US20210374541A1(19)UnitedStates(12)PatentApplicationPublication(lo)Pub.No.:US2021/0374541AlISHII(43)Pub.Date:Dec.2,2021(54)INFORMATIONPROCESSINGMETHODPublicationClassificationANDRECORDINGMEDIUM(51)Int.Cl.G06N3/08(2006.01)(71)Applicant:PanasonicIntellectualPropertyG06N3/04(2006.01)CorporationofAmerica,Torrance,CA(US)(52)U.S.Cl.CPC.........G06N3/08(2013.01);G06N3/04(72)Inventor:YasunoriISHII,Osaka(JP)(2013.01)(57)ABSTRACT(21)Appl.No.:17/404,312Inaninformationprocessingmethodtobeexecutedbya(22)Filed:Aug.17,2021computer,withafirstmodeltrainedthroughmachinelearningtooutputdatasimulatingnoise-reduceddatainresponseRelatedU.S.ApplicationDatatoinputnoise-containingdata,featuredataoffirstdatageneratedbythefirstmodelgeneratedviaprocessesleading(63)ContinuationofapplicationNo.PCT/JP2020/uptooutputofseconddatasimulatingnoise-reducedfirst015801,filedonApr.8,2020.dataofinputnoise-containingfirstdataisobtained;this(60)ProvisionalapplicationNo.62/854,673,filedonMayfeaturedataisinputtoasecondmodelthatisanestimation30,2019.model,andinferenceresultdatathatthesecondmodeloutputsinresponsetoaninputofthefeaturedatais(30)ForeignApplicationPriorityDataobtained;andthesecondmodelistrainedthroughmachinelearningbasedontheinferenceresultdataandreferenceDec.20,2019(JP)2019-229945datathatisformakinginferenceaboutthefirstdata.70%PatentApplicationPnblicationDec.2,2021Sheet1of6US2021/0374541Al-oo'O(SIl1w'‘f*.o'S’*oLT)'-POOtHscc.2.24-1'-4->(J(J□nU.i_4->4-1owuyi-cI-£=£o0)O,4->U"S-&sS'shdc2S"FIG.8&sS'sOInputtoDisplayofProportionofsalt-and-peppernoiseadded(%)KernelsizeofGaussianblur«ffirecognizernoiselabel10305070357911ls>ONoiseimage—0.8570.8190.7470.6800.8760.8770.8820.8840.881ls>GeneratedimageWith0.7320.6290.4920.2830.7210.6910.6660.6920.684STreLatentvariablenoise0.9090.8690.8050.5260.8500,8590.8560.8360.828reGeneratedimageWithout0.7460.7080.6510.5520.7640.7550.7540.7590.746oLatentvariablenoise0.9650.9580.9180.7140.9590.9610.9600.9610.964dvyls>ols>o<1■uU1■uUS2021/0374541AlDec.2,20211INFORMATIONPROCESSINGMETHODexecutedbyacomputer,andtheinformationprocessingANDRECORDINGMEDIUMmethodincludes:obtainingfirstsensingdatacontainingnoise;executingfirsttrainingthroughmachinelearning,theCROSSREFERENCETORELATEDfirsttrainingtrainingafirstmodeltooutput,inresponsetoAPPLICATIONSaninputofsensingdatacontainingnoise,simulatedsensingdatathatsimulatessensingdatatobeobtainedbyreducing[0001]ThisisacontinuationapplicationofPCTInternationalApplicationNo.PCT/JP2020/015801filedonApr.8,thenoiseinthesensingdatacontainingthenoise,and2020,designatingtheUnitedStatesofAmerica,whichisinputtingthefirstsensingdatatothefirstmodelandobtainbasedonandclaimspriorityofU.S.ProvisionalPatentingfirstfeaturedata,thefirstmodelgeneratingfeaturedataApplicationNo.62/854,673filedonMay30,2019andofthesensingdatacontainingthenoisegeneratedviaJapanesePatentApplicationNo.2019-229945filedonDec.processesleadinguptooutputofthesimulatedsensingdata20,2019.Theentiredisclosuresoftheabove-identifiedinresponsetotheinputofthesensingdatacontainingtheapplications,includingthespecifications,drawingsandnoise,thefirstfeaturedatabeingfeaturedataofthefirstclaimsareincorporatedhereinbyreferenceintheirentirety.sensingdata,thefirstfeaturedatabeinggeneratedviaprocessesleadinguptooutputofthefirstsimulatedsensingFIELDdatainresponsetoaninputofthefirstsensingdata,thefirstsimulatedsensingdatabeingthesimulatedsensingdata[0002]Thepresentdisclosurerelatestoaninformationsimulatingthefirstsensingdatatobeobtainedbyreducingprocessingmethodtobeexecutedbyacomputer.thenoiseinthefirstsensingdata;inputtingthefirstfeaturedatatoasecondmodeltobesubjectedtosecondtrainingBACKGROUNDthroughmachinelearningandobtainingfirstinferenceresult[0003]Techniquesrelatedtoareconstructingprocessofdata,thesecondtrainingtrainingthesecondmodeltooutputreconstructinganimage(hereinafter,apre-reconstructioninferenceresultdatainresponsetoaninputofthefeatureimage)basedonthefeatureamountofthispre-reconstruc-data,thefirstinferenceresultdatabeingtheinferenceresulttionimagearebeingstudied(see,forexample.NonPatentdatathatthesecondmodeloutputsinresponsetoaninputofLiterature1).Inoneconceivablecase,forexample,anthefirstfeaturedata;andexecutingthesecondtrainingimagereconstructedthroughareconstructingprocessmaybasedonthefirstinferenceresultdataandreferencedata,thebesubjectedtoimagerecognition,andthismayindirectlyreferencedatabeingformakinginferenceaboutthefirstyieldtheresultofimagerecognitionthatwouldbeobtainedsensingdata.fromtheimagerecognitionperformedonthepre-recon-[0008]Inaddition,aninformationprocessingmethodstructionimage.accordingtoanotheraspectofthepresentdisclosureisamethodtobeexecutedbyacomputer,andtheinformationCITATIONLISTprocessingmethodincludes:obtainingfirstsensingdatacontainingnoise;executingfirsttrainingthroughmachineNonPatentLiteraturelearning,thefirsttrainingtrainingafirstmodeltooutput,in[0004]NPL1:DiederikP.KingmaandMaxWelling,responsetoaninputofsensingdatacontainingnoise,“Auto-EncodingVariationalBayes,”arXivpreprintarXiv:simulatedsensingdatathatsimulatessensingdatatobe1312.6114,Dec.20,2013.obtainedbyreducingthenoiseinthesensingdatacontainingthenoise,andinputtingthefirstsensingdatatothefirstSUMMARYmodelandobtainingfirstfeaturedata,thefirstmodelgeneratingfeaturedataofthesensingdatacontainingtheTechnicalProblemnoisegeneratedviaprocessesleadinguptooutputofthesimulatedsensingdatainresponsetotheinputofthesensing[0005]Withtheexistingtechniquedescribedabove,howdatacontainingthenoise,thefirstfeaturedatabeingfeatureever,whenthereconstructingprocessbecomesharder,thedataofthefirstsensingdata,thefirstfeaturedatabeingresultofinferencemadethroughimagerecognitionorthegeneratedviaprocessesleadinguptooutputofthefirstlikeondata,suchasapre-reconstructionimage,maysensingdatainresponsetoaninputofthefirstsensingdata,substantiallydeteriorate.Forexample,whenthequalityofthefirstsimulatedsensingdatabeingthesimulatedsensinganimageoutputthroughareconstructingprocessislow,thisdatasimulatingthefirstsensingdatatobeobtainedbydegradestheresultofimagerecognitionperformedonthereducingthenoiseinthefirstsensingdata;inputtingthefirstimageoutputthroughthereconstructingprocess.Therefore,featuredatatoasecondmodeltobesubjectedtoseconditispossibletosaythattheresultofimagerecognitionthattrainingthroughmachinelearningandobtainingfirstinferwouldbeobtainedfromtheimagereconstructionperformedenceresultdata,thesecondtrainingtrainingthesecondonthepre-reconstructionimagemaysubstantiallydeteriomodeltooutputinferenceresultdatainresponsetoaninputrateinturn.ofthefeaturedata,thefirstinferenceresultdatabeingthe[0006]Thepresentdisclosureprovidesaninformationinferenceresultdatathatthesecondmodeloutputsinprocessingmethodthatmakesitpossibletokeeptheresultresponsetoaninputofthefirstfeaturedata;andoutputtingofinferenceaboutpre-reconstructiondatafromsubstantiallythefirstinferenceresultdata.deterioratingevenwhenthereconstructingprocessbecomes[0009]Anon-transitorycomputer-readablerecordingharder.mediumaccordingtoanotheraspectofthepresentdisclosure,thenon-transitorycomputer-readablerecordingSolutiontoProblemmediumhavingrecordedthereonaprogramthat,upon[0007]Aninformationprocessingmethodaccordingtoexecutedbyaprocessorincludedinacomputer,causestheoneaspectofthepresentdisclosureisamethodtobeprocessortoexecuteinthecomputer:obtainingfirstsensingUS2021/0374541AlDec.2,20212datacontainingnoise;executingflrsttrainingthrough[0018]FIG.6isaflowchartillustratinganexampleofamachinelearning,theflrsttrainingtrainingaflrstmodeltoprocedureofanimagerecognizingmethodinwhichaoutput,inresponsetoaninputofsensingdatacontainingrecognizertrainedinthestatedinformationprocessingnoise,simulatedsensingdatathatsimulatessensingdatatomethodisused.beobtainedbyreducingthenoiseinthesensingdata[0019]FIG.7isadiagramfordescribinganoverviewofcontainingthenoise,andinputtingtheflrstsensingdatatoaninformationprocessingmethodaccordingtoavariationtheflrstmodelandobtainingflrstfeaturedata,theflrstoftheembodiment.modelgeneratingfeaturedataofthesensingdatacontaining[0020]FIG.8isatablesummarizingtheresultofanthenoisegeneratedviaprocessesleadinguptooutputoftheexperimentconductedbytheinventor.simulatedsensingdatainresponsetotheinputofthesensingdatacontainingthenoise,theflrstfeaturedatabeingfeatureDESCRIPTIONOFEMBODIMENTSdataoftheflrstsensingdata,theflrstfeaturedatabeinggeneratedgeneratedviaprocessesleadinguptooutputoftheUnderlyingKnowledgeflrstsensingdatainresponsetoaninputoftheflrstsensingdata,theflrstsimulatedsensingdatabeingthesimulated[0021]Thepresentinventorfoundthefollowingproblemsensingdatasimulatingtheflrstsensingdatatobeobtainedwithrespecttotheimagerecognitiontechniquedescribedinbyreducingthenoiseintheflrstsensingdata;inputtingtheBackgroundArt.flrstfeaturedatatoasecondmodeltobesubjectedtosecond[0022]Animagecapturedbyamonitoringcameraorthetrainingthroughmachinelearningandobtainingflrstinferlikeinstalledataprivatehomeorinapubliclocationmayenceresultdata,thesecondtrainingtrainingthesecondbesubjectedtoimagerecognitionforsecuritypurposeorthemodeltooutputinferenceresultdatainresponsetoaninputlike.Thisimagerecognitionprocessmaybeperformedonofthefeaturedata,theflrstinferenceresultdatabeingtheimagedataonacloudserveraftertheimagedatahasbeeninferenceresultdatathatthesecondmodeloutputsinoutputfromthecameraanduploadedtothecloudserver.Inresponsetoaninputoftheflrstfeaturedata;andoutputtingthiscase,duetotheprivacyprotectionrequirement,noise,theflrstinferenceresultdata.suchasblurring,maybeaddedinadvancetotheimageto[0010]Itistobenotedthatgeneralorspeciflcembodibesubjectedtotheimagerecognitionprocess.Inothermentsoftheabovemaybeimplementedintheformofanwords,theimagerecognitionmayneedtobeperformedonapparatus,asystem,anintegratedcircuit,oracomputerareduced-qualityimageforprivacyprotectionpurpose.Yet,readablerecordingmedium,suchasaCD-ROM,orthroughsuchanimagedegradedwithaddednoisetendstoresultinanydesiredcombinationofanapparatus,asystem,alowimagerecognitionaccuracy.Therefore,areconstructingmethod,anintegratedcircuit,acomputerprogram,andaprocessofimprovingtheimagequalitybyreducingtherecordingmedium.noiseisperformedasapreprocesstotheimagerecognitionprocess.AdvantageousEffects[0023]However,animageaddedwithmoreintensenoisesoastoprotecttheprivacymorereliablyrendersitharderto[0011]Theuseoftheinformationprocessingmethodandperformthereconstructingprocessthereonwithhighaccutheprogramaccordingtothepresentdisclosuremakesitracy.FIG.1isatableillustratingexamplesofimagesaddedpossibletokeeptheresultofinferenceaboutpre-reconstruc-withnoiseatdifferentintensitylevelsandexamplesoftiondatafromdeterioratingsubstantiallyevenwhentheresultsofthereconstructingprocessperformedonthesereconstructingprocessbecomesharder.images.Inthisexample,fourimagesobtainedbyaddingsalt-and-peppernoiseataproportionof10%,30%,50%,BRIEFDESCRIPTIONOFDRAWINGSand70%toimagescapturingahandwrittendigit“9”includedintheModifiedNationalInstituteofStandardsand[0012]TheseandotheradvantagesandfeatureswillTechnology(MNIST)databasearearrangedintheupperbecomeapparentfromthefollowingdescriptionthereofhalf,andimagesobtainedasaresultofsubjectingtheabovetakeninconjunctionwiththeaccompanyingDrawings,byfourimagestothereconstructingprocessarearrangedinthewayofnon-limitingexamplesofembodimentsdisclosedlowerhalf.Suchareconstructingprocesscanbeperformedherein.byuseofamodeltrainedthroughmachinelearningfor[0013]FIG.1isatableillustratingexamplesofimagesremovingorreducingtargetnoise(hereinafter,regardlessofwithdifferentimagequalitiesandexamplesofresultsofawhethernoiseisremovedorreducedasanactualeffect,thereconstructingprocessperformedontheseimages.term“reduce”isused).Thatis,forexample,thereconstruct[0014]FIG.2isadiagramfordescribinganoverviewofingprocesscanbeperformedbyuseofanautoencoder.InaninformationprocessingmethodaccordingtoanembodithereconstructingprocessintheexampleillustratedinFIG.ment.1,aconvolutionalautoencoderisused.Withreferenceto[0015]FIG.3isaflowchartillustratinganexampleofaFIG.1,whentheproportionofthenoiseisupto30%,theprocedureoftheinformationprocessingmethodaccordingimagesobtainedthroughreconstructioneachincludethetotheembodiment.handwritten“9”thatcanberecognizedbyhumanvision.Thisallowsforspeculatingthattheappearanceofthe[0016]FIG.4isaflowchartillustratinganexampleofaimagesobtainedthroughthereconstructionisclosetohowprocedureofamethodoftrainingavariationalautocoderintheimagesappearedbeforethenoisewasadded.Inthisthestatedinformationprocessingmethod.example,sincethedigit“9”inthepre-reconstructionimages[0017]FIG.5isaflowchartillustratinganexampleofacanberecognizedrelativelyeasilywithhumanvision,theprocedureofamethodoftrainingarecognizerinthestatednoiseatsuchintensitylevelsmaynotbesufficientforinformationprocessingmethod.protectingtheprivacy.US2021/0374541AlDec.2,20213[0024]Yet,adigitinapre-reconstructionimagebecomesdispersiondataofthefirstsensingdata.Thefeaturedatamaymoredifficulttorecognizewithhumanvisionastheprobealatentvariablepertainingtoapriordistributionoftheportionofnoiseincreases.Inotherwords,ifthistechniquefirstsensingdata.isappliedtoapictureofahuman,forexample,onecan[0028]Inthismanner,theinformationprocessingmethodexpectanadvantageouseffectofmorereliableprivacyaccordingtooneaspectofthepresentdisclosurecanuse,forprotection.However,whentheproportionofthenoiseexample,intermediatedataofanautoencoderoravariareachesorexceeds50%,theoverallcontrastisreducedintionalautoencoderconventionallyusedtoreducenoiseinthereconstructedimages,theoutlinesbecomemoreblurred,imagedata.Therefore,inacasewhereanautoencoderforandthewhitelinesthatwoulddepictthedigit“9”arereducingnoiseinanimageisalreadyusedforimagepartiallycutoffordeformed.Therefore,eveniftheseimagesrecognition,forexample,anadditionaluseofarecognizeraresubjectedtoimagerecognition,whethertheimagemakesitpossibletoconstructanenvironmentwheretherecognitioncanproduceanaccurateresultisuncertain.Ininformationprocessingmethodaccordingtooneaspectofthismanner,thereisatrade-offbetweenintensifyingthethepresentdisclosureistobeexecuted.Inotherwords,innoiseforprotectingtheprivacyandimprovingtheaccuracythiscase,theinformationprocessingmethodaccordingtoinreconstructingtheimage.Accordingly,tryingtoincreaseoneaspectofthepresentdisclosurecanbeintroducedtheprivacyprotectionmayresultinsacrificingtheimagewithoutanincreaseintheamountofprocessingandthecostrecognitionperformance,andthismakesitdifficulttoofhardware.Moreover,whenintermediatedataofwhichtheenhancethesecuritywhileutilizingtheimagerecognitiontendencyofinputdataisorganized(inotherwords,theresult,forexample.intermediatedatainwhichthefeaturesoftheinputdataare[0025]Aninformationprocessingmethodaccordingtorepresentedbyapredeterminedstructure),insteadofinteroneaspectofthepresentdisclosureconceivedofinordertomediatedataofasimpleencoder,isinputtothesecondsolvetheaboveproblemisamethodtobeexecutedbyamodel,theperformance(theaccuracyinparticular)ofthecomputer.Theinformationprocessingmethodincludesinferringprocessofthesecondmodelcanbeimproved.obtainingfirstsensingdatacontainingnoise;executingfirst[0029]Thefirstsensingdataandthefirstsimulatedsenstrainingthroughmachinelearning,thefirsttrainingtrainingingdatamaybeobtained,andthefirsttrainingmaybeafirstmodeltooutputsimulatedsensingdatainresponsetoperformedbasedonthefirstsensingdata,thefirstsimulatedaninputofsensingdatacontainingnoise,thesimulatedsensingdata,andthefirstfeaturedata.Moreover,retrainingsensingdatasimulatingsensingdatatobeobtainedbymaybeexecutedafterthesecondtraining.Theretrainingreducingthenoiseinthesensingdatacontainingthenoise;mayincludefurtherexecutingthefirsttraining,obtaininginputtingthefirstsensingdatatothefirstmodelandobtainsecondfeaturedatathatisthefeaturedatageneratedbytheingfirstfeaturedata,thefirstmodelgeneratingfeaturedatafirstmodeltrainedfurther,obtainingsecondinferenceresultofthesensingdatacontainingthenoisegeneratedviadatathatistheinferenceresultdatathatthesecondmodelprocessesleadinguptooutputofthesimulatedsensingdataoutputsinresponsetoaninputofthesecondfeaturedata,inresponsetotheinputofthesensingdatacontainingtheandfurtherexecutingthesecondtrainingbasedonthenoise,thefirstfeaturedatabeingthefeaturedataofthefirstsecondinferenceresultdata.Furthermore,anevaluationonsensingdata,thefirstfeaturedatabeinggeneratedgeneratedaninferenceresultbythesecondmodelindicatedbytheviaprocessesleadinguptooutputofthefirstsensingdataininferenceresultdatamaybeobtained,andtheretrainingresponsetoaninputofthefirstsensingdata,thefirstmayberepeateduntiltheevaluationsatisfiesapredetersimulatedsensingdatabeingthesimulatedsensingdataminedstandard.simulatingthefirstsensingdatatobeobtainedbyreducing[0030]Itishighlylikelythattheperformanceofanthenoiseinthefirstsensingdata;inputtingthefirstfeatureestimatorimprovesastheperformanceofanautoencoderdatatoasecondmodeltobesubjectedtosecondtrainingimproves.Therefore,theperformanceoftheestimatoristhroughmachinelearningandobtainingfirstinferenceresultexpectedtoimproveastheestimatoristrainedinaccordancedata,thesecondtrainingtrainingthesecondmodeltooutputwiththetrainingoftheautoencoderasdescribedabove.inferenceresultdatainresponsetoaninputofthefeatureMoreover,asthetrainingofthefirstmodelthroughmachinedata,thefirstinferenceresultdatabeingtheinferenceresultlearningisexecutedinparallel,forexample,theoutcomeofdatathatthesecondmodeloutputsinresponsetoaninputofthetrainingorthetimingtostopthetrainingcanbethefirstfeaturedata;andexecutingthesecondtrainingdeterminedbyusingtheaccuracyoftheinferencemadebybasedonthefirstinferenceresultdataandreferencedatathatthesecondmodelasanindexoftheresultoftrainingoftheisformakinginferenceaboutthefirstsensingdata.firstmodel.[0026]Thismethodmakesitpossibletokeeptheresultof[0031]Thesensingdatacontainingthenoisemaybeinferenceaboutpre-reconstructiondatafromsubstantiallyimagedata.deterioratingevenwhenthereconstructingprocessbecomes[0032]Thismakesitpossibletoobtainarecognitionhardertoperformwithhighaccuracy.Inotherwords,amodelthatcanexhibithigherrecognitionperformancewithrecognitionmodelthatcanexhibithigherrecognitionperrespecttoanoise-containing,low-qualityimage.formanceonnoise-containingsensingdatacanbeobtained.[0033]Inaddition,aninformationprocessingmethod[0027]Thefirstmodelmayincludeanencoderandaaccordingtoanotheraspectofthepresentdisclosureisadecoder.Theencodermayoutputthefeaturedataofthemethodtobeexecutedbyacomputer.Theinformationsensingdatacontainingthenoiseinresponsetotheinputofprocessingmethodincludesobtainingfirstsensingdatathesensingdatacontainingthenoise.Thedecodermaycontainingnoise;executingfirsttrainingthroughmachinegeneratethesimulatedsensingdatainresponsetoaninputlearning,thefirsttrainingtrainingafirstmodeltooutputofthefeaturedataoutputbytheencoderandoutputthesimulatedsensingdatainresponsetoaninputofsensinggeneratedsimulatedsensingdata.Thefeaturedatamaybeadatacontainingnoise,thesimulatedsensingdatasimulatinglatentvariable.ThefeaturedatamaybemeandataandsensingdatatobeobtainedbyreducingthenoiseintheUS2021/0374541AlDec.2,20214sensingdatacontainingthenoise;inputtingthefirstsensinggenerativemodel.TheVAEisatypeofaneuralnetwork.Indatatothefirstmodelandobtainingfirstfeaturedata,theFIG.2,thefirstmodelreceivesaninputofimagedataasanfirstmodelgeneratingfeaturedataofthesensingdataexampleofsensingdata.containingthenoisegeneratedviaprocessesleadingupto[0039]Theotheroftheaforementionedtwomodelsisaoutputofthesimulatedsensingdatainresponsetotheinputneuralnetworkinferencemodelthatfunctionsasarecogofthesensingdatacontainingthenoise,thefirstfeaturedatanizer.Therecognizerreceivesaninputofintermediatedatabeingthefeaturedataofthefirstsensingdata,thefirstthatarisesintheprocessperformedbythefirstmodel,featuredatabeinggeneratedviaprocessesleadinguptoperformsrecognitionthroughinferenceaboutthereceivedoutputofthefirstsensingdatainresponsetoaninputoftheintermediatedata,andoutputstheresultoftherecognition.firstsensingdata,thefirstsimulatedsensingdatabeingtheInFIG.2,thismodelcorrespondstoasecondmod
本文档为【INFORMATION_PROCESSING_METHOD_AND_RECORDING_MEDIUM】，请使用软件OFFICE或WPS软件打开。作品中的文字与图均可以修改和编辑，
图片更改请在作品中右键图片并更换，文字修改请直接点击文字进行修改，也可以新增和删除文档中的内容。
该文档来自用户分享，如有侵权行为请发邮件ishare@vip.sina.com联系网站客服，我们会及时删除。
[版权声明] 本站所有资料为用户分享产生，若发现您的权利被侵害，请联系客服邮件isharekefu@iask.cn，我们尽快处理。
本作品所展示的图片、画像、字体、音乐的版权可能需版权方额外授权，请谨慎使用。
网站提供的党政主题相关内容(国旗、国徽、党徽..)目的在于配合国家政策宣传，仅限个人学习分享使用，禁止用于任何广告和商用目的。
浙公网安备 33021202002483