You are on page 1of 4

iFACE iMASK iDETECTION i

iShubham iPawar i Ananya iDey i Deepdev ivyas i Gaurang igupta i


iSELECT iEEE iSELECT iEEE iSELECT iEEE iSELECT iEEE

VIT iVELLOE VIT iVELLOE VIT iVELLOE VIT iVELLOE


INDIA INDIA INDIA INDIA
ishubham.2018@vitstudent.ac.in ananya.dey2018@vitstudent.ac.in deepdev.vyas2018@vitstudent.ac gaurang.gupta2018@vitstudent.a
.in c.in
i

Abstract—The icorona ivirus iCOVID-19 ipandemic iis imasks. iThe imodel iis iintegration ibetween ideep ilearning
icausing ia iworld ihealth icrisis itherefore ithe ieffective iand iclassical imachine ilearning itechniques iwith iOpenCV,
iprotection imethods iis iwearing ia iface imask ipublicly iareas ipytorch, itensor iflow iand ikeras. iwe've igot iused ideep
iper ithe iworld iHealth iOrganization i(WHO). iThe iCOVID-19 itransfer ilearning ifor ifeature iWe iachieved ithe ivery ibest
ipandemic iforced igovernments iacross ithe iglobe ito iimpose
iaccuracy iand iconsumed ithe ismallest iamount itime iwithin
ilockdowns ito istop iVirus itransmissions. iReports iindicate ithat
iwearing ifacemasks iwhile iat iwork iclearly ireduces ithe ichance ithe iprocess iof itraining iand idetection. iMachine iLearning:
iof itransmission. iAn iefficient iand ieconomic iapproach iof iMachine ilearning iis ian iapplication iof iAI i(AI) ithat ihas
iusing iAI ito imake ia isecure ienvironment iin ia ivery isystems ithe iflexibility ito iautomatically ilearn iand iimprove
imanufacturing isetup. iA ihybrid imodel iusing ideep iand ifrom iexperience iwithout ibeing iexplicitly iprogrammed.
iclassical imachine ilearning ifor iface imask idetection iare igoing iMachine ilearning ifocuses ion ithe ievent iof icomputer
ito ibe ipresented. iA imask idetection idataset iconsists iof iwith iprograms iwhich iwill iaccess idata iand iuse iit ilearn ifor
imask iand iwithout imask iimages, iwe iare ivisiting iuse iOpenCV ithemselves
ito itry iand ido ireal-time iface idetection ifrom ia ilive istream . iMachine ilearning ialgorithms iare iemployed iin ia igood
ivia iour iwebcam. iwe iare igoing ito iuse ithe idataset ito icreate
iform iof iapplications, ilike iemail ifiltering iand icomputer
ia iCOVID-19 iface imask idetector iwith icomputer ivision iusing
iPython, iOpenCV, iPytorch, iTensor iFlow iand iKeras. iOur ivision, iwhere iit's idifficult ior iinfeasible ito idevelop
igoal iis ito ispot iwhether ithe iperson ion iimage/video istream iis iconventional ialgorithms ito iperform ithe ineeded itasks.
iwearing ia imask ior inot iwith ithe iassistance iof icomputer iMachine ilearning iis iclosely iassociated iwith icomputational
ivision iand ideep ilearning.. i(Abstract) i istatistics, iwhich ifocuses ion imaking ipredictions iusing
icomputers. i
Keywords—corona ivirus, icovid, iPython, iMask, iArtificial Machine ilearning iapproaches iare itraditionally idivided iinto
iintelligence. ithree ibroad icategories: i
I. INTRODUCTION I • iSupervised ilearning: ithe ipc iis ipresented iwith iexample
iinputs iand itheir idesired ioutputs, igiven iby ia i"teacher",
We iare iwithin ithe imidst iof ia iworldwide ipandemic, iwhich iand itherefore ithe igoal iis ito ibe itold ia igeneral irule ithat
ihas icaused imajor ilifestyle ichanges. iIt's ibeen imade imaps iinputs ito ioutputs. i• iUnsupervised ilearning: iNo
imandatory ito iwear ifacemasks iwhen ione isteps imove iinto ilabels iare igiven ito ithe itraining ialgorithm, ileaving iit ion
ipublic. iRegardless, iwe idiscover ithat iseveral ipeople iits iown ito iseek iout istructure iin iits iinput. iUnsupervised
idisregard irules iand irefuse ito iwear ia imask ithereby iputting ilearning iwill ibe ia igoal iin iitself i(discovering ihidden
ithemselves iand ialso ithe ilives iof iothers iin iperil. iProper ipatterns iin idata) ior ia iway itowards ian iend i(feature
ifacemask iwearing imay ibe ieffective iin icontrolling ithe ilearning). i• iReinforcement ilearning: iA icomputer iprogram
ispread iof ithe idisease. iit's iimpossible ifor ihumans ito iinteracts iwith ia idynamic ienvironment iduring iwhich iit
ieffectively imonitor ieveryone. iover ifive imillion icases iwere imust iperform ia ispecific igoal i(such ias idriving ia ivehicle
iinfected iby iCOVID-19 iin iless ithan i6 imonths iacross i188 ior iplaying ia igame iagainst ian iopponent). ibecause iit
icountries. iThe ivirus ispreads ithrough iclose icontact iand iin inavigates iits iproblem ispace, ithe iprogram iis iprovided
icrowded iand iovercrowded iareas. iHere iwe iintroduce ia ifeedback ithat's ianalogous ito irewards, iwhich iit itries ito
imask iface idetection imodel ithat's isupported icomputer imaximize i6 iDeep iLearning: iDeep ilearning imethods iaim
ivision iand ideep ilearning. iThe iproposed imodel iwill ibe iat ilearning ifeature ihierarchies iwith ifeatures ifrom ihigher
iintegrated iwith isurveillance icameras ito iimpede ithe ilevels iof ithe ihierarchy iformed iby ithe icomposition iof
iCOVID-19 itransmission iby iallowing ithe idetection iof ilower-level ifeatures. iAutomatically ilearning ifeatures iat
iindividuals iwho iare iwearing imasks inot iwearing iface imultiple ilevels iof iabstraction iallow ia isystem ito ifind iout
icomplex ifunctions imapping ithe iinput ito ithe ioutput imitigate ithe iimpact iof iCOVID-19 ioutbreak ihas ialso
idirectly ifrom idata, iwithout idepending icompletely ion ibeen iproposed.[1]
ihuman-crafted ifeatures. iDeep ilearning iallows
icomputational imodels ithat iare icomposed iof imultiple Guanjun iGuo iet ial i[5] iproposed ia ifast iface idetection
iprocessing ilayers ito ifind iout irepresentations iof iknowledge imethod ibased ion idiscriminative icomplete ifeatures
iwith imultiple ilevels iof iabstraction. iPyTorch iis ian iopen- i(DCFs) iextracted iby ian ielaborately idesigned
source imachine ilearning ilibrary isupported ithe iTorch iconvolutional ineural inetwork, iwhere iface idetection iis
ilibrary, iused ifor iapplications ilike icomputer ivision iand idirectly iperformed ion ithe icomplete ifeature imaps ias
inatural ilanguage iprocessing, iprimarily ideveloped iby iDCFs ihave ishown ithe iability iof iscale iinvariance, iwhich
iFacebook's iAI iscience ilab. iOpen iCV: iOpenCV i(Open- iis ibeneficial ifor iface idetection iwith ihigh ispeed iand
Source iComputer iVision iLibrary) iis ian iopen-source ipromising iperformance.
icomputer ivision iand imachine ilearning isoftware ilibrary.
iOpenCV iwas ibuilt ito isupply ia istandard iinfrastructure ifor Walid, iHariri i[6] iproposed ia ireliable imethod ibased ion
icomputer ivision iapplications iand ito iaccelerate ithe idiscard imasked iregion iand ideep ilearning-based ifeatures
iutilization iof imachine iperception iwithin ithe icommercial iin iorder ito iaddress ithe iproblem iof ithe imasked iface
iproducts. iBeing ia iBSD-licensed iproduct, iOpenCV imakes irecognition iprocess. iThe ifirst istep iin ihis imodel iwas ito
iit ieasy ifor ibusinesses ito iutilize iand imodify ithe icode. idiscard ithe imasked iface iregion. iNext, ithey iapplied ia ipre-
iThe ilibrary ihas iquite i2500 ioptimized ialgorithms, iwhich trained ideep iConvolutional ineural inetwork i(CNN) ito
ihas ia icomprehensive iset iof iboth iclassic iand istate-of ithe- iextract ithe ibest ifeatures ifrom ithe iobtained iregions
art icomputer ivision iand imachine ilearning ialgorithms. i(mostly ieyes iand iforehead iregions). iFinally, ithe iBag-of-
features iparadigm iwas iapplied ion ithe ifeature imaps iof ithe
ilast iconvolutional ilayer iin iorder ito iquantize ithem iand ito
II. RELATED IWORK iget ia islight irepresentation icomparing ito ithe ifully
iconnected ilayer iof iclassical iCNN.
Shuang iLi iet ial. i[2] iproposed ia imethod iset iup iby iHGL
itechnique ito imanage ithe ihead ipose iorder iby iembracing B.QIN iand iD.Lin i[7] iproposed ian ialgorithm icontaining
ishading isurface iinvestigation iof ipictures iand iline ifour imain isteps: iimage ipre iprocessing, iface idetection
irepresentation. iThe iproposed iHGL istrategy iconsolidates iand icrop, iimage isuper iresolution, iand iface imask
ithe iH-channel iof ithe iHSV ishading ispace iwith ithe iface iwearing iconditions iidentification. iThey ideveloped ia
ipotrait iand igrayscale ipicture, iand itrain ithe iCNN ito inew ifacemask iwearing icondition iidentification imethod
iextract ifeatures ifor iclassification. iin icombination iwith iimage isuper iresolution iwith
iclassification inetwork i(SRCNet).
Bensenane iHamdan iet ial. i[4] iproposed ianother istrategy
ifor ithe idiscovery iof iprogrammers iwho iutilize i3D icovers Inamdar iet ial i[8] iproposed ia iface imask idetector iwhich
ito imisdirect iface iacknowledgment iframeworks. iThis iuses i iFacemasknet, iusing ideep ilearning itechniques ito
itechnique iutilizes ithe iAngular iRadial iTransformation isuccessfully itest iwhether ia iperson iis iwearing ia iface
i(ART) ito iseparate iappropriate ihighlights ithat iare itaken imask ior inot. iTheir imanuscript ipresents ithree-class
icare iof iinto ia iclassifier ito ichoose iwhether ithe icaught iclassification inamely iperson iis iwearing ia imask, ior
ipicture iaddresses ia iface ipicture. iimproperly iworn imasks ior ino imask idetected. iUsing ia
ideep ilearning imethod icalled iFacemasknet, ithey ican iwork
Nizam iUd iDin iKamran iJaved iet ial. i[3] iproposed ia iwith istill iimages ias iwell ias iwith ia ilive ivideo istream.
imodel iwith itwo istages: imask iobject idetection iand iimage
icompletion iof ithe iremoved imask iregion. iThe ifirst istage Ejaz iet ial i[9] iproposed ia imodel ithat iconsists iof ifirstly
iof ihis imodel iautomatically iproduces ibinary isegmentation idetecting ithe ifacial iregions. iThen ithey iapproached ithe
ifor ithe imask iregion. iThen, ithe isecond istage iremoves ioccluded iface idetection iproblem iusing iMulti-Task
ithe imask iand isynthesizes ithe iaffected iregion iwith ifine iCascaded iConvolutional iNeural iNetwork i(MTCNN). iThen
idetails iwhile iretaining ithe iglobal icoherency iof iface ithey iperformed ifacial ifeatures iextraction iusing ithe iGoogle
istructure. iFor ithis, ihe ihas iemployed ia iGAN-based iFaceNet iembedding imodel. iAnd ifinally, ithey iperformed
inetwork iusing itwo idiscriminators iwhere ione ithe iclassification itask iby ithe ihelp iof iSupport iVector
idiscriminator ihelps ilearn ithe iglobal istructure iof ithe iface iMachine i(SVM).
iand ithen ianother idiscriminator icomes iin ito ifocus
ilearning ion ithe ideep imissing iregion. i

Use iof itechnologies isuch ias ithe iInternet iof iThings i(IoT),
iUnmanned iAerial iVehicles i(UAVs), iblockchain, iArtificial
iIntelligence i(AI), iand i5G, iamong iothers, ito ihelp
III. PROPOSED IWORK iApplies ia ilinear itransformation ito ithe iincoming idata: iy i=
iThe imethodology ifocuses ion ihow ito iidentify ithe iperson ixAT i+ ib`. iThis icriterion i(CrossEntropyLoss) icombines
ion iimage/video istream iwearing iface imask iwith ithe ihelp inn.LogSoftmax() iand inn.NLLLoss() iin ione isingle iclass.
iof icomputer ivision iand ideep ilearning ialgorithm iby iusing iAdagrad iis ian ioptimizer iwith iparameter-specific ilearning
ithe iOpenCV, iTensor iflow, iKeras iand iPyTorch ilibrary. i irates, iwhich iare iadapted irelative ito ihow ifrequently ia
1. iTrain iDeep ilearning imodel i iparameter igets iupdated iduring itraining iis iused.
2. iApply imask idetector iover iimages i
3. iLive ivideo istream IV. REAL ILIFE IAPPLICATION/ IFUTURE ISCOPE I
Flowchart: i
Businesses iare iconstantly ioverhauling itheir iexisting
iinfrastructure iand iprocesses ito ibe imore iefficient, isafe,
iand iusable ifor iemployees, icustomers, iand ithe icommunity.
iWith ithe iongoing ipandemic, iit’s ieven imore iimportant ito
ihave iadvanced ianalytics iapps iand iservices iin iplace ito
imitigate irisk. iFor ipublic isafety iand ihealth, iauthorities iare
irecommending ithe iuse iof iface imasks iand icoverings ito
icontrol ithe ispread iof iCOVID19. iSo, iwe iimplemented ia
isystem iwhich iallows ius ito idetect iwhether ia iperson iis
iwearing ia imask ior inot. iThis isystem ihas ia ilot iof iscope
iin ithe ireal iworld. iIt ican ibe iused ias ithe ifirst istep ito
ireduce ithe ispread iof icovid-19. iThis isystem ican ibe iused
iDatasets: iThe imajority iof ithe iimages iwere iaugmented ito idetect iwhether ia iperson iis iwearing ia imask ior inot,
iby iOpenCV. iThe iset iof iimages iwere ialready ilabelled iwhen ithis isystem iis iimplemented iinto ithe icity iCCTV
i“mask” iand i“no imask”. iThe iimages ithat iwere ipresent icameras ior ithe itraffic icameras, ithis iwill ihelp iin ireducing
iwere iof idifferent isizes iand iresolutions, iprobably iextracted ithe ipeople iwho iare inot iwearing imasks.
ifrom idifferent isources ior ifrom imachines i(cameras) iof Results
idifferent iresolutions. iIt iconsists iof i194 iimages iin itest iset Without imask:
iand i1315 iimages iin itraining iset. iData ipre-processing:
iData ipre-processing isteps ias imentioned ibelow iwas
iapplied ito iall ithe iraw iinput iimages ito iconvert ithem iinto
iclean iversions, iwhich icould ibe ifed ito ia ineural inetwork
imachine ilearning imodel.
i1. iResizing ithe iinput iimage i(224x224) i
2. iApplying ithe icolour ifiltering i(RGB) iover ithe ichannels
i3. iScaling i/ iNormalizing iimages iusing ithe istandard imean
iof iPyTorch ibuild iin iweights i
4. iCentre icropping ithe iimage iwith ithe ipixel ivalue iof
i224x224x3 i5. iFinally iConverting ithem iinto itensors
i(Similar ito iNumPy iarray)
i6. iSaving ithe imodel iusing iPyTorch i7. iUsing ithe imodel
iin ithe ilive iwebcam ito idetect ithe imask iand iwithout imask
ipeoples.
Deep iLearning iFrameworks: iTo iimplement ithis ideep
ilearning inetwork, iwe ihave ithe ifollowing ioptions. i
1. iTensor iFlow i
2. iKeras i With imask:
3. iPyTorch i
4. iTensorflow i
We iare iusing ithe iPyTorch ibecause iit iruns ion iPython,
iwhich imeans ithat ianyone iwith ia ibasic iunderstanding iof
iPython ican iget istarted ion ibuilding itheir ideep ilearning
imodels, iand ialso iit ihas ithe ifollowing iadvantage
icompared iwith iTensor iFlow i
1. iData iParallelism i
2. iIt ilooks ilike ia iFramework. i
PyTorch: iResNet-101 iis ia iconvolutional ineural inetwork
ithat iis i101 ilayers ideep iis iused ito ibuild ithis iNN imodel.
iEpidemic(https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumb
er=9130585) i
3. iA iNovel iGAN-Based iNetwork ifor iUnmasking iof
iMasked iFace(https://ieeexplore.ieee.org/document/9019697)
i
4. iThe idetection iof ispoofing iby i3D imask iin ia i2D
iidentity irecognition
isystem(https://www.sciencedirect.com/science/article/pii/S11
10866517303067) i
5. iGuanjun iGuo, iHanzi iWang, iYan iYan, iJin iZheng, iBo
iLi, iA ifast iface idetection imethod ivia iconvolutional ineural
inetwork, iNeurocomputing, iVolume i395, i2020, iPages i128-
137, iISSN i0925- i2312(https://arxiv.org/abs/1803.10103)
6. iWalid, iHariri. i(2020). iEfficient iMasked iFace
iRecognition iMethod iduring ithe iCOVID-19 iPandemic.
ihttps://doi.org/10.21203/rs.3.rs-39289/v1
7. iB. iQIN iand iD. iLi, iIdentifying ifacemask-wearing
icondition iusing iimage isuper-resolution iwith iclassification
inetwork ito iprevent iCOVID-19, iMay i2020, idoi:
i10.21203/rs.3.rs-28668/v1.
CONCLUSION
(https://www.researchsquare.com/article/rs-28668/v1)
i8. iInamdar, iMadhura iand iMehendale, iNinad, iReal-Time
As ithe itechnology iare iblooming iwith iemerging itrends ithe
iFace iMask iIdentification iUsing iFacemasknet iDeep
isupply iso iwe've inovel iface imask idetector iwhich imight
iLearning iNetwork i(July i29, i2020). iAvailable iat iSSRN:
ipossibly icontribute ito ipublic ihealthcare. iWe iused
i(https://ssrn.com/abstract=3663305)
iOpenCV, itensor iflow, ikeras, iPytorch iand iCNN ito idetect
9. iEjaz, iM. iS., i& iIslam, iM. iR. i(2019). iMasked iFace
iwhether ipeople iwere iwearing iface imasks ior inot. iThe
iRecognition iUsing iConvolutional iNeural iNetwork. i2019
imodels iwere itested iwith iimages iand ireal-time ivideo
iInternational iConference ion iSustainable iTechnologies ifor
istreams. iThe iaccuracy iof ithe imodel iis iachieved iand, ithe
iIndustry4.0(STI)
ioptimization iof ithe imodel icould ibe ia icontinuous iprocess
(https://doi.org/10.1109/STI47673.2019.9068044)
iand ithat iwe iare ibuilding ia ihighly iaccurate isolution iby
ituning ithe ihyper iparameters. iThis ispecific imodel ican ibe
iused ias ia iuse icase ifor iedge ianalytics. iBy ithe ievent iof
iface imask idetection, iwe iwill idetect iif ithe iperson iis
iwearing ia iface imask iand ipermit itheir ientry iwould ibe iof
igreat ihelp ito ithe isociety.

REFERENCES i
1. iA iComprehensive iReview iof ithe iCOVID-19 iPandemic
iand ithe iRole iof iIoT, iDrones, iAI, iBlockchain, iand i5G iin
iManaging iIts iImpact
i(https://ieeexplore.ieee.org/document/9086010) i
2. iMulti-angle iHead iPose iClassification iwhen iWearing ithe
iMask ifor iFace iRecognition iunder ithe iCOVID-19
iCoronavirus

You might also like