简介:语义图象分割是一项任务为每个图象象素预言一个范畴标签。它的关键挑战是设计一个强壮的特征代表。在这份报纸,我们作为特征表示熔化神经网络(CNN)展示的层次convolutional和基于区域的特征。层次特征包含更全球的信息,当基于区域的特征包含更多的本地信息时。这些二种特征的联合显著地提高特征表示。然后,熔化特征被用来训练一个softmax分类器生产每象素标签任务概率。并且一块充分连接的有条件的随机的地(CRF)被用作一个processing以后方法改进标记的一致性。我们进行实验在上筛流动数据集。象素精确性和班精确性分别地是84.4%和34.86%。
简介:Analgorithmforfacedescriptionandrecognitionbasedonmulti-resolutionwithmulti-scalelocalbinarypattern(multi-LBP)featuresisproposed.Thefacialimagepyramidisconstructedandeachfacialimageisdividedintovariousregionsfromwhichpartialandholisticlocalbinarypatter(LBP)histogramsareextracted.AllLBPfeaturesofeachimageareconcatenatedtoasingleLBPeigenvectorwithdifferentresolutions.ThedimensionalityofLBPfeaturesisthenreducedbyalocalmarginalignment(LMA)algorithmbasedonmanifold,whichcanpreservethebetween-classvariance.Supportvectormachine(SVM)isappliedtoclassifyfacialimages.ExtensiveexperimentsonORLandCMUfacedatabasesclearlyshowthesuperiorityoftheproposedschemeoversomeexistedalgorithms,especiallyontherobustnessofthemethodagainstdifferentfacialexpressionsandposturesofthesubjects.
简介:ThisletterstudiesonthedetectionoftexturefeaturesinSyntheticApertureRadar(SAR)images.ThroughanalyzingthefeaturedetectionmethodproposedbyLopes,animprovedtexturedetectionmethodisproposed,whichcannotonlydetecttheedgeandlinesbutalsoavoidstretchingedgeandsuppressinglinesoftheformeralgorithm.ExperimentalresultswithbothsimulatedandrealSARimagesverifytheadvantageandpracticabilityoftheimprovedmethod.
简介:Inthisresearch,acontent-basedimageretrieval(CBIR)systemforhighresolutionsatelliteimageshasbeendevelopedbyusingtexturefeatures.Theproposedapproachusesthelocalbinarypattern(LBP)texturefeatureandablockbasedscheme.Thequeryanddatabaseimagesaredividedintoequallysizedblocks,fromwhichLBPhistogramsareextracted.TheblockhistogramsarethencomparedbyusingtheChi-squaredistance.ExperimentalresultsshowthattheLBPrepresentationprovidesapowerfultoolforhighresolutionsatelliteimages(HRSI)retrieval.
简介:ThepaperaddressestheproblemoftargetrecognitionusingHigh-resolutionRadarRangeProfiles(HRRP).Anovelapproachoffeatureextractionanddimensionreductionbasedonextendedhighordercentralmomentsisproposedinordertoreducethedimensionofrangeprofiles.FeaturesextractedfromradarHRRPsarenormalizedandsmoothed,andthencomparativeanalysisofthesimilarapproachesisdone.Therangeprofilesareobtainedbystepfrequencytechniqueusingthetwo-dimensionalbackscattersdistributiondataoffourdifferentaircraftmodels.Thetemplatematchingmethodbynearestneighborrules,whichisbasedonthetheoryofkernelmethodsforpatternanalysis,isusedtoclassifyandidentifytherangeprofilesfromfourdifferentaircrafts.Numericalsimulationresultsshowthattheproposedapproachcanachievegoodperformanceofstability,shiftindependenceandhigherrecognitionrate.Itishelpfulforreal-timeidentificationandtheengineeringimplementsofautomatictargetrecognitionusingHRRP.Thenumberofrequiredtemplatescouldbereducedcon-siderablywhilemaintaininganequivalentrecognitionrate.
简介:Handwrittensignaturerecognitionispresentedbasedonananglefeaturevectorbyusingtheartificialneuralnetwork(ANN)inthisresearch.Eachsignatureimagewillberepresentedbyananglevector.ThefeaturevectorwillconstitutetheinputtotheANN.Thecollectionofsignatureimagesisdividedintotwosets.OnesetwillbeusedfortrainingtheANNinasupervisedfashion.TheothersetwhichisneverseenbytheANNwillbeusedfortesting.Aftertraining,theANNwillbetestedbyrecognizingthesignatures.Whenasignatureisclassifiedcorrectly,itisconsideredcorrectrecognition,otherwiseitisafailure.Theachievedrecognitionrateofthissystemis94%.
简介:TosolvetheproblemsoftheAMR-WB+(ExtendedAdaptiveMulti-Rate-WideBand)semi-open-loopcodingmodeselectionalgorithm,featuresforACELP(AlgebraicCodeExcitedLinearPrediction)andTCX(TransformCodedeXcitation)classificationareinvestigated.11classifyingfeaturesintheAMR-WB+codecareselectedand2novelclassifyingfeatures,i.e.,EFM(EnergyFlatnessMeasurement)andstdEFM(standarddeviationofEFM),areproposed.Consequently,anovelsemi-open-loopmodeselectionalgorithmbasedonEFMandselectedAMR-WB+featuresisproposed.TheresultsofclassifyingtestandlisteningtestshowthattheperformanceofthenovelalgorithmismuchbetterthanthatoftheAMR-WB+semi-open-loopcodingmodeselectionalgorithm.