简介:无
简介:Thispaperconcernstheproblemofobjectsegmentationinreal-timeforpickingsystem.Aregionproposalmethodinspiredbyhumanglancebasedontheconvolutionalneuralnetworkisproposedtoselectpromisingregions,allowingmoreprocessingisreservedonlyfortheseregions.Thespeedofobjectsegmentationissignificantlyimprovedbytheregionproposalmethod.Bythecombinationoftheregionproposalmethodbasedontheconvolutionalneuralnetworkandsuperpixelmethod,thecategoryandlocationinformationcanbeusedtosegmentobjectsandimageredundancyissignificantlyreduced.Theprocessingtimeisreducedconsiderablybythistoachievetherealtime.Experimentsshowthattheproposedmethodcansegmenttheinterestedtargetobjectinrealtimeonanordinarylaptop.
简介:Synchronouschipsealisanadvancedroadconstructingtechnology,andthegravelcoveragerateisanimportantindicatoroftheconstructionquality.Inthispaper,anovelapproachforgravelcoverageratemeasurementisproposedbasedondeeplearning.Convolutionalneuralnetwork(CNN)isusedtosegmenttheimageofgroundcoveredwithgravels,andthegravelcoveragerateiscomputedbythepercentageofgravelpixelsinthesegmentedimage.Thegravelcoverageratedatasetformodeltrainingandtestingisbuilt.Theperformanceoffullyconvolutionalneuralnetwork(FCN)andU-Netmodelinthedatasetistested.AbettermodelnamedGravelNetisconstructedbasedonU-Net.Thescaledexponentiallinearunit(SELU)isemployedintheGravelNettoreplacethepopularcombinationofrectifiedlinearunit(ReLU)andbatchnormalization(BN).Dataaugmentationandalphadropoutareperformedtoreduceoverfitting.Theexperimentalresultsdemonstratetheeffectivenessandaccuracyofourproposedmethod.OurtrainedGravelNetachievesthemeangravelcoveragerateerrorof0.35%ontestdataset.
简介:Deepconvolutionalneuralnetworks(DCNNs)haveshownoutstandingperformanceinthefieldsofcomputervision,naturallanguageprocessing,andcomplexsystemanalysis.Withtheimprovementofperformancewithdeeperlayers,DCNNsincurhighercomputationalcomplexityandlargerstoragerequirement,makingitextremelydifficulttodeployDCNNsonresource-limitedembeddedsystems(suchasmobiledevicesorInternetofThingsdevices).NetworkquantizationefficientlyreducesstoragespacerequiredbyDCNNs.However,theperformanceofDCNNsoftendropsrapidlyasthequantizationbitreduces.Inthisarticle,weproposeaspaceefficientquantizationschemewhichuseseightorlessbitstorepresenttheoriginal32-bitweights.Weadoptsingularvaluedecomposition(SVD)methodtodecreasetheparametersizeoffully-connectedlayersforfurthercompression.Additionally,weproposeaweightclippingmethodbasedondynamicboundarytoimprovetheperformancewhenusinglowerprecision.Experimentalresultsdemonstratethatourapproachcanachieveuptoapproximately14xcompressionwhilepreservingalmostthesameaccuracycomparedwiththefull-precisionmodels.TheproposedweightclippingmethodcanalsosignificantlyimprovetheperformanceofDCNNswhenlowerprecisionisrequired.
简介:ThemaingoalofroutingsolutionsistosatisfytherequirementsoftheQualityofService(QoS)foreveryadmittedconnectionaswellastoachieveaglobalefficiencyinresourceutilization.InthispaperproposesasolutionbasedonHopfieldneuralnetwork(HNN)todealwithoneofrepresentativeroutingproblemsinuni-castrouting,i.e.themulti-constrained(MC)routingproblem.ComputersimulationshowsthatwecanobtaintheoptimalpathveryrapidlywithournewLyapunovenergyfunctions.
简介:Basedoncurrentresearchonapplicationsofchaoticneuronnetworkforinformationprocessing,thestabilityandconvergenceofchaoticneuronnetworkareprovedfromtheviewpointofenergyfunction.Moreover,anewauto-associativematrixisdevisedforartificialneuralnetworkcomposedofchaoticneurons,thus,animprovedchaoticneuronnetworkforassociativememoryisbuiltup.Finally,theassociativerecallingprocessofthenetworkisanalyzedindetailandexplanationsofimprovementaregiven.
简介:为网络交通,WPANFIS,它为多决定分析依靠小浪包变换(WPT)和适应neuro模糊的推理系统(ANFIS)的预言的新奇方法论在这篇文章被建议。在网络交通的自我类似的普遍存在在更早的研究被表明了,它展出长期的依赖(LRD)和短范围依赖(SRD)。另外,小浪分解是为LRDdecorrelation的一个有效工具,这被显示出。新方法把WPT用作装decoorrelateLRD并且让更多精确在原来的交通的高周波的节划分的小浪变换的扩展。然后,能从原来的交通提取有用信息的ANFIS为每个分解非静止的小浪系数的更好的预言性能在这研究被实现。模拟结果证明建议WPANFIS能在真实网络交通环境完成高预言精确性。
简介:在这份报纸,我们建议了能与深convolutional从食物图象识别盘子类型,食物成分,和煮的方法的一个多工系统神经网络。我们为每个班与至少500幅图象建立了不同食物的360个班的数据集。到数据的噪音,它是从因特网收集了的还原剂,孤立点图象通过与深convolutional特征训练的一个类的SVM被检测并且消除。我们同时训练了一个盘子标识符,一个煮的方法识别器,和一个多标签成分察觉者。他们在深网络体系结构分享一些低级的层。建议框架与手工制作的特征,和识别器和成分察觉者能被用于没在训练数据集被包括为用户提供引用信息的盘子的煮的方法比传统的方法显示出更高的精确性。
简介:在这份报纸,我们比较了联合网络隧道编码的表演(JNCC)为多点传送当独占时,用低密度同等值支票(LDPC)的继电器网络作为隧道代码编码,Convolutional编码或(XOR)编码的网络在中间的继电器节点使用了。多点传送继电器传播是二个固定继电器节点在第二在作出贡献的传播计划的一种类型在基础收发器车站(BTS)和一双活动车站之间的端对端的传播跳跃。我们认为一个方法和二个方法多点传送评估位错误率(BER)和产量性能的情形。是否使用XOR网络在中间的继电器节点编码,被看了那,一样的传播因此在更少的时间槽变得可能产量性能能被改进。而且我们也在建议系统模型,差异和multiplexing获得在被考虑了讨论了二种可能的情形。它值得通知那BER和产量为LDPC代码完成了比对讨论的所有计划的Convolutional代码好。
简介:Thispaperdescribestheinverstigationdevotedtoestablishsuitableweightsinafeed-forwardneuralnetworkrealizingthenarrow-bandfilteringmapinthecaseofadaptivelineenhancement(ALE)bytheutilityoftheoptimumcommonlearningratebackpropagation(OCLRBP)algorithm.Itisfoundthatafeed-forwardnetworkwith64linearinputandoutputneurons,and8oddsigmoidneuronsinthehiddenlayer,i.e.an(64→8→64)architecture,couldestablishthespecificinput-outputfunctioninthecaseofrelativelylowsignal-to-noiseradio.Onlyisaninputsignalconsistingofmixedperiodicandbroad-bandcomponentsavailabletothenetworksystem.Afterlearning,boththe"fanning-in-connectionpatterns",eachofwhichconsistsofweightsfanningintoahidden-neuronFromalltheoutputsofinput-neurons,andthe"fanning-out-connectionpatterns",eachofwhichconsistsofweightsfanningoutfromahidden-neurontoalltheinputsofoutput-neurons,aretunedtotheperiodicsignals.Thenonline
简介:Gatematrixlayoutproblemplaysanimportantroleinintegratedcircuitdesign,butitsoptimizationisNP-hard.Inthispaper,typicalgatelayoutproblemisanalysedandadaptedtoneuralnetworkrepresentation,furthermorethesimulatedresultsaregiven.
简介:Onthebasisofanalyzingsomeneuralnetworkstoragecapacityproblemsanetworkmodelcomprisinganewencodingandrecallingschemeispresented.Byusingsomelogicaloperationswhichoperateonthebinarypatternstringsduringinformationprocessingprocedurethemodelcanreachahighstoragecapacityforacertainsizeofnetworkframework.
简介:Inthecontextoftowermeasuredradiationdatasets.followingthecorrectionprinciplemeetingadiagnosticequationindataqualitycontrolandintermsofatechniqueformodelconstructionondataandANN(artificialneuralnetwork)retrievalforBPcorrectionofradiationmeasurementswithrougherrorsavailable,aBPmodelispresented.Evidencesuggeststhatthedevelopedmodelworkswellandissuperiortoaconvenientmultivariatelinearregressionmodel,indicatingitswideapplications.
简介:NewalgorithmsbasedonartificialneuralnetworkmodelsarepresentedforcubicNURBScurveandsurfaceinterpolation.Whenallthknotspansareidentical,theNURBScurveinterpolationproceduredegeneratesintothatofuniformrationalB-splinecurves.Ifalltheweightsofdatapointsareidentical,thentheNURBScurveinterpolationproceduredegeneratesintotheintegralB-splinecurveinterpolation.