原型基于颜色的图像检索及MATLAB.docx

原型基于颜色的图像检索及MATLAB.docx

《原型基于颜色的图像检索及MATLAB.docx》由会员分享，可在线阅读，更多相关《原型基于颜色的图像检索及MATLAB.docx（19页珍藏版）》请在冰豆网上搜索。

原型基于颜色的图像检索及MATLAB

PrototypingColor-basedImageRetrievalwithMATLAB®

PetteriKerminen1,MoncefGabbouj2

1TampereUniversityofTechnology,Pori,Finland

2TampereUniversityofTechnology,SignalProcessingLaboratory,Tampere,Finland

Abstract

Content-basedretrievalof（image）databaseshasbeemorepopularthanbefore.Algorithmdevelop-mentforthispurposerequirestesting/simulationtools,buttherearenosuitablemercialtoolsonthemarket.

Asimulationenvironmentforretrievingimagesfromdatabaseaccordinghistogramsimilaritiesispresentedinthispaper.Thisenvironmentallowstheuseofdifferentcolorspacesandnumbersofbins.ThealgorithmsareimplementedwithMATLAB.Eachcolorsystemhasitsownm-files.

Thephasesofthesoftwarebuildingprocessarepre-sentedfromsystemdesigntographicaluserinterface（GUI）.ThefunctionalityisdescribedwithsnapshotsofGUI.

1.Introduction

Nowadaystherearethousandsorhundredsofthousandsofdigitalimagesinanimagedatabase.Iftheuserwantstofindasuitableimageforhis/herpurposes,he/shehastogothroughthedatabaseuntilthecorrectimagehasbeenfoundoruseareferencebookorsome“intelligent”program.Videoondemand（VoD）servicesalsorequiresanintelligentsearchsystemforend-users.VoDsystems’searchmethodsdifferslightlyfromimagedatabase’smethods.

Areferencebookisasuitableoption,iftheimagesarearrangedwithausefulmethod,forexample:

1）categories:

animals,flags,etc,2）names（requiresagoodnamingtechnique）or3）dates.Anexperiencedusercanusethesesystemsaswellastextualsearches（keywordshavetobeinsertedinadatabase）efficiently.Therearesituationswhenamulti-languagesystemhastobeused.Therealanguageindependentsearchsystem’sbestpropertiescanbeutilized.Atoolwhichisbasedontheimages’propertiescanbemade

languageindependent.Thesepropertiescanbeforexamplecolor,shape,texture,spatiallocationofshapeetc.

IntheMuVi-project[1]thiskindoftoolisunderconstruction.Itwillcoverthepropertiespresentedabove.Researchworkoncontent-basedimageretrievalhasbeendonein[2–6].Thesystem,whichispresentedinthispaper,isasimulationenvironment,whereMuVi’scolorcontentbasedretrievalhasbeendevelopedandtested.

2.Systemdevelopment

MATLABisanefficientprogramforvectorandmatrixdataprocessing.Itcontainsreadyfunctionsformatrixmanipulationsandimagevisualizationandallowsaprogramtohavemodularstructure.BecauseofthesefactsMATLABhasbeenchosenasprototypingsoftware.

2.1Systemdesign

Beforeanym-fileshavebeenwritten,thesystemdesignhasbeendone.AsystemdesignfortheHSV（hue,saturationandvalue）colorsystembasedretrievalprocessispresentedinFigure1.Similardesignhasbeendoneforallusedcolorsystems.

Figure1:

FunctionchartforHSVcolorspacewith27binshistogram.

Tesths27isthemainfunctionforthiscolorsystemandthisnumberofbins.Itcallsotherfunctions（hs27read,dif_hsvandimage_pos）whenneeded.Each

colorsystemhasamainfunctionofitsownandvariablenumber（2–3）ofsub-functions.Ifthereisnoneedforcolorspaceconversionthereare2functions,otherwise3functionsonthefirstbranchofthefunctionchart.

Thefunctioncallofthemainfunctionis:

matches=tesths27（imagen,directory,num）

Thevariableimagenspecifiesthequeryimage’snameandpath.Thedirectoryisapathoftheimagedatabaseandnumisadesirednumberofretrievedimages.

2.2Functions

Atthismomenttherearefunctionsimplementedforfourcolorspaces:

HSV,L*a*b*,RGBandXYZ[7].Eachcolorspacehasfrom2to4implementationsfordifferentnumbersofbins.Therearealtogether14mainfunctions.

Forsomecolorsystemsitispossibletomakethesefunctionsdynamic,i.e.dynamichistogramcalculation.Everycolorsystem/binbinationrequiresitsownhistogramsandthesecanbemadeonlywithanexhaustivemethod（pixelbypixel）.Histogramcalculationtakes½-5minutesperimage,each

approximately320×240pixels,dependingontheplexityofthecolorspaceon150MHzPentium.Thusitisnotreasonabletolettheuserselectabinnumberfreely,especiallyinthecaseoflargedatabases.

Thefunctionshavebeennamedsothatthenamescontaininformationofthecolorspaceused,thepurposeofthefunctionsandthenumberofusedbins.Somefunctions,forexampleimage_pos,havebeenusedbymanyorallmainfunctionsandthesefunctionshavenotbeennamedasdescribedabove.

Themainfunctionchecks,ifthefunctioncalliscorrect.Ifthequeryimage’snamedoesn’tcontainapath,thefunctionassumesthattheimageissituatedinthedatabasedirectory.Inadditiontothis,themainfunctionchecks,ifthequeryimagealreadyhasahistograminthecurrentlyuseddatabase.Iftherequiredhistogramisnotthere,theimageread（forexamplehs27read）functioniscalled.Thisfunctionalsonormalizespixelvaluesandarrangesimagematrixdatatoavectorformat.Afterthatstageacolorspaceconversionfunction（ifneeded）iscalled.Finallyaquantizationfunctionbuildsthehistogramwiththecorrectnumberofbins.

Thehistogramwillthenbesavedintothedatabasedirectory.Ifthehistogramalreadyexiststhere,thethreepreviousstepswillnotbeexecuted.Nowthequeryimagehasbeenanalyzed.Thenthemainfunctionwillgothroughallimagesinthedatabasedirectorywithanalmostsimilaralgorithmasinthecaseofthequeryimage.Thedifferenceisthatnowtherewillbea

histogramdifferencecalculationbetweenthequeryimage’sandcurrentimage’shistogram.Finallytheimage_posfunctionwillbeusedtoputaqueryimageandthedesirednumberofbestmatchimagesonthedisplay.

2.3Linking

Itisnotpossibletouseaprogrambeforethemainfunctionandsub-functionsareconnectedtoeachother.Themainfunctionwillbecalledfromthemandlineorthroughthegraphicaluserinterface,whichwillbepresentedlaterinthispaper.Inbothcasesthefunctioncallwillcontainthesamearguments.Formulti-levelsearchpurposesseparatemainfunctionshavebeenimplemented,butitispossibletoutilize“normal”functionsandaddoneparameter,wherethebestmatchesarraycanbetransferredforsecondastageparisonfunction.

Themainfunctioncallsanimagereadfunctionwiththeimage’sname.Thehistogramwillbereturnedtothemainfunction.Ifacolorspaceconversionisneeded,theconversionfunctionwillbecalledfromthereadfunctionwithr,gandb–vectors.Thehistogramwillbereturnedtothecallingfunction.Finallythehistogrambuildfunctionwillbecalledwithconvertedcolorvectors.Thisfunctionreturnsaquantizedhistogram,whichwillgothroughallfunctionsuntilitachievesthemainfunction.

Themainfunctioncallsthehistogramdifferencefunctionwithtwohistogramvectorsandwillgetadifferencevalueasaresponse.ThedifferencefunctionusesEuclidean-distancecalculation,butitcanbeeasilychangedto

anotheralgorithmduetothemodularityoftheprogram.Ifthedifferenceissmallerthanlargestdifferenceonabestmatchtable,thecurrentresultwillbewrittenoverthelastresultonthebestmatchtable.Afterthatthetableisarrangedagaininanascendingorderofdistance.Whenalltheimageshavebeenanalyzed,thesortedbestmatchtable,thenumberofdesiredoutputimages,thequeryimage’sname,thesearchimage’spathandthedatabase

patharetransferredtotheimage_posfunction.Thesevaluescanbetransferredintolargerponents（vectors/containers）.Nowtheprogramworksfasterwithseveralinputarguments,becausethereisnoneedfor

pickingupvariablesfromacontainer.

2.4Graphicaluserinterface

Thegraphicaluserinterface（GUI）isanimportantpartofsoftwaredevelopment.ThedesigningoftheGUIhavetosolvethefollowingproblems:

learningtime,speedofperformance,rateoferrorsbyusers,retentionovertime,andsubjectivesatisfaction[9].Thissoftwareis,atthemoment,intendedtobeusedonlyfortestingpurposes.Themostimportantpropertyofthissoftwareisthattheresultsofdifferenttestqueriescanbeseenquicklyandthe

resultscanbesavedsafelyonadisk.Thusthevisuallayoutisnotasimportantasincaseofamercialsoftwareproduct.

InFigure2thefirstscreenonGUIispresented.Thepurposesofthebuttons,menusandotherponentswillbepresentedlater.Ifthissoftwareisdevelopedintoamercialproduct,themenubarwillbedisabledinthefutureandtheexitandhelpbuttonswillbeaddedonthecanvas.

Figure2:

GUIbeforethesearchimageselection.

InFigure3thesearchscreenispresentedjustbeforestartingasearch.Theuserisshownasearchimage,andinthiswayhe/shecanbesurethatthesearchwillbemadewiththecorrectimage.

Figure3:

GUIjustbeforerunningaquery.

TheresultsofthequerywillbepresentedonthescreenintheformatwhichispresentedinFigure6.

3.Usingthesoftware

ThefirstscreenhasalreadybeenpresentedinFigure2.Theusercanchoosefrompop-upmenus（seeFigure4）,ifthesearchismadewithoneacolorsystemorasamulti-levelsearch.Inaone-levelsearcharoughlyquantizedoramoreaccuratehistogramisusedinoneloop（onecolorsystem）.

Figure4:

Colorsystemselectionfromapopupmenu.

Thesecondmenuisdisabledbecauseaone-levelsearchisselected.

Inamulti-levelsearchtwodifferentcolorsystems/histogramsareused.Duringthefirstlooptheroughlyquantizedhistogramsareusedandduringthesecondloop.

themoreaccuratehistogramsareutilizedforthebestmatchesfromthefirstloop.Thecolorsystemonthesecondloopcanbeeitherthesameasonthefirstlooporadifferentone.Forquerieswithone-levelsearchtheselectionofasecondcolorsystemisdisabled.Theusercanselectthenumberofretrie