Target detection (2) Feature extraction of traditional target detection and recognition – Viola Jones detection principle based on Haar-like features

TableofContents

Introduction

integralplot

Trainoptimalweakclassifier

Trainastrongclassifier

strongclassifier

Cascadeclassifier

Introduction

TheViolaJonesdetectorconsistsofthreecoresteps,namelyHaar-likefeatureandintegralmap,Adaboostclassifierandcascadeclassifier.Supposethatduringtargetdetection,suchasub-windowisneededtocontinuouslyslideandmoveintheimagetobedetected.Everytimethesub-windowreachesaposition,thecharacteristicsoftheareawillbecalculated,andthenthetrainedcascadeclassifierwillbeusedtoFeaturesarefiltered,andaslongasthefeaturepassesthescreeningofallstrongclassifiers,theareaisdeterminedtobethetargetarea.

AsshowninFigure9.3,thereare5differentHaar-likefeatureoperators.Assumethatthetotalgrayvalueoftheblackareaineachpictureis?,andthetotalgrayvalueofthewhiteareais.Theresultobtainedis:Haar-likeeigenvaluesofthesub-windowarea.

Figure9.3haarfeatures

Therectangularfeaturecanbelocatedatanypositionintheimagewindow,anditssizecanalsobechangedatwill.Therefore,therectangularfeaturevalueisdeterminedbythethreefactorsofhaarfeatureoperatorcategory,rectangularpositionandrectangularsize.Therefore,changesincategory,size,andposition,sothatsmallerpictureswillalsocontainmanyrectangularfeatures.

Takinga24×24windowasanexample,fivedifferentHaar-likefeatureoperatorsareusedforcalculationinFigure9.3.ThenumberofeigenvaluesofthefiveHaar-likefeatureoperatorsare:43200,43200,27600,27600,20736,atotalof160381.Theimagewindowof24×24sizealonehasmorethan160,000featurevalues.Now,wearefacedwithtwoproblems:

    Facedwithsomanyeigenvalues,howtooptimizecalculationsandreducetheamountofcalculations? Therearetoomanyeigenvalues.Theremustbesomeeigenvaluesthatarebetteratidentifyingpositivesamplesandnegativesamples,andsomethatarenotgoodatdistinguishingpositivesamplesfromnegativesamples.Sample.Sohowtofindthesegood,excellentfeatures,thatis,theoptimalweakclassifier.
IntegralChart

First,tosolvethefirstproblem,youneedtousetheintegralmap.Foragrayscaleimage,thevalueofanypointintheintegralmapisthegrayscaleofallpointsintherectangularareaformedbythispointfromtheupperleftcorneroftheoriginalimage.Thesumofthevalues,asshowninFigure9.4:theleftpictureistheoriginalimage,therightpictureistheintegralimage,the3rdrowand4thcolumnintheintegralimagearethepixelsumoftheyellowboxareaintheoriginalimage,the5thintheintegralimageThesecondcolumnoftherowisthepixelsumofthepurpleboxareaintheoriginalimage.

Figure9.4Integralplot

TheformulashowninFigure9.4canconstructanintegralgraph,butgenerallythefollowingformulaisusedtoconstructanintegralgraph:

    representsthecumulativesumofthepixelvaluesofthefirstjrowoftheimagei,initialized Userepresentsanintegralimage,initialized Scantheimagelinebylineandcalculatetheaccumulationofeachpixelii,jand,thecalculationmethodsareasfollows:

    Scantheimageonce.Whenthepixelinthelowerrightcorneroftheimageisreached,theintegralmapii(i,j)iscompleted.Howtocalculatethesumofpixelsinacertainrectangularareaoftheimagethroughtheintegralmap?

AsshowninFigure9.5:Intheoriginalimage,therearefourareasA,B,C,andD,wherea,b,c,anddrepresentthefourverticesoftheDarea.HowtocalculatetheintegraldiagramofareaD?

Figure9.5Imagearea

Use,,,representa,binFigure9.5,respectively.Theintegralvaluesofthefourpointscandd.Usethefollowingequation9.3tocalculatethetotalgrayvalueofthepixelsinareaD:

Accordingtotheabovetheory,itcanbefoundthataftertheintroductionofintegralmaptechnology,theHaar-likerectangularfeaturevalueofanimageisonlyaffectedbythevalueofitscorrespondingintegralmap,andthechangeofthecoordinatesoftheimage’spositionhasnocorrelationwithitsvalue.Inthisway,thetimetakentocalculatefeaturevaluescanbeshortenedwhencalculatingrectangularfeatures.

Traintheoptimalweakclassifier

Nowlet’ssolvethesecondproblem.Therearetoomanyfeatures,soweneedtoscreenthefeaturesandselecttheoptimalweakclassifier.TheinitialweakclassifiermaybejustabasicHaar-likefeature.ItcalculatestheHaar-likefeaturevalueoftheinputimageandcomparesitwiththefeaturevalueoftheinitialweakclassifiertodeterminewhethertheinputimageisaface.However,thisweakclassifierItistoocrudeandmaynotbebetterthanrandomjudgment.Theincubationofweakclassifiersistotrainweakclassifierstobetheoptimalclassifier.Notethattheoptimalclassifierisnotastrongclassifier,butaweakclassifierwitharelativelylowerror.,trainingaweakclassifierisactuallytheprocessofsettinguptheclassifier.Asforhowtosetuptheclassifierandwhattoset,let’sfirstlookatthemathematicalstructureoftheweakclassifierasshowninEquation9.4:

Theparameterxisthecharacteristicwindow,prepresentsthedifferentdirectionsoftheinequality,withvaluesof1and-1,andgistherectangularwindow.ThecorrespondingHaar-likefeaturevalue,θ,isthejudgmentthresholdoftheweakclassifier.

ThemostbasicweakclassifieronlycontainsoneHaar-likefeature,whichmeansthatthedecisiontreehasonlyonelayer,calledastump.Tocomparethefeaturevaluesoftheinputimageandtheweakclassifierfeatures,athresholdisneeded.Whenthefeaturevalueoftheinputimageisgreaterthanthethreshold,itisdeterminedtobeaface.Theprocessoftrainingtheoptimalweakclassifierisactuallytofindasuitableclassifierthresholdsothattheclassifierhasthelowestjudgmenterrorforallsamples.Thespecificoperationprocessisasfollows:

Findtheoptimalthresholdthatminimizestheclassificationerror.Withthisthreshold,thefirstoptimalweakclassifierisborn,ThebirthofthefirstoptimalweakclassifiermeansthatthebestHaar-likefeaturevaluehasbeenscreenedout.

Trainstrongclassifier

Weakclassificationisonlyaclassifierwithacertainclassificationability,anditsratiocannotmeetthebasicrequirementsforaclassifier,soitneedstobefurtherstrengthenedtobuildastrongclassifierbasedontheweakclassifier.Thespecifictrainingstepsareasfollows:

    Initializethesampletrainingset

Trainingsampleset(x1,y1),…,(xn,yn),whereyi∈0,1,yiisthepositiveandnegativesampleidentifier,1ispositive,otherwiseitisnegative.

    Initializesampleweightsbasedonthenumberofpositiveandnegativesamples

Theinitialweightofthepositivesampleis,theinitialweightofthenegativesampleis,wherem,listhetotalnumberoffacesamplesandthetotalnumberofnon-facesamples,andthetotalnumberofallsamplesisnem>,andsatisfytheconditionofn=m+l,isusedtorepresenttheweight,wheretreferstothetthalgorithmiteration,jreferstothejthoneinthetrainingsample

    Iterativetrainingofweakclassifiers

StrongClassifier

Thefinalstrongclassifierisrecordedas:

Amongthem,indicatesthemaximumnumberoftrainingtimes.

CascadeClassifier

AstrongclassifiercanbeconstructedthroughthetrainingoftheaboveAdaBoostalgorithm,buttheaccuracyofsuchaclassifierstillneedstobeimproved,sothestrongclassifierisnotthefinalstructureoftheclassifier.Inordertomaketheaccuracyanddetectionefficiencyoftheclassifierbetter,youcanUsethecascademethodtocombinestrongclassifiersintoacascadeclassifier.

Figure9.6Cascadeclassifier

Infact,thepurposeoftrainingacascadeclassifieristomakedetectionmoreaccurate.ThedetectionsystemoftheHaarclassifiertakesanimageinrealityasinput,andthenperformsmulti-regionandmulti-scaledetectionontheimage.Theso-calledmulti-Regionistodividetheimageintomultipleblocksanddetecteachblock.

Sincethephotosusedfortrainingareonlysmallimagesofabout20*20,multi-scaledetectionisrequiredforlargefaces.Multi-scaledetectionmechanismsgenerallyhavetwostrategies.Oneisnottochangethesizeofthesearchwindow.However,bycontinuouslyscalingtheimage,thismethodrequiresthecalculationofregionalfeaturevaluesforeachscaledimage,whichisnotefficient.

Anothermethodistocontinuouslyinitializethesearchwindowsizetotheimagesizeduringtraining,continuouslyexpandthesearchwindow,andperformsearches,whichsolvesthedisadvantagesofthefirstmethod.Duringtheprocessofareaenlargement,thesamefacewillbedetectedmultipletimes,whichrequiresthemergingofareas.Nomatterwhichsearchmethodisused,alargenumberofsub-windowswillbeoutputfortheinputimage.Thesesub-windowswillbecontinuouslyfiltered,discardedorpassedbythefirstnodethroughthefilteringcascadeclassifier,asshowninFigure9.6.

UsetheViolaJonesclassifierprovidedbyopencvtodetectfacesandeyes.Theimplementationcodeisasfollows:

importcv2
#Loadclassifier
eye_cascade=cv2.CascadeClassifier(cv2.data.haarcascades,'haarcascade_eye.xml')
img=cv2.imread('image/lenna.bmp')#Loaddetectionimages
gray=cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)
#Performtargetdetectiononimagesthroughtheclassifier
eyes=eye_cascade.detectMultiScale(gray,scaleFactor=1.3,minNeighbors=5)
#Marktheareawheretheeyesare
for(x,y,w,h)ineyes:
img=cv2.rectangle(img,(x,y),(x+w,y+h),(0,255,0),2)
cv2.imshow('demo',img)#Outputdisplay
cv2.waitKey(0)

Figure9.7Runningresults

Asshownabove,opencvalreadycontainsmanytrainedclassifiers,including:faces,eyes,smiles,etc.Thefunctionsusedtocallthetrainedcascadeclassifierinpythonareasfollows:

cv2.CascadeClassifier(cv2.data.haarcascades,’haarcascade_eye.xml’):Theparameterswhencallingmainlyincludethecategoryoftheclassifier.haarcascade_eye.xmlisacascadeclassifierthathasbeentrainedbyopencvtodetecthumaneyes.OthercommonlyusedTheclassificationisasfollowsinTable9.1:

Table9.1CascadeClassifier

Facedetector(default)

haarcascade_frontalface_default.xml

FaceDetector(FastHarr)

haarcascade_frontalface_alt2.xml

Facedetector(sideview)

haarcascade_profileface.xml

Eyedetector(lefteye)

haarcascade_lefteye_2splits.xml

Eyedetector(righteye)

haarcascade_righteye_2splits.xml

mouthdetector

haarcascade_mcs_mouth.xml

nosedetector

haarcascade_mcs_nose.xml

bodydetector

haarcascade_fullbody.xml

detectMultiScale(image,scaleFactor=1.1,minNeighbors=3):whereimageisthegrayscaleimagetobedetected,scaleFactoristheproportioncoefficientofthesearchwindowintwoconsecutivescans,thedefaultis1.1,thatis,thesearchwindowisexpandedby10%eachtime,minNeighborsistheminimumnumberofadjacentrectanglesthatmakeupthedetectiontarget.Ifthesumofthenumberofsmallrectanglesthatmakeupthedetectiontargetislessthanminneighbors-1,theywillbeexcluded.

Theknowledgepointsofthearticlematchtheofficialknowledgefiles,andyoucanfurtherlearnrelatedknowledge.OpenCVskilltreeHomepageOverview23957peoplearelearningthesystem