多尺分割原理與應(yīng)用

HierarchicalSegmentationReadingReportHSeg當(dāng)前1頁(yè)，總共19頁(yè)。CONTENTSPARTONEAbouttheIntroduction.PARTTWOTheHsegSegmentation.PARTTHREEApplicationinENVI.當(dāng)前2頁(yè)，總共19頁(yè)。1PARTONEAbouttheIntroduction.當(dāng)前3頁(yè)，總共19頁(yè)。AbstractRecentadvancesinspectral–spatialclassificationofhyperspectralimagesarepresentedinthispaper.Severaltechniquesareinvestigatedforcombiningbothspatialandspectralinformation.Highlighttheimportanceofspectral–spatialstrategiesfortheaccurateclassificationofhyperspectralimagesandvalidatetheproposedmethods.當(dāng)前4頁(yè)，總共19頁(yè)。IntroductionAdvantagesofhyperspectralremotesensortechnology:Thedetailedspectralinformationincreasesthepossibilityofmoreaccuratelydiscriminatingmaterialsofinterest.Thefinespatialresolutionofthesensorsenablestheanalysisofsmallspatialstructuresintheimage.Manyoperationalimagingsystemsarecurrentlyavailableprovidingalargeamountofimagesforvariousthematicapplications.But,italsobringssomeproblem:theCurseofDimensionalityandtheneedforspecificspectral–spatialclassifiers.PS:維數(shù)災(zāi)難(CurseofDimensionality):通常是指在涉及到向量的計(jì)算的問(wèn)題中，隨著維數(shù)的增加，計(jì)算量呈指數(shù)倍增長(zhǎng)的一種現(xiàn)象。維數(shù)災(zāi)難在很多學(xué)科中都可以碰到，比如動(dòng)態(tài)規(guī)劃，模式識(shí)別和影像識(shí)別等。當(dāng)前5頁(yè)，總共19頁(yè)。IntroductionTheCurseofDimensionalityofhyperspectralremotesensortechnology:Inhigh-dimensionalspaces,normallydistributeddatahaveatendencytoconcentrateinthetails,

whichseemstobecontradictorywithitsbell-shapeddensityfunction.therateofconvergenceofthestatisticalestimationdecreaseswhenthedimensiongrowswhileconjointlythenumberofparameterstoestimateincreases,makingtheestimationofthemodelparametersverydifficult.withalimitedtrainingset,beyondacertainlimit,theclassificationaccuracyactuallydecreasesasthenumberoffeaturesincreases.當(dāng)前6頁(yè)，總共19頁(yè)。IntroductionHowtobuildaccurateclassifiersforhyperspectralimages?SVMsperformanonlinearpixel-wiseclassificationbasedonthefullspectralinformationwhichisrobusttothespectraldimensionofhyperspectralimages.IterativestatisticalclassifierbasedonMarkovrandomfield(MRF)modeling.NotethatrecentlyadaptiveMRFhavebeenintroducedinremotesensing.Useadvancedmorphologicalfiltersasanalternativewayofperformingjointclassification.PS:魯棒性(Robust):即系統(tǒng)的健壯性，是在異常和危險(xiǎn)情況下系統(tǒng)生存的關(guān)鍵。當(dāng)前7頁(yè)，總共19頁(yè)。Introduction當(dāng)前8頁(yè)，總共19頁(yè)。2PARTTWOTheHsegSegmentation.當(dāng)前9頁(yè)，總共19頁(yè)。TheHsegSegmentationTheHSegalgorithmisasegmentationtechniquecombiningregiongrowing,usingthehierarchicalstepwiseoptimization(HSWO)method,whichproducesspatiallyconnectedregions,withunsupervisedclassification,thatgroupstogethersimilarspatiallydisjointregions.Thealgorithmcanbesummarizedasfollows.Initialization:Initializethesegmentationbyassigningeachpixelaregionlabel.Ifapresegmentationisprovided,labeleachpixelaccordingly.Otherwise,labeleachpixelasaseparateregion.當(dāng)前10頁(yè)，總共19頁(yè)。TheHsegSegmentationIIIIIIIVCalculatethedissimilaritycriterionvaluebetweenallpairsofspatiallyadjacentregionsStopifconvergenceisachieved.Otherwise,returntostepIMergespatiallynonadjacentregionsMergespatiallyadjacentregions當(dāng)前11頁(yè)，總共19頁(yè)。TheHsegSegmentationI計(jì)算每一對(duì)空間相鄰區(qū)域的相異準(zhǔn)則值（dissimilaritycriterionvalue），例如計(jì)算向量?；蛘邊^(qū)域平均向量間的光譜角填圖SAM。SAM是基于物理的一種光譜分類，利用n維角度來(lái)匹配兩組像元光譜，將光譜看成是維數(shù)與波段數(shù)相等的空間里的向量，計(jì)算光譜間的角度的算法，決定了兩個(gè)光譜之間的相似性，也就是通過(guò)計(jì)算兩向量之間的廣義角來(lái)確定它們的相似性（角度越小越相似）。這種方法充分利用了光譜維的信息。當(dāng)前12頁(yè)，總共19頁(yè)。TheHsegSegmentationII找到最小的相異準(zhǔn)則值并設(shè)為閾值，如果兩相鄰區(qū)域的相異準(zhǔn)則值等于閾值則合并這兩個(gè)區(qū)域。即dissim_val=thresh_val時(shí)，合并兩相鄰區(qū)域。當(dāng)前13頁(yè)，總共19頁(yè)。TheHsegSegmentationIII如果Swght>0.0，則將dissim_val<Swght*thresh_val的不相鄰區(qū)域合并（Swght是衡量基于光譜信息的聚類相對(duì)于區(qū)域增長(zhǎng)的相對(duì)重要性的選擇性參數(shù)，當(dāng)Swght=0.0時(shí)，僅空間上相鄰的區(qū)域才允許合并，

當(dāng)0<Swght<1時(shí)，則空間相鄰區(qū)域相對(duì)于不相鄰區(qū)域有一1/Swght的系數(shù)更可能被合并）當(dāng)前14頁(yè)，總共19頁(yè)。TheHsegSegmentationIV如果合并完成則結(jié)束，否則回到步驟I。通過(guò)迭代運(yùn)算，最終形成從細(xì)到粗的多尺度、多層次的分割結(jié)果。當(dāng)前15頁(yè)，總共19頁(yè)。TheHsegSegmentation因?yàn)橐獙?duì)空間上非相鄰的區(qū)域進(jìn)行合并，運(yùn)算量變得非常巨大。為解決運(yùn)算量問(wèn)題，發(fā)展出了迭代分而治之的多尺度分割近似計(jì)算及其相應(yīng)的有效實(shí)現(xiàn)方式。多尺度分割能輸出一個(gè)從初始化分割直至一個(gè)區(qū)域的分割的多尺度序列，在這個(gè)序列中，一個(gè)特定的對(duì)象既可以表示成幾個(gè)區(qū)域，從而具有較好的細(xì)節(jié)信息，也可以與其他對(duì)象一起被一個(gè)區(qū)域吸收。在實(shí)際應(yīng)用中，可以通過(guò)HSegViewer程序交互式的選擇適當(dāng)?shù)乃交蛘哌x擇其他為