简介:Basedonhreedifferentimplementationschemes,thispaperstronglydemonstratesthattheperformanceoftheHoughtransformdependscruciallyonitsimplementationscemewhenitisusedforlinedetection.Moreover,theobtainedresultscanbeusedasatheoreticalbasistopredicttheperformanceoftheHoughtransformaswellastoeliminatethenoiseinHoughspacecomingfromimagenoise.
简介:Givennpropositionalvariables,letKn(i,j),0≤i≤j≤n,betheset(ordisjunction)ofallconjunctionsofiliteralsofwhichexactlyjliteralsarenegative.DunhamandWangconjecturedthatitmayrequireexponentialtimetodecidethateverydisjunctionKn(i,j)isnotvalidbytheresolutionmetho.Thispapergivesaproofoftheconjectureandthenexhibitsanewcounterexampletothefeasibilityoftheresolutionorconsensusmethod.
简介:与exascale来超级计算的时代,电源效率成为了最重要的障碍造一个exascale系统。Dataflow建筑学在为科学应用完成高电源效率有本国的优点。然而,最先进的dataflow体系结构没能为循环处理利用高并行。处理这个问题,我们建议一个pipelining环优化方法(PLO),它在处理元素(PE)在环流动做重复dataflow的数组加速器。这个方法由二种技术,帮助建筑学的硬件重复和帮助说明的软件重复组成。在硬件重复执行模型,一个在薄片上循环控制器被设计产生循环索引,减少计算内核并且打为pipelining执行的一个好基础的复杂性。在软件重复实行模型,另外的环指令被论述解决重复相关性问题。经由这二种技术,准备好了每周期执行的指令的平均数字被增加使浮点联合起来忙。当这二种技术的硬件费用是可接受的时,模拟结果证明分别地,我们的建议方法平均由2.45x和1.1x在浮点效率超过静电干扰和动态循环执行模型。
简介:主成分分析(PCA)和线性判别分析(LDA)是统计模式识别地里的二种流行特征抽取技术。不能直接由于小样品尺寸问题LDA被用于基于外观的面貌识别任务。作为后果,很多基于LDA的面部特征抽取技术被建议相继地处理这个问题。Nullspace方法是在他们之中的最有效的方法之一。Nullspace方法试着发现在在内班scatter矩阵的零空间最大化在班之间scatter的一套判别式向量。它的判别式向量的计算将涉及在一个高度维的矩阵上执行奇异值分解。它通常消费记忆、费时间。在统计分析在Nullspace方法和变化的系数的概念借关键想法我们在场新奇面部特征抽取法,即,判别式基于在这篇论文的变化(DCV)的系数。在FERET和AR脸图象数据库上执行的试验性的结果证明DCV是与Eigenfaces,Nullspace方法,和另外的最先进的美容比较的一种有希望的技术特征抽取法。电子增补材料这篇文章(doi:10.1007/s11390-007-9070-2)的联机版本contatins增补材料,它对授权用户可得到。
简介:AnewmethodforrecognizingChinesecharactersisproposed.Itisbasedontheso-calledfeaturepointsofChinesecharacters.Thefeaturepointsweuseincludethoseonthestrokeofacharacter.i.e.,endpoints.turningpoints,forkpointsandcrosspoints.andthekeypointsonthebackgroundofcharacter.ThismethoddiffersfromthepreviousonesforitcombinesthefeaturepointsonstrokewiththoseonbackgroundanditusesfeaturepointstorecognizeChinesecharactersdirectly.AChinesecharacterrecognitionsystembasedtotop-downdynamicalmatchingoffeaturepointisdeveloped.Thesystemcanrecognizenotonly6763printedsampleSongfontChinesecharactersofsize5.6×5.6mm^2withhighrecognitionrate,butalsothegeneralprintedbooks,magazinesanddocumentswithasatisfactoryrecognitionrateandspeed.
简介:Streamprocessingapplicationscontinuouslyprocesslargeamountsofonlinestreamingdatainrealtimeornearrealtime.Theyhavestrictlatencyconstraints.However,thecontinuousprocessingmakesthemvulnerabletoanyfailures,andtherecoveriesmayslowdowntheentireprocessingpipelineandbreaklatencyconstraints.Theupstreambackupschemeisoneofthemostwidelyappliedfault-tolerantschemesforstreamprocessingsystems.Itintroducescomplexbackupdependenciestotasks,whichincreasesthedifficultyofcontrollingrecoverylatencies.Moreover,whendependenttasksarelocatedonthesameprocessor,theyfailatthesametimeinprocessor-levelfailures,bringingextrarecoverylatenciesthatincreasetheimpactsoffailures.Thispaperstudiestherelationshipbetweenthetaskallocationandtherecoverylatencyofastreamprocessingapplication.Wepresentacorrelatedfailureeffectmodeltodescribetherecoverylatencyofastreamtopologyinprocessor-levelfailuresunderataskallocationplan.Weintroducearecovery-latencyawaretaskallocationproblem(RTAP)thatseekstaskallocationplansforstreamtopologiesthatwillachieveguaranteedrecoverylatencies.WediscussthedifferencebetweenRTAPandclassictaskallocationproblemsandpresentaheuristicalgorithmwithacomputationalcomplexityofO(nlog2n)tosolvetheproblem.Extensiveexperimentswereconductedtoverifythecorrectnessandeffectivenessofourapproach.Itimprovestheresourceusageby15%-20%onaverage.
简介:针对HEVC中SATD率失真代价计算的特点,本文提出利用向量SIMD(单指令多数据流)技术,设计哈达玛变换的并行化方案.该方案采用多加法器和多乘法器协同工作模式,发挥处理器的并行性,通过合理的数据安排,很好地实现了多个宏中数据的并行计算,增大DSP的数据吞吐率,提高数据处理速度.实验结果表明其在单核BWDSP1041上的并行加速比达到87.9,证明了优化工作的有效性.
简介:KL变换是一种适用于任意概率密度函数的正交变换,它能消除各分量之间的相关性.根据协方差矩阵特征值和特征向量有效地进行信息压缩等。相同类的指纹图像的特征码具有较强的相似性.不同类指纹图像的特征码却有一定差异。采用对特征码进行KL变换降维,得到的新的特征码仍然具有同样的相似性和差异性。证明可以通过KL变换后的特征向量进行指纹识别是可行且具有一定意义和研究应用价值。
简介:Thispaperpresentsafactoringalgorithmforcomputingsource-to-Kterminal(SKT)reliability,theprobabilitythatasourceacansendmessagetoaspcifiedsetofterminalsK,inacyclicdirectednetworks(AD-networks)inwhichbothnodesandedgescanfail,BasedonPivotaldecompositiontheorem,anewformulaisdevivedforcomputingtheSKTreliabilityofAD-networks.ByestablishingatopologicalpropertyofAD-networks,itisshownthattheSKTreliabilityofAD-networkscanbecomputedbyrecursivelyapplyingthisformula,TwonewReliabilityPreservingReductionsarealsointroduced.Therecursiontreegeneratedbythepresentedalgorithmhasatmost2^(|V|-|K|-|C|leafnodes,where|V|and|K|arenodessatisfyingsomespecifiedconditions.ThecomputationcomplexityofthenewalgorithmisO(|E|·|V|·2^(|V|-|K|-|C|)intheworstcasewhere|E|isthenumberofedges.Forsource-to-all-terminal(SAT)reliability,itscomputationcomplexityisO(|E|).ComparisonofthenewalgorithmwiththeexistingonesindicatesthatthenewalgorithmismoreefficientforcomputingtheSKTreliabilityofAD-networks.
简介:OneofthecentralquestionsinCAGE^[1]isblendingsurfaceswhichprovidesthetheoreticalbasisforthedesigntechnologyofspacesurfaces.Wewilldiscussthegeneraltheoriesandalgorithmsformultivariatehyperfiniteinterpolationandtheirapplicationtotheblendingofimplicitalgebraicsurfaces,andinvestigatetheexistenceconditionsofhyperfiniteinterpolation.BasedonWu'stheoryonblendingimplicitalgebraicsurfaces,theproblemofblendingtwoquadricsurfacesisstudied.TheconditionsforthecoefficientofgiunderwhichthereexiststhecubicblendingsurfaceS(f)(thelowestdegree)areobtainedandtheconcreteexpressionsoffarepresentedistheyexist.TheseresultscanbeapplieddirectlytoCAGD.
简介:Thecombinationofvisualandtextualinformationinimageretrievalremarkablyalleviatesthesemanticgapoftraditionalimageretrievalmethods,andthusithasattractedmuchattentionrecently.Imageretrievalbasedonsuchacombinationisusuallycalledthecontent-and-textbasedimageretrieval(CTBIR).Nevertheless,existingstudiesinCTBIRmainlymakeeffortsonimprovingtheretrievalquality.Tothebestofourknowledge,littleattentionhasbeenfocusedonhowtoenhancetheretrievalefficiency.Nowadays,imagedataiswidespreadandexpandingrapidlyinourdailylife.Obviously,itisimportantandinterestingtoinvestigatetheretrievalefficiency.Tothisend,thispaperpresentsanefficientimageretrievalmethodnamedCATIRI(content-and-textbasedimageretrievalusingindexing).CATIRIfollowsathree-phasesolutionframeworkthatdevelopsanewindexingstructurecalledMHIM-tree.TheMHIM-treeseamlesslyintegratesseveralelementsincludingManhattanHashing,Invertedindex,andM-tree.TouseourMHIM-treewiselyinthequery,wepresentasetofimportantmetricsandrevealtheirinherentproperties.Basedonthem,wedevelopatop-kqueryalgorithmforCTBIR.ExperimentalresultsbasedonbenchmarkimagedatasetsdemonstratethatCATIRIoutperformsthecompetitorsbyanorderofmagnitude.
简介:多重消息拷贝保证消息交货的在容忍的网络通常利用的延期的路由协议,以便克服无法预言的节点活动性和容易打断的连接。一个store-carry-and-forward范例也被建议进一步改进消息交货。然而,过多的消息拷贝导致缓冲区和带宽的缺乏。水花并且等待路由协议被建议了减少不受限制的消息拷贝的缓冲区和传播引起的网络超载。然而,当一个节点缓冲区相当被抑制时,仍然在那里存在拥挤问题。在这份报纸,我们在水花上建议安排的一条消息和落下策略并且等待路由协议(SDSRP)。为了改进交货比率,首先,SDSRP由评估在交货比率上复制并且落下一个消息拷贝的影响计算每条消息的优先级。随后,安排并且落下决定根据优先级被做。为了推进,增加交货比率,我们在水花上建议安排的一条改进消息和落下策略并且通过提高估计参数的精确性等待路由协议(ISDSRP)。最后,我们在一个基于合成、真实的踪迹进行广泛的模拟。结果证明与另外的缓冲区管理策略相比,ISDSRP和SDSRP完成更高的交货比率,类似的平均hopcounts,和更低的架空的比率。
简介:Sentimentanalysis,ahotresearchtopic,presentsnewchallengesforunderstandingusers'opinionsandjudg-mentsexpressedonline.Theyaimtoclassifythesubjectivetextsbyassigningthemapolaritylabel.Inthispaper,weintroduceanovelmachinelearningframeworkusingauto-encodersnetworktopredictthesentimentpolaritylabelatthewordlevelandthesentencelevel.Inspiredbythedimensionalityreductionandthefeatureextractioncapabilitiesoftheauto-encoders,weproposeanewmodelfordistributedwordvectorrepresentation"PMI-SA"usingasinputpointwise-mutual-information"PMI"wordvectors.Theresultedcontinuouswordvectorsarecombinedtorepresentasentence.Anunsupervisedsentenceembeddingmethod,calledContextualRecursiveAuto-Encoders"CoRAE",isalsodevelopedforlearningsentencerepresentation.Indeed,CoRAEfollowsthebasicideaoftherecursiveauto-encoderstodeeplycomposethevectorsofwordsconstitutingthesentence,butwithoutrelyingonanysyntacticparsetree.TheCoRAEmodelconsistsincombiningrecursivelyeachwordwithitscontextwords(neighbors'words:previousandnext)byconsideringthewordorder.Asupportvectormachineclassifierwithfine-tuningtechniqueisalsousedtoshowthatourdeepcompositionalrepresentationmodelCoRAEimprovessignificantlytheaccuracyofsentimentanalysistask.Experimentalresultsdemon-stratethatCoRAEremarkablyoutperformsseveralcompetitivebaselinemethodsontwodatabases,namely,SanderstwittercorpusandFacebookcommentscorpus.TheCoRAEmodelachievesanefficiencyof83.28%withtheFacebookdatasetand97.57%withtheSandersdataset.
简介:Amethod,calledTwo-DimensionalExtendedAttributeGrammars(2-DEAGs)fortherecognitionofhand-printedChinesecharactersispresented.Thismethodusesdirectlytwodimensionalinformation,andprovidesaschemefordealingwithvariouskindsofspecificcasesinauniformway.Inthismethod,componentsaredrawninguidedandredundantwayandreductionsaremadelevelbylevejustinaccordancewiththecomponentcombinationrelationsofChinesecharacters.Themethodprovidsalsopolysemousgrammars,coexistinggrammarsandstructureinferringswhihconstrainredundantrecognitionbycomparisonamongsimilarcharactersofcomponentsandgreatlyincreasethetoleranceabilitytodistortion.
简介:Inthispaper,ageometry-basedpointcloudreductionmethodisproposed,andareal-timemobileaugmentedrealitysystemisexploredforapplicationsinurbanenvironments.Weformulateanewobjectivefunctionwhichcombinesthepointreconstructionerrorsandconstraintsonspatialpointdistribution.Basedonthisformulation,amixedintegerprogrammingschemeisutilizedtosolvethepointsreductionproblem.Themobileaugmentedrealitysystemexploredinthispaperiscomposedoftheofttineandonlinestages.Attheofflinestage,webuildupthelocalizationdatabaseusingstructurefrommotionandcompressthepointcloudbytheproposedpointcloudreductionmethod.Whileattheonlinestage,wecomputethecameraposeinrealtimebycombininganimage-basedlocalizationalgorithmandacontinuousposetrackingalgorithm.Experimentalresultsonbenchmarkandrealdatashowthatcomparedwiththeexistingmethods,thisgeometry-basedpointcloudreductionmethodselectsapointcloudsubsetwhichhelpstheimage-basedlocalizationmethodtoachievehighersuccessrate.Also,theexperimentsconductedonamobileplatformshowthatthereducedpointcloudnotonlyreducesthetimeconsumingforinitializationandre-initialization,butalsomakesthememoryfootprintsmall,resultingascalableandreal-timemobileaugmentedrealitysystem.
简介:Inthispaperanewtext-independentspeakerverificationmethodGSMSVisproposedbasedonlikelihoodscorenormalization.Inthisnovelmethodaglobalspeakermodelisestablishedtorepresenttheuniversalfeaturesofspeechandnormalizethelikelihoodscore.Statisticalanalysisdemonstratesthatthisnormalizationmethodcanremovecommonfactorsofspeechandbringthedifferencesbetweenspeakersintoprominence.Asaresulttheequalerrorrateisdecreasedsignificantly,verificationprocedureisacceleratedandsystemadaptabilitytospeakingspeedisimproved.
简介:Animportantconceptproposedintheearlystageofrobotpathplanningfieldistheshrinkingofarobottoapointandmeanwhiletheexpandingofobstaclesintheworkspaceasasetofnewobstacles.TheresultinggrownobstaclesarecalledtheConfigurationSpace(Cspace)obstacles.Thefind-pathproblemisthentransformedintothatoffindingacollision-freepathforapointrobotamongtheCspaceobstacles.However,theresearchexperienceshaveshownthattheCspacetransformationisveryhardwhenthefollowingsituationsoccur:1)boththerobotandobstaclesarenotpolygons,and2)therobotisallowedtorotate.Thissituationgetsevenworsewhentherobotandobstaclesarethreedimensional(3D)objectswithvariousshapes.Forthisreason,directpathplanningapproacheswithouttheCspacetransformationisquiteusefulandexpected.Motivatedbythepracticalrequirementsofrobotpathplanning,ageneralizedconstrainedoptimizationproblem(GCOP)withnotonlylogicANDbutalsologicORrelationshipswasproposedandamathematicalsolutiondevelopedpreviously.Thispaperinheritsthefundamentalideasofinequalityandoptimizationtechniquesfromthepreviouswork,convertstheobstacleavoidanceproblemintoasemi-infiniteconstrainedoptimizationproblemwiththehelpofthemathematicaltransformation,andproposesadirectpathplanningapproachwithoutCspacecalculation,whichisquitedifferentfromtraditionalmethods.Toshowitsmerits,simulationresultsin3Dspacehavebeenpresented.