基于小波变换的SIFT和SURF描述符的分数融合(IJIGSP-V9-N10-3)

I.J.Image,GraphicsandSignalProcessing,2017,10,22-28PublishedOnlineOctober2017inMECS()DOI:10.5815/ijigsp.2017.10.03Copyright©2017MECSI.J.Image,GraphicsandSignalProcessing,2017,10,22-28ScoreFusionofSIFT&SURFDescriptorsforFaceRecognitionUsingWaveletTransformsMusaM.AmeenIshikUniversity/ComputerEngineeringDepartment,Erbil,44001,IraqEmail:[e-mail]rsity/ElectricalandElectronicsEngineeringDepartment,Trabzon,61000,TurkeyEmail:[e-mail]:09June2017;Accepted:12September2017;Published:08October2017Abstract—Automaticfacerecognitionisamajorresearchareaincomputervisionwhichaimstorecognizehumanfacewithouthumanintervention.Significantdevelopmentsinthisfieldhaveshownthatinmanyfacerecognitionapplicationstheautomatedtechniquesoutperformhumans.TheconventionalScale-InvariantFeatureTransform(SIFT)andSpeeded-UpRobustFeatures(SURF)areusedinfacerecognitionwheretheyprovidehighperformances.However,thisperformancecanbeimprovedfurtherbytransformingtheinputintodifferentdomainsbeforeapplyingSIFTandSURFalgorithms.Hence,weapplyDiscreteWaveletTransform(DWT)orGaborWaveletTransform(GWT)attheinputfaceimages,whichprovidesdenserandextrainformationtobeusedbytheconventionalSIFTorSURFalgorithms.MatchingscoresofSIFTorSURFfromeachsubimageisfusedbeforemakingfinaldecision.SimulationsshowthattheproposedapproachesbasedonwavelettransformsusingSIFTorSURFprovidesveryhighperformancecomparedtotheconventionalalgorithms.IndexTerms—Speeded-UpRobustFeatures,Scale-InvariantFeatureTransform,DiscreteWaveletTransform,GaborWaveletTransform.I.INTRODUCTIONFacerecognitionisoneofthemostcommonbiometricsystems.Duetoitshigheracceptabilityrate,researchershavedevelopedvariousalgorithmsforfacerecognitionpurpose.Theprocessofrecognitionusingthesealgorithmshasbeendescribedasadifficulttaskbecauseofthesimilaritynatureorshapesofhumanfaces[1].Despitethedifficultiesencounteredindesigningthesesystems,severalreasonscontributedtotheenormousattentioninautomaticdigitalimageprocessingandvideoprocessingindifferenttypesofapplications,whichincludewideavailabilityofpowerfulandlow-costdesktopandembeddedcomputingsystems.Also,ithasbeendescribedasoneofthebestapplicationsofimageprocessingandanalysis[2].DifferentstatisticalmethodsandalgorithmssuchasPrincipalComponentAnalysisorEigenface(PCA)[3],LocalBinaryPattern(LBP)[4],IndependentComponentAnalysis(ICA)[5],andtriplethalfbandwaveletfilterbank(TWFB)[6]algorithmshavebeendevelopedforfacerecognitionpurposes.In[7]Speed-UpRobustFeature(SURF)andLinearDiscriminantAnalysis(LDA)areusedtoimprovethequalityparametersoffacerecognitionandoptimizingtheresult.Duetocontinuousresearch,asignificantimprovementinrecognitionperformanceisobtainedoveryears[8],[10].Characteristicfacesaremoreeasilyrecognizedthantypicalfaces.Lowfrequencybandscontaininformationthatdeterminesthesexofthespecificsubjects,whilerecognitionofindividualsdependsonthehighfrequencyfeatures.Theglobaldescriptionisdeterminedbythelowfrequency,whilethefinerdescriptionshighfrequencymodulesgivetothefinerinformationrequiredfortheidentificationprocedure[11],[13].Thecoretaskofthispaperworkistoinvestigatehowtherecognitionperformancecanbeenhancedandspeededup.Therefore,imagetransformationapproachisusedasapre-processingstagebeforethefeatureextractionstage.Therestofthepaperisorganizedasfollows;Section2describesmaterials,whilesection3explainsthemethods.Section4showstheresults.Finally,section5includesthediscussionandconclusion.II.MATERIALSThebasicsoffeatureextractionisthedimensionalityreduction,bychoosingsomedominantordistinctfeaturesthatcanbestrepresentsthefaceimagewithlessdistortiontotheoriginalimage.Appropriatealgorithmsareusedtoextractthesalientfeaturesfromtherelevantpatterns.Thefacerepresentationisdoneintwoways:thefirstwayistheappearance(holistic)texturefeaturesandisappliedtothewholefaceimage;thesecondwayisthecomponentbasedwhichutilizesthelinearrelationshipsbetweenthefacialfeaturessuchaseyes,mouth,andnose.Theunpopularcomponent(feature)basedapproachesutilizesomespecialfacialpoints,andcharacterizethembyScoreFusionofSIFT&SURFDescriptorsforFaceRecognitionUsingWaveletTransforms23Copyright©2017MECSI.J.Image,GraphicsandSignalProcessing,2017,10,22-28applyingabankoffilterswhichextractthetypicaltexturearoundthem[10].Theholisticapproachesattractmoreattentionthanthecomponentbasedmethods.Inthispaper,twoofthepopularholisticorappearancebasedmethodsarestudiedbrieflytoextractfeaturesfromfaceimages.A.Scale-InvariantFeatureTransformScale-InvariantFeatureTransform(SIFT)wasdevelopedbyD.Lowe[14].SIFTcandetectandextractdistinctivefeaturesfromdifferentfaceimagestoachieverobustandstablematchingbetweendifferentfaceimagesofthesamesubject(person)withvariousfacialexpressions,faceposes,andthefeaturesextractedfromfaceimagesarescale,illuminationandrotationinvariance.Fig1.showsfourimportantstagesinvolvedfordetectingkeypointsintheSIFTalgorithm.Fig.1.SIFTfeaturesextractionprocess.Intheinitialstage,adifferenceofGaussian(DoG)[8]wasusedtodetectspecificfeaturesandpointswhichar