Automated Target Detection in Aerial Images using Convolutions Neural Networks (CNN)

Abstract

In iComputer ivision, iobject idetection iand iclassification iare iactive ifields iof iresearch. iApplications iof iobject idetection iand iclassification iincludes ia idiverse irange iof ifields isuch ias isurveillance, iautonomous icars, irobotic ivision, isearch iand irescue, idriver iassistance isystems iand imilitary iapplications. iIn ithe ilast icouple iof idecades, iConvolution iNeural iNetwork i(CNN) iemerged ias ithe imost iactive ifield iof iresearch. iThere iare ia inumber iof iapplications iof iCNN, iand iits iarchitectures iare iused ifor ithe iimprovement iof iaccuracy iand iefficiency iin ivarious ifields. iIn ithis iresearch, iautomated iimage iinterpretation ifor itarget idetection iand irecognition iin isatellite/aerial iimages iis ipresented ithat iattracted imuch iattention iin ipast ifew iyears. iBut ilarge iimages iwith icomplex ibackground iand ithe iuneven idistribution iof itrainings isamples imake iit imore ichallenging, iparticularly iwith ismall iand idense iobjects. iRecently ivarious ideep ilearning itechniques imainly ibased ion ithe iCNN iare iproposed. iThe iperformance iof iall ithese itechniques, ihowever, idepends ion ithe isituations ithey iare iuse. iHowever, iin ithe icontext iof iobject idetection ifrom isatellite iimages iwe iexamine ithe iperformance iof ithe ilatest iCNN ialgorithms. iThis iresearch idetails ithe iprocedure iand iparameters iused ifor ithe itraining iof iconvolutional ineural inetworks i(CNNs) ion ia iset iof iaerial iimages ifor iefficient iand iautomated iobject idetection. iPotential iapplication iareas iin ithe itransportation iand imany iother ifields iare ialso ihighlighted. iThe iaccuracy iand ireliability iof iCNNs idepend ion ithe inetwork’s itraining iand ithe iselection iof ioperational iparameters. iThe iobject idetection iresults ishow ithat iby iselecting ia iproper iset iof iparameters, ia iCNN ican idetect iand iclassify iobjects iwith ia ihigh ilevel iof iaccuracy iand icomputational iefficiency. iFurthermore, iusing ia iconvolutional ineural inetwork iimplemented iin ithe i“YOLOv3” i(“You iOnly iLook iOnce”) iplatform, iwe idemonstrated ithat iYOLOv3 inot ionly iexceeds iin ithe isensitivity iand iprocessing itime iof iother iCNN ialgorithms ibut ialso iin idetecting ismall iand idense itargets. iThe ieffectiveness iof ithe iYOLOv3 iframework ihas ibeen idemonstrated ithrough iextensive iexperiments iand icomprehensive ievaluations ion iDOTA: iA iLarge iScale iDataset ifor iObject iDetection iin iAerial iImages. iThis idataset icontains ihigh iresolution iimages ithat iare icollected ifrom ithe iGoogle iEarth, isome iare itaken iby isatellite iJL-1, iand ithe iothers iare itaken iby isatellite iGF-2 iof ithe iChina iCentre ifor iResources iSatellite iData iand iApplication.YOLOv3 iachieves imAp iof i61.94% ithat iis i1.08% ihigher ias icompared ito iother idetecting imethods.