniqueparameters,i.e.,anequivalenceclasssizeofone).Shownontherightisthecumulativeprobabilitydistribution.Lastly,showninFigure5istheprobabilitydistributionsbasedonusingonlythefullresolutionimagesizeandquantizationtable.Inthiscase,40:4%ofthecamerasareinanequivalenceclassofsizeone,18:8%areinanequivalenceclassofsizetwo,18:4%areinanequivalenceclassofsizethree,andthelargestequivalenceclassisofsize6,with9:8%ofthecameras.Notethattheadditionofthethumbnailparameters,Figure4,signicantlyimprovesthedistinctivenessofthesignature.Tofurthertesttheutilityofthumbnailparametersforauthentication,weanalyzedthethumbnailparametersusedbyPhotoshop(CS3).AnimagewassavedwithPhotoshopateachof13possibleJPEGqualitysettings.Thethumbnailandimageparameterswereestimatedforeachimageandcomparedtotheremaining245cameraclasses.NoneofthePhotoshopthumbnailparametersweresharedbyanyofthecameraclasses.WealsoindividuallycomparedthePhotoshopparameterstothecameraclasses:Nonesharedthesameimagequantizationtable,thumbnailquantizationtable,orpre/postlter;elevenofthecameraclassesusedthesamecropparameters:nikond300andcoolpixl4;fujilmnepixs5700/s700,s100fs,andz20fd;canonpowershotsd400ands2is;casioex-z77,ex-z1050,ex-s880,andex-z60;theolympuse-510andthesonydsc-w120.4.DISCUSSIONAnimagethumbnailiscreatedbyaseriesofsixsteps:crop,pre-lter,down-sample,post-lter,contrastandbrightnessadjustment,andJPEGcompression.Wehavedescribedhowtheseparameterscanbeestimated,andhaveshownthattheyvarysignicantlybetweencameramanufacturersandphoto-editingsoftware.Thedistinc-tivenessoftheseparametersbecomesevenmorepronouncedwhentheyarecombinedwiththefullresolutionimagesizeandcompressionparameters.Assuch,thesethumbnailandimageparameterscanbeusedforimageauthentication(determiningifanimagewasalteredinanywayfromthetimeofitsrecording).Theeectivenessofthisapproachdependsonbuildingalibraryofthumbnailandimageparametersextractedfromalargecollectionofimages.Wehavecollected1:2million\original"imagesfromwhichweplantobuildthislibrary.Fromourinitialtests,wehavefoundthat83%oftheimageshaveathumbnail.Fortheimageswithoutthumbnails,thefullresolutionimageparameterscanstillbeusedforauthentication.ACKNOWLEDGMENTSThisworkwassupportedbyagiftfromAdobeSystems,Inc.,agiftfromMicrosoft,Inc.,andagrantfromtheNationalScienceFoundation(CNS-0708209). REFERENCES1.J.Fridrich,D.Soukal,andJ.Lukas,\Detectionofcopymoveforgeryindigitalimages,"inProceedingsofDigitalForensicResearchWorkshop,August2003.2.A.PopescuandH.Farid,\Exposingdigitalforgeriesbydetectingduplicatedimageregions,"Tech.Rep.TR2004-515,DepartmentofComputerScience,DartmouthCollege,2004.3.Z.Lin,R.Wang,X.Tang,andH.-V.Shum,\Detectingdoctoredimagesusingcameraresponsenormalityandconsistency,"inComputerVisionandPatternRecognition,(SanDiego,CA),2005.4.M.JohnsonandH.Farid,\Exposingdigitalforgeriesincomplexlightingenvironments,"IEEETransactionsonInformationForensicsandSecurity3(2),pp.450{461,2007.5.J.Lukas,J.Fridrich,andM.Goljan,\Digitalcameraidenticationfromsensornoise,"IEEETransactionsonInformationSecurityandForensics1(2),pp.205{214,2006.6.H.Farid,\Asurveyofimageforgerydetection,"IEEESignalProcessingMagazine2(26),pp.16{25,2009.7.H.Farid,\DigitalimageballisticsfromJPEGquantization:Afollowupstudy,"Tech.Rep.TR2008-638,DepartmentofComputerScience,DartmouthCollege,2008.8.E.Simoncelli,HandbookofComputerVisionandApplications,ch.BayesianMulti-scaleDierentialOpticalFlow,pp.397{420.AcademicPress,1999.9.H.FaridandJ.Woodward,\Videostabilizationandenhancement,"Tech.Rep.TR2007-605,DepartmentofComputerScience,DartmouthCollege,2007. AppendixAEachentryinthistablecorrespondstoanequivalenceclassofidenticalthumbnailandimageparameters,Figure4.Thesizeoftheequivalenceclassisintherstcolumnandthemake,model,andresolutionareinthesecondcolumn.Forcompactness,allcamerasinanequivalenceclassofsizeonearelistedinasingleentry{forthesecameras,theparentheticalfollowingthemakeandmodelcorrespondstothenumberofinstancesofthesamemakeandmodelwithdierentthumbnailandimageparameters. 6 canonpowershotsd870is(8mp){canonpowershotsd850is(8mp){canonpowershots80(8mp){canonpowershota590is(8mp){canonpowershotsx100is(8mp){canonpowershota720is(8mp) 5 canonpowershotsd1100is(8mp){canonpowershots5is(8mp){canonpowershota630(8mp){canonpowershota580(8mp){canondigitalixus80is(8mp) 5 canonpowershotsd750(7mp){canonpowershotsd800is(7mp){canonpowershota570is(7mp){canonpowershotsd550(7mp){canondigitalixus850is(7mp) 5 canonpowershota710is(7mp){canondigitalixus70(7mp){canonpowershota620(7mp){canonpowershottx1(7mp){canonpowershota550(7mp) 4 canonpowershota530(5mp){canonpowershota610(5mp){canonpowershotsd400(5mp){canonpowershota460(5mp) 3 canonpowershotsd600(6mp){canonpowershota540(6mp){canonpowershotsd630(6mp) 3 canoneosdigitalrebelxti(10.1mp){canoneos400ddigital(10.1mp){canoneoskissdigitalx(10.1mp) 3 canoneosdigitalrebel(6.3mp){canoneos300ddigital(6.3mp){canoneos10d(6.3mp) 3 sonydsc-w55(7.2mp){sonydsc-h5(7.2mp){sonydsc-w35(7.2mp) 3 canonpowershots3is(6mp){canondigitalixus800is(6mp){canondigitalixus60(6mp) 2 canonpowershotsd600(6mp){canonpowershots3is(6mp) 2 canoneosdigitalrebelxt(8mp){canoneos350ddigital(8mp) 2 canoneos30d(8.2mp){canoneos20d(8.2mp) 2 canoneosdigitalrebelxsi(12.2mp){canoneos450d(12.2mp) 2 canoneosdigitalrebelxsi(12.2mp){canoneos450d(12.2mp) 2 canonpowershota510(3.2mp){canonpowershota95(5mp) 2 canonpowershotsd450(5mp){canonpowershots2is(5mp) 2 canonpowershotg9(12.1mp){canonpowershotsd950(12.1mp) 2 canonpowershota520(4mp){canonpowershota85(4mp) 2 canonpowershota75(3.2mp){canonpowershotsd200(3.2mp) 2 canonpowershotsd1000(7.1mp){canonpowershots230(3.1mp) 2 canonpowershota640(8mp){canonpowershotsd790is(10.1mp) 2 kodakeasysharem1033(10mp){kodakeasysharez1012(10mp) 2 canonpowershotg7(10mp){canonpowershotsd900(10mp) 1 appleiphone(2x){camera8mp-9hx(1x){canondigitalixus30(1x){canondigitalixus50(1x){canondigitalixus700(1x){canondigitalixus75(1x){canoneos20d(1x){canoneos30d(1x){canoneos400ddigital(5x){canoneos40d(5x){canoneos5d(1x){canoneosdigitalrebel(1x){canoneosdigitalrebelxsi(1x){canoneosdigitalrebelxt(5x){canoneosdigitalrebelxti(3x){canoneoskissx2(1x){canonpowershota400(1x){canonpowershota560(2x){canonpowershota650is(1x){canonpowershota70(1x){canonpowershota80(1x){canonpowershota95(1x){canonpowershotg9(1x){canonpowershots2is(1x){canonpowershots30(1x){canonpowershotsd1000(3x){canonpowershotsd1100is(1x){canonpowershotsd300(1x){canonpowershotsd400(1x){canonpowershotsd600canonpowershotsd700is(1x){canonpowershotsd750(1x){casioex-s880(1x){casioex-z1050(2x){casioex-z60(1x){casioex-z77(1x){casioex-z80(1x){casiolex-z850(1x){digitalcamera6mp-9fw(1x){fujilmnepixs100fs(1x){fujilmnepixs2000hd(1x){fujilmnepixs5700s700(7x){fujilmnepixz20fd(1x){hpphotosmartm537(1x){hpphotosmartr717(1x){hpphotosmartr740(1x){konicaminoltadimagez3(1x){nikoncoolpixl4(1x){nikond100(1x){nikond200(4x){nikond300(5x){nikond3(2x){nikond40(14x){nikond40x(2x){nikond50(12x){nikond60(4x){nikond70(4x){nikond70s(4x){nikond80(15x){nikond90(1x){olympuse-500(2x){olympuse-510(2x){olympusu1030sw,s1030sw(1x){olympusu850sw,s850sw(1x){panasonicdmc-fx9(1x){panasonicdmc-fz18(1x){panasonicdmc-fz20(1x){panasonicdmc-fz5(1x){panasonicdmc-fz7(1x){panasonicdmc-fz8(1x){panasonicdmc-tz1(1x){panasonicdmc-tz3(1x){panasonicdmc-tz4(1x){panasonicdmc-tz5(1x){pentaxk100d(1x){pentaxk100dsuper(1x){pentaxk10d(3x){pentaxoptios5z(1x){samsungtechwinkenoxs630/samsungs630&#x-5.1;䡣(1x){sanyovpc-s600(1x){sanyovpc-t700(1x){sonydsc-p200(1x){sonydsc-p92(1x){sonydsc-t300(1x){sonydsc-t70(1x){sonydsc-t9(1x){sonydsc-w120(1x){sonydsc-w50(1x){sonydsc-w55(1x){sonydsc-w70(1x){sonydsc-w80(1x){sonydslr-a100(1x){sonydslr-a200(2x){sonydslr-a350(1x)