端對(duì)端深度神經(jīng)網(wǎng)絡(luò)架構(gòu)在汽車自動(dòng)駕駛車速及轉(zhuǎn)向角預(yù)測(cè)的應(yīng)用

electronics

/journal/electronics

Electronics2021,10,1266.

/10.3390/electronics10111266

Article

End-to-EndDeepNeuralNetworkArchitecturesforSpeedandSteeringWheelAnglePredictioninAutonomousDriving

PedroJ.Navarro1,*,LeanneMiller1,FranciscaRosique1,CarlosFernindez-Isla1andAlbertoGila-Navarro2

checkfor

updates

Citation:Navarro,P.J.;Miller,L.;Rosique,F.;Fernández-Isla,C.;Gila-Navarro,A.End-to-EndDeep

NeuralNetworkArchitecturesfor

SpeedandSteeringWheelAnglePredictioninAutonomousDriving.Electronics2021,10,1266.

https://

/10.3390/electronics10111266

AcademicEditors:DongSeogHan,KalyanaC.VeluvoluandTakeoFujii

Received:13April2021

Accepted:18May2021

Published:25May2021Publisher’sNote:MDPIstaysneutralwithregardtojurisdictionalclaimsinpublishedmapsandinstitutionalaf?l-iations.

Copyright:?2021bytheauthors.LicenseeMDPI,Basel,Switzerland.Thisarticleisanopenaccessarticledistributedunderthetermsand

conditionsoftheCreativeCommons

Attribution(CCBY)license(https

://

/licenses/by/

4.0/).

1

2

*

Divisi6ndeSistemaseIngenieriaElectr6nica(DSIE),CampusMuralladelMar,s/n,UniversidadPolitécnicadeCartagena,30202Cartagena,Spain;ler@upct.es(L.M.);paqui.rosique@upct.es(F.R.);

carlos.fernandez@upct.es(C.F.-I.)

GenéticaMolecular,InstitutodeBiotecnologiaVegetal,Edi?cioI+D+I,PlazadelHospitals/n,UniversidadPolitécnicadeCartagena,30202Cartagena,Spain;alberto.gilan@um.es

Correspondence:pedroj.navarro@upct.es;Tel.:+34-968-32-6546

Abstract:Thecomplexdecision-makingsystemsusedforautonomousvehiclesoradvanceddriver-assistancesystems(ADAS)arebeingreplacedbyend-to-end(e2e)architecturesbasedondeep-neural-networks(DNN).DNNscanlearncomplexdrivingactionsfromdatasetscontainingthousandsofimagesanddataobtainedfromthevehicleperceptionsystem.Thisworkpresentstheclassi?cation,designandimplementationofsixe2earchitecturescapableofgeneratingthedrivingactionsofspeedandsteeringwheelangledirectlyonthevehiclecontrolelements.Theworkdetailsthedesignstagesandoptimizationprocessoftheconvolutionalnetworkstodevelopsixe2earchitectures.Inthemetricanalysisthearchitectureshavebeentestedwithdifferentdatasourcesfromthevehicle,suchasimages,XYZaccelerationsandXYZangularspeeds.Thebestresultswereobtainedwithamixeddatae2earchitecturethatusedfrontimagesfromthevehicleandangularspeedstopredictthespeedandsteeringwheelanglewithameanerrorof1.06%.Anexhaustiveoptimizationprocessoftheconvolutionalblockshasdemonstratedthatitispossibletodesignlightweighte2earchitectureswithhighperformancemoresuitableforthe?nalimplementationinautonomousdriving.

Keywords:autonomousdriving;end-to-endarchitecture;speedandsteeringwheelangleprediction;DNNforregression

1.Introduction

Autonomousdrivingtechnologyhasadvancedgreatlyinrecentyears,butitisstillanongoingchallenge.Traditionally,intelligentdecisionmakingsystemsonboardautonomousvehicleshavebeencharacterizedbytheirenormouscomplexity[

1

]andarecomposedofmultiplesubsystems,includingaperceptionsystem,globalandlocalnavigationsystems,acontrolsystem,asurroundingsinterpretationsystem,etc.,[

2

].Thesesubsystemsarecombinedaimingtocoverthecomplicateddecisionsandtaskswhichthevehiclemustperformwhilstdriving.Toobtaintheobjectivesofthevehicle,thesesubsystemsuseawiderangeoftechniqueswhichinclude:cognitivesystems[

3

],agentsystems[

4

],fuzzysystems[

5

],neuralnetworks[

6

],evolutionaryalgorithms[

7

]orrule-basedmethods[

8

].

Deeplearningtechniquesarebecomingincreasinglypopularandarenowavaluabletoolinawiderangeofindustries,includingtheautomotiveindustry,duetotheirpowerfulimagefeatureextraction.Thesetechniqueshaveallowedtheso-calledend-to-end(e2e)drivingapproachtoappear,simplifyingthetraditionalsubsystemsgreatlyandreducingthetasksofmodelingandcontrolofthevehicle[

9

](Figure

1

).TheappearanceofDNNsmeanthatdecision-makingsystemsonboardautonomousvehiclescanreplacemanyofthesubsystemsmentionedpreviouslywithneuralblocks[

10

].Theseneuralblocks,properlyinterconnectedandtrainedwiththecorrectdataarecapableofobtainingperformancesgreaterthan95%forthepredictionofvehiclecontrolvariables[

11

].Anadvantageofthesemodelsisthattheygenerallyrequirefeweronboardsensorsasthemainsourceof

2of21

Electronics2021,10,1266

informationfedtotheDNNsusuallyconsistsofRGBimagesandkinematicdatafromaninertialmeasurementunit(IMU)[

12

].Thismakesend-to-enddrivingsystemsmoreeasilyaccessiblethanthetraditionalperceptionsubsystemswithsensorssuchasLIDARwhichareverycostly.

Figure1.Traditionaldrivingsystemscomparedtoend-to-enddrivingsystem.

Deeplearningmethodsforautonomousdrivinghavegainedpopularitywithadvance-mentsinhardware,suchasGPUs,andmorereadilyavailabledatasets,bothforend-to-enddrivingtechniques[

13

]andtheuseofdeeplearninginindividualsubsystems[

14

].Therehavebeenavarietyofdifferentapproachesforthedevelopmentofdrivingapplicationsusingendtoendlearningtechniques.Inonestudy,a98%accuracywasobtainedusingconvolutionalneuralnetworks(CNN)togeneratesteeringanglesfromimagesgeneratedbyafrontviewcamera[

15

].Inasimilarwork,asequenceofimagesfromapublicdatasetwasusedasinputtotheCNN,topredictwhetherthevehiclewasaccelerating,deceleratingormaintainingspeedaswellascalculatingthesteeringangle[

16

].

AninterestingapproachdesignedaCNNtodevelopahuman-likeautonomousdrivingsystemwhichaimstoimitatehumanbehaviormeaningthevehiclecanbetteradapttorealroadconditions[

13

].Theauthorsused3DLIDARdataasinputtothemodelandgeneratedsteeringandspeedcommands,andinadrivingsimulationmanagedtodecreaseaccidentswiththeautonomoussystemten-foldcomparedwiththehumandriver.AdrivingsimulatorwasalsousedtotestaCNN-basedclosedloopfeedbacktocontrolthesteeringangleofthevehicle[

17

].TheauthorsdesignedtheirownCNN,DAVE-2SKY,usingtheCaffedeeplearningframeworkandtestedthesysteminalane-keepingsimulation.

Theresultswerepromising,althoughproblemsoccurredifthedistancetothevehicleinfrontbecamelessthan9m.

Variouslongshort-termmemory(LSTM)modelshavealsobeenstudied.Aconvolu-tionalLSTMmodelwithbackpropagationwastrainedtoobtainthesteeringanglefromvideoframesusingtheUdacitydataset[

18

].AnFCN-LSTMarchitecturewasusedtopredictdrivingactionsandmotionfromimagesobtainingalmost85%accuracy.Acon-volutionalLSTMmodelwasalsousedtopredictsteeringanglesfromastreamofimagesfromafrontfacingcamera[

19

],improvingontheresultsfrompreviousworks[

20

].

Anotherapproachconsistsinaddingmoresensors.Inoneworkadatasetwasobtainedusingsurroundviewcamerasinadditiontothetypicalfrontviewcamera[

21

].ThedataobtainedbythecameraswasusedtopredictthespeedandsteeringangleusingexistingpretrainedCNNmodels.Theuseofsurroundviewcamerasimprovedtheresultsobtainedatlowspeeds(<20km/h),butatgreaterspeedstheimprovementwaslesssigni?cant.

Inthiswork,wepresentadetailedstudyimplementingsixend-to-endDNNarchitec-turesforthepredictionofthevehiclespeedandthesteeringwheelangle.Thearchitectureshavebeentrainedandtestedusing78,011imagesfromrealdrivingscenarios,whichwerecapturedbytheCloudIncubatorCar(CIC)autonomousvehicle[

2

].

3of21

Electronics2021,10,1266

2.MaterialsandMethods

DNNend-to-endarchitecturesrequirelargevolumesofdataforthemodelstocon-vergecorrectly.ThedataneededtocreateDNNmodelsforautonomousdrivingorADAScanbeobtainedfromthreedifferenttypesofsources:

1.Adhoctests.Toperformthistypeoftesting,largeresourcesarerequired,intheformofoneormorevehicles,expensiveperceptionsystems(e.g.,LIDAR)andpersonnelcapableoftheinstallation,integrationandcommissioningofsophisticatedsensorsanddatarecordingsystems.Inaddition,thedatamustbepost-processed,andthesynchronizationofthedifferentvehicleinformationsourcesisrequired.

2.Publicdatasets.Therearedatasetsdevelopedbybusinessesanduniversitiesforau-tonomousdrivingwheredataobtainedfromtheperceptionsystemsoftheirvehiclescanbeaccessed[

10

].Someofthesepresentdiversescenarioswithdifferentlightandmeteorologicalconditions[

22

].Table

1

showssomerecentpublicdatasetsincludingnumberofsamples,typesofimagesavailableandtypesofvehiclecontrolactionsstored.

3.Simulators.Giventhecomplexityofconductingrealtests,autonomousdrivingsimulatorshavebecomeoneofthemostwidelyusedalternatives.Thesimulationindustryrangesfromsimulationplatforms,vehicledynamicssimulationandsensorsimulationtoscenariosimulationandevenscenariolibraries.Atpresent,therearemanyoptions,includinggenericsolutionswhichmakeuseofgamesandphysicenginesforsimulation[

23

]androboticssimulators[

16

].Recentlyonthemarketcompaniesthatdevelopsimulationproductsspeci?callydesignedtosatisfytheneedsofautonomousdrivinghaveappeared.SomeofthesecompaniesincludeCognata,CARLA,METAMOTO,etc.

Table1.Publicdatasetsforautonomousdriving.

Ref./Year

Samples

ImageType

LIDAR

RADAR

IMU

ControlActions

UPCT2019

78,000

RGB,Depth

Yes

No

Yes

Steeringwheel,Speed

LyftL5[

24

]/2019

323,000

RGB

Yes

No

Yes

-

nuScenes[

25

]/2019

1,400,000

RGB

Yes

Yes

Yes

-

Pandaset[

22

]/2019

48,000

RGB

Yes

No

Yes

-

Waymo[

23

]/2019

1,000,000

RGB

Yes

No

Yes

-

Udacity[

16

]/2016

34,000

RGB

Yes

No

Yes

Steeringwheel

GAC[

26

]/2019

3,240,000

RGB

No

No

No

Steeringwheel,Speed

Inthisworkadhocdatahasbeenchosen.Toobtainthedata,acustomdatasetwas

created,astheresultofadhocdrivingtestsperformedusingtheCloudIncubatorCarautonomousvehicle(CICar)[

2

](seeFigure

2

),anautonomousvehicleprototypebasedontheadaptionofthecommercialelectricvehicle,RenaultTwizy.Thevehiclehasbeenconvenientlymodi?edandhousesacompleteperceptionsystemconsistingofa2DLIDAR,3DHDLIDAR,ToFcameras,aswellasalocalizationsystemwhichcontainsareal-timekineticunit(RTK)andinertialmeasurementunit(IMU,seeFigure

2

c)andautomationofthedrivingelementsofthevehicle(accelerator,brake,steeringandgearbox).Allofthisiscomplementedwiththebiometricdataofthedriverstakenduringthedrivingtests.

2.1.DrivingTests

Agroupof30driversofdifferentageandgenderwereselectedtoperformthedrivingtests,ofwhich?vewerediscardedduetosynchronizationproblems,recordingfailureorincompletedata.ThedrivingtestswerecarriedoutinCartagenaintheRegionofMurcia,Spain,followingapreviouslyselectedroutewithrealtraf?c.

Thisrouteprovidesasigni?cantsetoftypicalurbandrivingscenarios:(a)junctionswithrightofwayandchangesofpriority;(b)incorporation,internalcirculationandexitingofaroundabout;(c)drivingalongaroadwithandparkingareas;(d)mergingtraf?c

4of21

Electronics2021,10,1266

situations.Inordertocontemplateagreatervarietyofenvironmentalconditions,eachdrivercompletedtheroutetwiceatdifferenttimesofday(morning,afternoonorevening).InFigure

3

asampleofsomeofthedatasetimagesisshown,wheresomeofthedifferentdrivingconditionscapturedduringthetestscanbeobserved.

Figure2.(a)CloudIncubatorCarautonomousvehicle(CiCar).(b)CiCarondataacquisitionmission.(c)Vehiclemodel

andIMUsensormeasurementdetails.

Figure3.Imagesfromdataset.(a)pedestriancrossing;(b)saturationoftheilluminationonroundabout;(c)carbraking;

(d)complexshadowsontheroad.

2.2.VehicleCon?guration

Asmentionedpreviously,thedatawascollectedusingtheCICarprototypevehicleinmanualmode,drivenbyahumandriver.InTable

2

thevariablesanddataacquiredduringthedrivingtestsareshown,aswellastheinformationaboutthedevicesandsystemsusedtoobtainthedata.

Eachsensorworkswithitsownsamplerate,andinmostcasesthisisdifferentbetweendevices.Toachievethecorrectdatasynchronizationandreconstructthetemporalsequencewithprecision,stampingtimeshavebeengeneratedforeachsensorandthesehavebeensynchronizedatthestartandendoftherecording.Therefore,allthedevicesarecontrolledbythecontrolunitonboardthevehicle,providingaperfecttemporalandspatialsynchronizationofthedataobtainedbythedifferentsensors.Thedatafromeachtestisdownloadedandstoredinthecentralserveroncethedrivehas?nished.

5of21

Electronics2021,10,1266

Table2.CICar.Sensordata.

Variable/Unit

Device/System

Frequency

Vehicleposition/(LLA)acceleration/(m/s2)

1

GNSS-IMU

4Hz

20Hz

angularspeed/(。/s)

20Hz

Steeringwheelangle/(。)

50Hz

Distance/(m)

compactRioControlunit

50Hz

Speed/(m/s)

50Hz

Frontalimage/

RGBDCamera

25fps

Driverattentionimage

RGBCamera

25fps

SurroundingsCloudPoints

LIDARs,ToFcameras

10Hz

1LLA—latitude,longitude,andaltitude.

2.3.DeepLearningEnd-to-EndArchitecturesClassi?cation

End-to-end(e2e)systemsbasedonDNNarchitecturesappliedtoautonomousdrivingcanmodelthecomplexrelationshipsextractedfromtheinformationobtainedfromthevehicleperceptionsystem.Thisisachievedusingdifferenttypesofneuralblocksgroupedintolayers(e.g.,convolutionallayers,fully-connectedlayers,recurrentlayers,etc.),withtheaimofgeneratingdirectcontrolactionsonthesteeringwheel,theacceleratorandthebrake.Theseactionsonthevehiclecontrolelementscanbecategorical,e.g.,increaseordecreasethespeed,ortheycangenerateasetpointonthecontroller,e.g.,turn13.6degreesorreach45km/h.

Themachinelearningalgorithmsthatareusedtomodeldrivingactionsbelongtothesetknownassupervisedlearning.Thesealgorithmsacquireknowledgefromadatasetofsamplespreviouslyacquiredduringdrivingtestswithapreviouslyconditionedvehicle[

2

]orfromdrivingsimulators[

27

].Thesedatasetsincludedatafromtheperceptionsystem,suchas:images(RGBoIR),LIDAR,RADAR,IMU,aswellastheactionsperformedbythedriveronthevehiclecontrolelements,suchasthesteeringwheel,theacceleratorandthebrake.

Thegenerationofdiscretevariablesbyamachinelearningalgorithmisknownasregressionandisawidelystudiedproblem[

28

].RegressionmodelsforDNNusethegradientdescentfunctiontosearchfortheoptimalweightsthatminimizethelossfunction.Thelossfunctionsusedforthesemodelsdifferfromthoseusedintheclassi?cationmodels,withthemostusedbeingthemeanabsoluteerror,meansquareabsoluteerrorormeanabsolutepercentageerror,amongothers.

Thisworkproposesaclassi?cationofe2earchitecturesbasedonthetypeofdatareceivedbytheDNNfromthevehicleperceptionsystem.Thisisdonebyconsideringtheimageprovidedbythevisualperceptionsystemofthevehicleasthemaindatasourceforthee2earchitecture.Basedonthetypeofnetworkinput,thearchitectureshavebeenclassi?edintothreetypes:(1)singledatae2earchitecture(SiD-e2e),(2)mixeddatae2earchitecture(MiD-e2e)andsequentialdatae2earchitecture(SeD-e2e).

2.3.1.SiD-e2eArchitecture

Thistypeofarchitectureusesasingledatasourcefortheinputlayertogeneratethesetpointsdirectlyforthecontrolelementsofthevehicle.TheSiDarchitecturesusethevisualinformationprovidedbyoneormorecameraslocatedonthefrontandperipheryofthevehicletocomposeasingleimageofthevehicles?eldofviewofthevehicleasavisualinputtothenetwork[

15

,

29

,

30

].BeforebeingprocessedbytheDNN,theimagesarereducedinsizeandnormalized.Subsequently,theimagesgothroughconvolutionallayersofdifferentkernelsize(kok)anddepth(d)whichallowtheimagefeaturesthatminimizethecostfunctiontobeextractedautomaticallyinsuccessivelayers.Aftertheconvolutionallayers,theresultingvectoristransformedintoonedimension(Flayer)andconnectedtoasetoffully-connectedlayers(FC)whichhavethedecision-makingcapacity.Lastly,theFClayersendinthenumberofneuronsequaltothenumberofvariablestobe

6of21

Electronics2021,10,1266

predicted[

15

,

28

].Figure

4

showsanexampleoftheSiDarchitecturewherethenormalizedimagefeedsagroupofconvolutionallayerswithdifferentkernelsizes,followedbyasetoffully-connectedlayersanda?naloutputlayer.

Figure4.Singledatae2earchitecture(SiD).

Thenumberofconvolutionallayers,theirsize,paddingandstride,aswellasthenumberofneuronsintheFClayersareadjustedempirically.Theseparametersarede-pendentonthetrainingdatasetandthesizeoftheinputimages.Thereareworkswherethearchitectureshavebeendesignedusingbanksofconvolutional?ltersofincreasingsize[

30

]andthereareotherswherethedesignistheopposite[

31

,

32

].Generallyspeaking,theconvolutionallayerswithasmallkernelsizeextractreducedspatialcharacteristics,suchastraf?csigns,traf?clightsorlaneseparationlines,whilethosewithagreaterkernelsizedetectlargerelementsintheimage,suchasvehicles,pedestriansortheroad[

31

].

2.3.2.MiD-e2eArchitecture

Mixeddataarchitecturesallowdifferentdatasourcesfromthevehicle,suchasRADAR,longitudinalandlateralaccelerations,angularvelocities,mapsorGPStobemergedtogetherwiththevisualinformationfromthevehicle’scameras.TheinclusionofmoreinformationsourcesintheDNNaimsto:(1)improvetheperformanceofthemodel,(2)improvethepredictionofspeci?ccasesorabnormaldriving;and(3)increasethetolerancetofailuresproducedbythedatasources[

21

,

29

,

33

].AsshowninFigure

5

,thistypeofarchitecturecombinestheresultsoftheSiD-e2e,suchasthoseshownintheprevious

Section

2.3.1

,withasetofFClayerswhichallowsthemappingofthecharacteristicsfromothervehicledatasourcesonalayerthatconcatenatesalltheinformation.

Figure

5

showsa?rstinputbranchwheretherelevantinformationisextractedfromtheimagewithasecondbranchthatextractsextrainformation,forexamplefromtheIMUorGPS.Theconcatenationlayerreceivesaspeci?ednumberofinputsfrombothbranchesofthemodel.Thenumberofconnectionsfromeachbranchisusuallydeterminedusingempiricaltechniques.MiDarchitectureishabituallyusedindatafusionintheperceptionsystemsofautonomousvehiclesorADAS.

7of21

Electronics2021,10,1266

Figure5.Mixeddatae2earchitecture(MiD).

2.3.3.SeD-e2eArchitecture

Drivingisataskwherethefutureactionsonthevehicle’scontrolelementsdependgreatlyonthepreviousactions,thereforethepredictionofthecontrolactionscanbemodeledasatimeseriesanalysis[

16

,

26

,

34

].Sequentialdatabasedarchitecturesaimtomodelthetemporalrelationshipsofthedatausingfeedbackneuralunits(seeFigure

6

),thesetypesofneuralnetworksareknownasrecurrentneuralnetworks(RNN)[

34

].BasicRNNscanlearntheshort-termdependenciesofthedatabuttheyhaveproblemswithcapturingthelong-termdependenciesduetovanishinggradientproblems[

35

].Tosolvethevanishinggradientproblems,moresophisticatedRNNarchitectureshaveappearedwhichuseactivationfunctionsbasedongatingunits.Thegatingunithasthecapacityofconditionallydecidingwhatinformationisremembered,forgottenorforpassingthroughtheunit.Thelongshort-termmemory(LSTM)[

36

]andGRU(gatedrecurrentunit)aretwoexamplesofthesekindsofRNNarchitectures[

37

].

Figure6.Sequentialdatae2earchitecture(SeD).

RNN[

15

],LSTM(longshort-termmemory)[

16

]andGRU(gatedrecurrentunit)arethemostusedformodelingthetemporalrelationshipsinthe?eldofe2earchitectures.TheuseofRNNine2earchitecturesrequiresthenetworkinputdatatobetransformedintotemporalsequencesintheformoftimesteps(ts).ThepartitioningoftheNinputsamples

8of21

Electronics2021,10,1266

ofthenetworkwillgenerate(N-ts)temporalsequencesthatwillcorrespondtoanoutputvectorfromthenetworkaccordingtoEquation(1):

S.input=<[I1,..,Its|,[I2,..,Its＋1|...,[In-1-ts,..,IN-1|},

output=<ots＋1,ots＋2,............,oN}

(1)

Figure

7

showstheprocedurestogenerateN-tssequencesofsizetsfromadatasetcomposedofNimagesandNpairsofoutputvalues(v:speed,9:steeringwheelangle).

Figure7.Compositionofsequentialimagesandoutputvaluesdata.

TocreateamodelfromtheSeD-e2earchitectures,thiswillbetrainedwithtemporalsequencesofsizets(I1toIts)andthenextoutputvectortopredict(vts+1,9ts+1)asitisshownintheFigure

7

.

2.4.ParemetersofDeepNeuralNetworkArchitectures

ThenumberofparameterswhichcomeintoplayduringthedesignprocessofaDNNisenormousandwecanseparatethemintothreetypes:

(1)Networkinputparameters.Theseparametersrefertothewaythenetworkinputvaluesarepresented.Fordataintheformofimages,theshapeparametersinclude:

·Normalization.Normalizationmustbeperformedonthedatabeforetrainingthe

DNN.Anadequatenormalizationcanimprovetheconvergenceandperformanceofthenetwork.Equations(2)and(3)showthemostcommontechniques.

Scaled(0,1)=(xi-min)/(max-min)(

2)

Standarized(╱=0,J=1)=(xi-╱)/(J)(

3)

whereminandmax,arethemaximumandminimumvaluespresentinthedatasetX={x1,...,xN},withuandobeingtheaverageandstandarddeviationofthedataset,respectively.Thereareothernormalizationtechniques,forexample,themeancanbesubstitutedforthemodeinEquation(3),forcasesinwhichthedatadistributiondoesnotalignbelowthemean.

·Resizing.Asageneralruleandespeciallyine2earchitecturesforautonomous

driving,theimagesizeisreducedbeforebeingprocessedbythenetwork.Themainreasonforthisistodecreasethenetworkprocessingtimeandtheresourcesinvolvedintheprediction.

·Colorspacetransformations.Itiscommontotransformtheinputimagetoa

colorspaceotherthantheonesuppliedbythecameratoimproveperformance,forexampleHSI,LAB,etc.,[

10

].

·Preprocessing.Whenthedataiscapturedfromdifferentsourcesordataset,these

tendtohavedisparatefeaturesfromthedeviseitselforfromthelightingof

9of21

Electronics2021,10,1266

thescenewheretheimageswerecaptured,thereforehistogramequalizationorimageenhancementalgorithmsareusuallyappliedtonormalizetheappearanceoftheentiredataset.

·Dataaugmentation.Thistechniqueconsistsinincreasingthesizeoftheoriginal

datasetinordertoachievehigherlevelsofgeneralizationandtoimprovetheperformanceofthenetwork[

38

].

(2)Architecturecon?gurationparameters.Theseparametersconstitutethecompositionofonearchitectureoranother,andtheseinclude:

·Typeoflayer.Thearchitecturescanstackdifferentsetsoflayersineachbranch:

FC,Convolutional,RNN,Concatenated,etc.

·Layersettings.Eachlayerhasspeci?ccon?gurations,forexample,convolutional

layerscanbecon?guredwithdifferenttypesof?lters,3o3,5o5,...,kok,theirdepthornumberoflayers.

·Layerdistribution.Thearchitecturescanconsistofasinglebranch,