歷求職職場(chǎng)實(shí)用文檔英文翻譯姓名：張衡學(xué)號(hào)：1102080234指導(dǎo)教師：孫中橋?qū)I(yè)：信息管理與信息系統(tǒng)班級(jí)：2011級(jí)時(shí)間：2015年6月25日信息管理系信息管理系英文翻譯評(píng)價(jià)表學(xué)生姓名張衡性別男學(xué)號(hào)1102080234外文文獻(xiàn)標(biāo)題FastComponent-BasedQRCodeDetectioninArbitrarilyAcquiredImages外文文獻(xiàn)出處以下內(nèi)容由指導(dǎo)教師填寫(xiě)（打勾“√”選擇）評(píng)價(jià)項(xiàng)目評(píng)價(jià)結(jié)論打勾評(píng)價(jià)結(jié)論打勾評(píng)價(jià)結(jié)論打勾是否外文期刊文獻(xiàn)是否與本人論文相關(guān)完全相關(guān)一般不相關(guān)翻譯工作量超負(fù)荷飽和不飽和翻譯態(tài)度認(rèn)真一般不認(rèn)真翻譯進(jìn)度按計(jì)劃執(zhí)行一般未按計(jì)劃執(zhí)行翻譯訓(xùn)練效果優(yōu)良中差綜合評(píng)語(yǔ)（是否完成了規(guī)定任務(wù)、效果是否符合要求等）指導(dǎo)教師簽名：2015年4月25日制表：李鐵治注1：此表與翻譯文本一起裝訂；注2：為了加強(qiáng)學(xué)生外語(yǔ)應(yīng)用能力的訓(xùn)練，每位同學(xué)至少選擇畢業(yè)論文中一篇外文參考文獻(xiàn)（10000袋即可。英文原文FastComponent-BasedQRCodeDetectioninArbitrarilyAcquiredImagesAbstractQuickResponse(QR)codesareatypeof2Dbar-codethatisbecomingverypopular,withseveralapplicationpossibilities.Sincetheycanencodealphanumericcharac-ters,arichsetofinformationcanbemadeavailablethroughencodedURLaddresses.Inparticular,QRcodescouldbeusedtoaidvisuallyimpairedandblindpeopletoaccesswebbasedvoiceinformationsystemsandservices,andau-tonomousrobotstoacquirecontext-relevantinformation.However,inordertobedecoded,QRcodesneedtobeprop-erlyframed,somethingthatrobots,visuallyimpairedandblindpeoplewillnotbeabletodoeasilywithoutguid-ance.Therefore,anyapplicationthataimsassistingrobotsorvisuallyimpairedpeoplemusthavethecapabilitytode-tectQRcodesandguidethemtoproperlyframethecode.Afastcomponent-basedtwo-stageapproachfordetectingQRcodesinarbitrarilyacquiredimagesisproposedinthiswork.Inthefirststage,regularcomponentspresentatthreecornersofthecodearedetected,andinthesecondstagege-ometricalrestrictionsamongdetectedcomponentsareveri-fiedtoconfirmthepresenceofacode.Experimentalresultsshowahighdetectionrate,superiorto90%,atafastspeedcompatiblewithreal-timeapplications.KeywordsQRcode?Component-baseddetection?Haar-likefeatures?Cascadeclassifier1IntroductionComparedtotraditional1Dbarcodes,2Dbarcodescanen-codealargeramountofdata,includingalphanumericchar-acters.QRcode,whichstandsforQuickResponseCode,isatypeoftwo-dimensionalcodeintroducedbyDensoWavein1994[12].Theyhavebeendesignedtobeeasilyfoundandtohaveitssizeandorientationdeterminedunderbadimagingconditions.Inaddition,ISO/IEC18004specifiesanerrorcorrectionschemethatcanrecoveratmost30%ofoccludedordamagedsymbolarea.ThesefeaturesmaketheQRcodeanextremelywellsucceededtechnologyintheareaofbarcodes.Figure1showssomeexamplesofQRcodes.Fig.1SamplesofQRcodeQRcodeswereinitiallyusedbytheautomotiveindustrytotrackvehiclepartsduringthemanufacturingprocess[12].Nowadays,QRcodesaremostcommonlyusedas“physicalhyperlinks”(encodedhyperlinks),notablyintheadvertisingindustry,toconnectplacesandobjectstowebsitescontain-ingadditionalcontextrelevantinformation.Applicationsineducationandentertainment[8,23,25],dataandsystemse-curity[9,16,19],specificserviceoffer[7,27],amongothersarealsoemerging.Anotherpossibleapplicationconsistsinhelpingvisuallyimpairedandblindpeople,orevenrobots,inseveraltaskssuchasindoornavigation,shopping,read-ing,andmuchmore[2,11,13].Theexistingdecoders,easilyfoundformobiledevices,areabletoworkcorrectlyonlyifcodesareproperlyframed,withcoderegioncorrespondingtoatleast30%oftheim-age.Whenexploringanenvironment,visuallyimpairedorrobotswillnotbeabletocapturesuchimagesunlesstheyaretoldwherethosecodesarelocated.Thus,inordertomakeusefulapplicationsforthemviable,detectingthepresenceofacodeinanimageisanecessarysteppriortothedecod-ingprocess.Inrelatedliterature,themajorityofworkthatmentionQRcoderecognitionordetectionisactuallyconcernedwithimprovingimagequalityordeterminingtheexactcontoursofthecoderatherthanfinding(i.e,decidingwhetherthereisorthereisnot)aQRcodesymbolinanimage[10,20,22].Infact,mostoftheimagesconsideredinthoseworksareimagesacquiredwiththespecificintentofcapturingonlythecodesymbol.SomefewworksthatdealwiththeproblemoffindingQRcodesproposesolutionsthatrelyonauxiliaryinforma-tionsuchasvisualcuesorRFIDtags[11,28].Althoughsuchapproachispossibleincontrolledenvironments,itisnotpracticalingeneralcontexts.ThisworkaddressestheproblemofdetectingQRcodesinarbitrarilyacquiredimageswithoutrelyingonauxiliarycues.Theaimisnotonlytodetectthepresenceofcodesymbols,butalsotodelimittheirpositionwithinanimageasaccuratelyaspossible.Thatwouldallow,additionally,toinstructanuserorrobottoproperlyapproachthecameratowardsthecode.Tothatend,acomponent-based,two-stagedetectionap-proachisproposed.Inthefirststage,asquarepattern,calledfinderpattern(FIP),locatedatthreecornersofanyQRcodeisdetectedusingacascadedclassifiertrainedaccordingtotherapidobjectdetectionmethodproposedin[26].Inthesecondstage,geometricalrelationshipsamongdetectedcan-didatefinderpatternsareanalyzedinordertoverifyiftherearesubgroupsofthreeofthemspatiallyarrangedasthreecornersofasquareregion.Areportonpreliminaryresultsofthisapproachhasbeenpreviouslypublishedin[3].Inthispaper,areformulatedreportoftheapproachandanextendedsetofresultsarepresented.Inparticular,adetailedaccountofthesecondstage,includingaformaldescriptionofthecomponentaggregationalgorithm,newresultsobtainedwithalargertrainingandtestsets,discussionsconcerningaddi-tionalparametersthathavenotbeenexaminedintheprevi-ouswork,andquantitativeresults,notpresentedbefore,onQRcodedetectionperformancearepresented.Thistextisorganizedasfollows.InSect.2,importantstructuralaspectsofQRcodeswhichwillbeexploredintheproposedapproacharefirstdescribed.Then,anoverviewoftheproposedtwo-stageapproachispresented,togetherwithabriefreviewonViola-Jonesmethodforrapidobjectde-tection,usedinthefirststageoftheapproach.InSect.3,theproblemofaggregatinginformationobtainedinthefirststageofthedetectionprocessisexpressedingraph-basedformalismandasimplealgorithmtosolvetheproblemisproposed.TheproceduretodetermineappropriatevaluesforthetrainingparametersoftheFIPdetector,inanOpenCVimplementationofViola-Jones’method,isdiscussedanddescribedinSect.4.Then,inSect.5experimentalresultstoassessinfluenceofsomeparametersinFIPaswellasQRcodedetection,andframeratesachievedinvideoprocess-ingarereported.Althoughextensiveexperimentalvariationwascarriedout,notallpossibilitiesarisingfromparametervariationswereevaluated.Nevertheless,QRcodedetectionratesuperiorto90%wasobservedontestimages,suggest-ingthatevenbetterratescanbeachieved.Finally,inSect.6asummaryofthemaincontributionsofthisworkandsomeissuesforfutureinvestigationarelisted.2Component-BasedApproachforQRCodeDetectionInthissection,somestructuralaspectsofQRcodesanddetectionrequirementsinareal-timeapplicationscenariothathavemotivatedtheproposedtwo-stageapproacharefirstdescribed.Then,anoverviewoftheproposedapproachfollowedbyabriefreviewonViola-Jones’methodforobjectdetectionispresented.2.1QRCodeAQRcodesymbol,accordingtotheISO/IECstandard18004,hasageneralstructurethatcomprisesdata,versioninformation,errorcorrectioncodewords,andthefollowingregionsillustratedinFig.2:Fig.2QRcodestructure(FIP:FinderPattern/TP:TimingPattern/AP:AlignmentPattern)–aquietzonearoundthesymbol,–threefinderpatterns(FIP)inthecorners,–twotimingpatterns(TP)betweenthefinderpatterns,and–acertainnumberofalignmentpatterns(AP)insidethedataareaTheinternalorganizationofacodewithrespecttoitsstructuralcomponentsmayvarydependingontheamountofdataencoded.Thesmallestunits(blackorwhitesquares)thatcomposeaQRcodearecalledmodules.Therearevariousversionsofthesymbols,from1to40,havingdistinctinformationstoragecapabilitiesthataredeterminedbypredefinednumberofmodulesinsymbolarea.Figure3showsanexampleofaversion1code,highlightingthenumberofmodules.Figure4showsadditionalexamplesthatillustratetherelationbetweennumberofmodulesandnumberofencodedcharacters.Notethatthealignmentpatternsoccurmoretimesinlargerversionsofthecode.Fig.3QRcodeversion1iscomposedof21×21modulesFig.4DifferentversionsofQRcodesymbolsandrespectivenumberofencodedalphanumericcharactersThereisavisuallydistinctivepatternatthreecornersofQRcodes,knownasfinderpattern(FIP).FIPshavebeenspeciallydesignedtobefoundinanysearchdirectionasthesequenceofblack(b)andwhite(w)pixelsalonganyscanlinethatpassesthroughitscenterpreservethespecialsequenceandsizeratiow:b:w:bbb:w:b:w,ascanbeseeninFig.5.Fig.5Blackandwhitepixelproportionsinfinderpatternsfordiagonal(d),vertical(v)andhorizontal(h)scanlines2.2DetectionRequirementsOneofthefirstrequirementsconsideringreal-timeapplicationsisfastdetection.Moreover,sincethereisnopriorknowledgeaboutthesizeofacode,anotherimportantfeatureofthedetectorisscaleinvariance.Inaddition,inordertosucceedinuncontrolledenvironments,thedetectorshouldbealsorobustwithrespecttovariationinillumination,rotation,andperspectivedistortion.Inordertoestablishsomelimitstothedetectiontask,inthisworkattentionispaidtoscaleinvarianceandrobustnesswithrespecttovariationinilluminationcondition,whilerotationsandperspectivedistortionsareassumedtobesmall.Furtherthanthat,itisassumedthatQRcodesareprintedinblackandwhite(oratleastdarkforegroundandlightbackground).Colorisnotanissuesinceallprocessingconsidersgraytoneimages.2.3OverviewoftheProposedApproachDirectdetectionofthewholecodemaynotbeaneasytaskbecausetheinternalstructuralorganizationmayvaryfromcodetocode,dependingoncodeversion.Moreover,extractionoffeaturesofthecodearea,suchashistogramsorfrequencyinformation,canbegreatlyaffectedbyimageresolution,codescale,ornoise.Forinstance,algorithmsforrecognitionofQRcodesfrequentlyrelyontheregularityofscanlinesovertheFIPs(seeFig.5)tofindthemandthendeterminetheexactcontoursofthecode.However,thatregularitymaybedisruptedinlowqualityimages.Thus,inordertodealwithunconstrainedimages,theprocessingalgorithmshouldnotdependonfinedetails.Ontheotherhand,FIPsaretypicallythelargeststructureswithinacode,haveafixedshape,andappearinexactlythreecornersofallcodes.ThesecharacteristicsofFIPssuggestacomponent-baseddetectionapproachtotheproblemofdetectingQRcodes.Themainideaofcomponent-baseddetection[21]isthedetectionofpartsoftheobjectandthentheanalysisofdetectedpartsinordertofindacoherentarrangementformingtheintegralobject.Arrangementofthepartscantakeintoconsiderationgeometricalrestrictions.Whentherelationbetweenpartscannotbeeasilystated,machinelearningbasedapproachescanbeusedasin[1,14].Takingtheaboveobservationsintoconsideration,theproposedapproachfordetectingQRcodesconsistsofatwostagemethod.InthefirststagetheaimistodetectFIPs,andinthesecondstagethegoalistointegrateinformationaboutdetectedFIPsanddecidewhethersubsetsofthreeofthemcorrespondtoaQRcodeornot.Inthefirststage,itisdesirabletomaximizetruepositives(TP)whilemaintainingacontrollednumberoffalsepositives(FP).Thus,forthedetectionofFIPs,somemechanismtocontrolTPandFPshouldbeavailable,andthedetectionprocessshouldfulfilltherequirementslistedabove.Inparticular,itshouldbefastandinvarianttoscaleandvariationofilluminationconditions.Tomeettheserequirements,Viola-Jones’rapidobjectdetectionmethod[26]isproposedforthefirststage.Inthesecondstage,thealgorithmshouldbeabletofindsetsofthreeFIPsthatarethecornersofasameQRcode.Tothatend,geometricalrestrictionsaswellassizerelatedrestrictionsareconsidered,asdetailedinSect.3.2.4AReviewonViola-Jones’ObjectDetectionFrameworkViola-Jones’objectdetectionframeworkhasthefollowingcharacteristics:–itusessimplefeatures(Haar-like);–itperformsfastcomputationoffeaturesusingintegralimage;–itisbasedonacascadedapproachtotrainaclassifier:eachstageistrainedinsuchawayastoachievefixedTPandfalsealarm(FA)rates,usingaboostingalgorithm;–itisveryfastinpracticesincefeaturesarecomputedrapidlyandthecascadeisdesignedtodiscardthemajorityofnegativesamplesinitsearlystages,eliminatingtheneedforcalculatingresponsesofallstagestoeverysampleconsidered.TheViola-Jonesframeworkcanbeillustratedwiththefollowingconcisedescription:ThedetectionprocessconsistsinacquiringacompletesamplesetfromtheinputimageandsubmittingeachsampletoacascadeclassifierwhosestageshavebeentrainedbyaboostingschemethataggregatesweakclassifiersmadeofHaar-likefeatures.Thesefeaturesarecalculatedinconstanttimeusingamatrixcalledintegralimage.Belowsomekeyconceptsarereviewedandthedetectionprocessisexplained.SamplesetAsampleconsistsofasub-regionoftheimagerestrictedtoawindowandthecompletesamplesetoftheimageisobtainedbyslidingthewindowontheimageandvaryingitssizefromagivenminimumtothelargestpossiblesizeintheimage.CascadeclassifierAcascadeconsistsofaseriesofconsecutiveclassifiers(stages)trainedtorejectsamplesthatdonotmatchthesearchedpatternwhileacceptingthemajorityofpositivesamplesandpassingthemtothenextstage.Asampleissaidtobedetectedwhenitisacceptedfromthefirsttothelastcascadestages,withoutrejection.Eachstageisbuiltfromasetofweakclassifiersaggregatedinacommitteebyaboostingalgorithmandcanbeseenasanindependentclassifierdesignedtoobtainaveryhighhitrate(typically99%ormore)withanacceptablefalsealarmrate(typicallybetween40%and60%).Figure6illustratestheconceptofacascadeclassifier.Fig.6IllustrationofthecascadeapproachtodetectionBoostingBoostingworksinrounds,iterativelytrainingasetofweakclassifierswhilereweightingthetrainingsamplessothat“hard”sampleswillhaveincreasedrelativeimportanceintheset.Ineachroundthebestweakclassifierisselectedtocomposetheresultingclassifier[15].Thus,boostingcanachievethespecifiedhitrateandfalsealarmrateasitincreasestheweakclassifiersinthecombination(aslongasthefeaturesusedhaveenoughrepresentationalpower).EveryweakclassifieristypicallytheresultofthecomputationofaHaar-likefeaturefollowedbyabinarythreshold,althoughtheycanhavemoresophisticatedformslikesimpletrees.Haar-likefeaturesThefeaturesusedbytheclassifiersproposedin[26],inspiredfrom[24],arebasedonfeatureprototypesshowninFig.7.In[18],thissethasbeenextendedwithadditional45degreerotatedprototypes,showninFig.8.Fig.7BasicfeatureprototypesetFig.8AdditionalrotatedfeatureprototypesintheextendedsetGivenasquarewindow,allfeaturesusedarederivedfromtheseprototypesbydefiningitswidth,heightandrelativepositionwithinthewindow.SomeexamplesareshowninFig.9.Note,forinstance,thatfeaturesinFig.9(a)andFig.9(b)arebothbasedonthesameprototype.Fig.9ExamplesofHaar-likefeaturesFeaturevaluesarecomputedwithrespecttoasamplethatcorrespondstoasub-regionoftheimageunderaslidingevaluationwindow.Bydefinition,thevalueofafeatureforeachsampleisthesumofthepixelvaluesinthewhiterectangleareasubtractedfromthecorrespondingsummationintheblackrectanglearea.3AggregationalgorithmOnceFIPcandidatesaredetectedinthefirststage,arrangementofthreecandidateFIPsthatarelikelytocorrespondtothecornersofaQRcodemustbeexaminedinthesecondstage.Inthissection,analgorithmthatconsiderssize,distanceandanglerestrictionstoperformsuchexaminationisproposed.Notethatbesidesthesegeometricalrestrictions,additionalinformationsuchasthenumberofoverlappingdetectionsortexturearoundtheFIPscouldbealsousedtoruleoutsomesubsetsorsupportothers.However,takingintoconsiderationtherequirementforfastprocessing,onlysize,distance,andangleinformationareconsidered.TheproposedsolutionforaggregatingFIPcandidatesispresentedinAlgorithm1.Lines1to3correspondtovertexsetcreation.Lines4to13correspondtothecreationofedges,wheneverapairofverticessatisfythesizeandthedistancecriteria,withrespectivetolerances.Theloopstructurefromline14to23isthemainpartofthealgorithm,wherecliquesofsizethreewhoseedgespairwisesatisfytheorientationcriterionarefound.Foreachedge(u,v),alledges(u,v)thatareadjacenttooneofitsextremities(u)areverified.Foreachofsuchpairsofedges(i.e.,{(u,v),(u,v)}),firsttheexistenceofthethirdedge(i.e.,edge(v,v))isverified.Ifthethirdedgeisnotthere,thatmeansthateitherthetwovertices,vandvdonotsatisfythesizeordistancecriteria,orthatthethirdedgehasalreadybeenprocessed(inthiscase,thetriplet{u,v,v}isalreadyintheoutputlist).Ifthethirdedge(v,v)ispresent,andifthefirsttwosatisfytheorientationcondition,thenthesubsetofthethreevertices{u,v,v}isaddedtotheoutputlist.Notethatthereisnoneedtoexamineedgesthatareadjacenttotheotherextremity(v)sincethosewillbeevaluatedlater.Thealgorithmendswhenalledgesareprocessed.4ProcedureforTrainingandEvaluatingFIPDetectorsClassifiertrainingandevaluationtoolsthatimplementtheViola-Jones’methodareavailablerespectivelyasopencv-haartrainingandopencv-performanceutilitiesinOpenCV2.0[4].Forthetrainingofacascadedclassifier,asetofpositivesamples(croppedimageregionscontainingonlyatargetobject)andnegativebackgroundimages(withnotargetobjects)shouldbeprovided.Foreachstage,thetrainingalgorithmrandomlyselectsnegativesamplesfromthebackgroundimagesthathavebeenmisclassified(aspositives)bytheprecedingstages.AnotherutilityavailableinOpenCVisopencv-createsamples,thatallowsapplicationoftransformationsonpositivesampleswithcontrolledrandomvariationsinintensity,rotationandperspectivedistortion.ExamplesoftransformedsamplesobtainedfromapositivesampleareshowninFig.13.Fig.13Examplesofsamplesgeneratedbyrandomtransformationsonapositivesample4.1MainTrainingParametersThetrainingprocessimplementedinopencv-haartraininginvolvesseveralparametersthatneedtobetunedforeachapplication.Amongthem,thefollowingparametershavebeenconsideredinthiswork:–Featureset:ItcanbethebasicsetshowninFig.7(MODE=Basic)orthebasicsetinconjunctionwiththeextendedsetshowninFig.8(MODE=Extended/All).–Symmetry:Whenthetargetpatternissymmetrical,thetrainingalgorithmcanberestrictedtoconsideronlyhalfpartofthepositivesamples(SYM=Symmetric/Y)inordertoreduceprocessingtimeduringtraining,ornot(SYM=Asymmetric/N).–Classifiertopology:Ratherthancascadedstages(MTS=Cascade),itispossibletoallowsplits(MTS=Tree)thatturnstheclassifierintoasimpleCARTtree[5].–Weakclassifiernumberofsplits:AweakclassifierinitssimplestformisjustasingleHaar-likefeatureandabinarythreshold.Theyarecombinedwithotherweakclassifiersbytheboostingalgorithmtoformastrongclassifiercorrespondingtoonestageofthecascade.Itispossibletoallowweakclassifierstolearnmorecomplexrelationshipsinthetrainingpatternbylettingthemtobenotjustasinglefeature(NS=1)butasimpleCARTtreewithalimitedsmallnumberofsplits(NS>1).Forfacedetection,empiricalobservations[17]indicatethatthereisanincreaseincascadeperformancewhensplitsintheweakclassifiersareallowed.–Maximumfalsealarmrate/Minimumhitrate:Eachcascadestagemustcomplywithamaximumfalsealarmrate(FA)andwithaminimumhitrate(HR)forthesamplessuppliedtoitinordertobeconsideredtrained.Atoolowfalsealarmraterequirementmaycausethestagetobecomeoverlycomplexdiminishingthebenefitsofthecascadeapproach.Averyhighminimumhitraterestrictionmayhaveasimilareffect.–Numberoftrainingsamples:Thenumberofpositivetrainingsamples(SAMPLES).–Sizeoftrainingsamples:Alltrainingsamplesareresizedtoafixeddimension(SIZE).Forfacedetection,ithasbeenobservedthat20×20isanappropriatesize[17].4.2DetectionEvaluationMetricsAclassifiertrainedusingtheopencv-haartrainingutilitycanbeevaluatedusingtheopencv-performanceutility.Theperformanceevaluationutilityrequirestestimageswithground-truthdata(i.e.,positionsofthetargetsamples).Suchimagescanbeartificiallygeneratedbytheopencv-createsamplesutility,thatinsertsonetargetsampleineachbackgroundimage.This,inconjunctiontoopencv-performance,isusefulforautomatingtheperformancetests,byprovidingtothematestsetofpositivesamplesandnegativebackgroundimagesastheonesprovidedfortraining.Figure14showsanexampleoftargetsampleinsertedintoabackgroundimage.Fig.14AbackgroundimagewithaFIPinsertedarounditscentralpositionScalingfactorForeachpositionoftheimage,severalsamplesareobtainedbyvaryingthewindowsizefromaminimumsizetothemaximumpossiblesizeinthatposition.Sinceallsamplesareresizedtothesizeofthedetector,thisminimumsizeshouldbetheSIZEofthedetector(i.e.thesizeofthesamplesusedfortrainingthedetector).Thescalingfactor(SF)determineshowwindowsizevariesforsampling.Forinstance,SF=1.10meansthatthesizeoftheprocessingwindowwillbeincreasedin10%ateachsamplingroundinasameposition.Thelargerthescalefactor,lesssamplesareconsideredineachposition.Notethattheclassifiercanbeconsideredscaleinvariantduetothisresizingofallsamplestoafixedsize.MinimumoverlappingdetectionsOftenmorethanonesampleamongthosecorrespondingtosmallshiftsofthewindowaroundthetargetobjectresultinpositivedetection.Insuchcases,itisconvenienttoestablishsomecriteriainordertoconsolidatemultipleredundantdetectionsintoasingledetectionrepresentingallofthem.Naturally,thelargerthenumberofdetectionsthatoverlapinagivenneighborhood,thelargertheevidencethatthereisatargetobjectinthatposition.Therefore,aminimumnumberofoverlaps(ND)inthesetofoverlappingdetectionscanbeestablishedasdeterminingasingledetection,thatis,asetofoverlappingdetectionsisconsideredadetectionifandonlyifthenumberofdetectionsinthesetisatleastND.Intheimplementationusedinthiswork,twodetectionsareconsideredoverlappingiftheirpositioninxandinycoordinatesdifferbynomorethan20%ofthewidthofthesmallerdetectedsampleandtheirwidthsdifferbynomorethan20%.4.3EstimationofTrainingParameterValuesViola-Jonesframeworkiswidelyknownandverypopularforfacedetectionandtheliteratureconcerningthetrainingofthecascadeclassifieroftenrecommendstrainingpractices,parametersandsets,suitableforthatapplicationdomain[6].SinceFIPsaresimplerandmuchdifferentfromfacepatterns,variationofparametervaluesforthecascadetrainingofFIPdetectorsisasubjectofstudyinthiswork.Duetothelargenumberofparametersinthetrainingprocess,anexhaustiveassessmentofpossiblecombinationsofparametervaluesisnotfeasible.TheapproachfordeterminingtheparametervaluesusedinthisworkforthecascadetrainingofaFIPdetectorissimilartotheoneusedin[17]forfacedetection.Themethodologyconsistsinfirstestablishingasequentialorderaswellasinitialvaluesfortheparameterstobeevaluated.Then,followingtheestablishedorder,theeffectofindividualparametervariationisassessedwhilethevaluesofallotherparametersarekeptfixed.Afterassessmentofaparameter,itsinitialvalueisreplacedbythebestvaluefound,andtheassessmentproceedstothenextparameter.Althoughthisanalysiscannotbeconsideredexhaustive,itprovidesaverygoodindicationoftheparametervariationinfluenceinfinaldetectionperformancewhileavoidingthecombinatorialexplosionthatwouldarisefromexperimentingallpossibilities.ParametervaluesfortrainingFIPdetectorswerechosenfollowingthestrategydescribedabove.Abasesetconsistingof380FIPsamplescroppedfromavarietyofdifferentimages,andasetof1500backgroundimages(withoutQRcodes)forthenegativesamplesweredividedintotrainingandtestsets.Allpositivesamplesfortrainingweregeneratedfromthe285FIPsinthebaseset,applyingtransformationsusingtheopencv-createsamplesutilitywithintensityvariationlimitedto±80androtationlimitedinx,yto±23?andinz(perspective)to±34?.Thetestsamplesweregeneratedfromtheremaining95FIPsapplyingthesamevariationinintensityandinxandy,andinzlimitedto±23?.Thebackgroundimagesweresplitintotwosubsetsofequalsize(750imageseach),beingusedrespectivelyfortrainingandtesting.InitialvaluesofparametersweresettoHR=0.998,ST=15,SYM=N(Asymmetric),MTS=Cascade,NS=2,FA=0.5,SAMPLES=4000(positivesandnegatives,4000each),andSIZE=20×20.RecallandFPshownforeachtrainingisrelativetothetestset,computedasdescribedinSect.4.2usingdetectorapplicationparametersSF=1.10andND=1.Darknodesindicatethevalueschosenfortherespectiveparameters.BeyondconsideringjustrecallandFP,thechoiceofthebestparametervaluesweremadetakingintoaccountthesimplicityoftheresultingclassifierandthebalanceofitscascade(i.e.uniformityoffeaturequantityamongconsecutivestages).Moreover,recallvalueswereprioritizedoverFPbecausewhiletheFIPaggregationalgorithmcannotrecoverundetectedFIPs,itdoescandisregardfalseFIPdetections.5ExperimentalResultsInthissection,experimentalresultsregardingFIPandQRcodedetect