多线性子空间可伸缩视频编码方法-豆柴文库

您所在位置：网站首页 / 多线性子空间可伸缩视频编码方法.docx / 文档详情

在线预览结束，喜欢就下载吧，查找使用更方便

5 金币

下载文档

/ 2

下载提示

如果您无法下载资料，请参考说明：

1、部分资料下载需要金币，请确保您的账户上有足够的金币

2、已购买过的文档，再次下载不重复扣费

3、资料包下载后请先用软件解压，在使用对应软件打开

文本预览

多线性子空间可伸缩视频编码方法
Introduction
Multilinearsubspacescaling(MSS)isatechniqueforvideocompressionthatisspecificallyoptimizedforvideoswithalow-rankstructure.Inthispaper,wepresentanoverviewoftheMSSvideoencodingmethod.WedescribethemathematicalbackgroundofMSS,andcompareittoothercommonlyusedvideocompressiontechniques.WealsopresentexperimentalresultsthatdemonstratetheeffectivenessofMSSinreducingthesizeofvideodata.
Background
Regularvideocompressiontechniquesrelyonatransformationoftheindividualframesorgroupsofframesintoasparserepresentationthatcanbeefficientlycoded.Thisapproachworkswellforvideoswheretheinformationisspreadoutacrossmanyindependentframes,butcanbelesseffectiveforvideosthathavealow-dimensionalstructure,suchasvideosofrotatingobjectsorvideoswithafixedcameraviewpoint.
MSSisarecenttechniquethataddressesthisproblembyexploitingthelow-rankstructureofvideos.Inthecontextofvideocompression,avideosequencecanbeviewedasatensor(i.e.amultidimensionalarray)ofpixelvalues.Ifthevideohasalow-rankstructure,i.e.ifthetensorcanbeapproximatelyexpressedasaproductoflower-dimensionalmatrices,thentheMSSmethodcanexploitthisstructuretoachievebettercompression.
Method
ThebasicideabehindMSSistofactorizethetensorrepresentingthevideosequenceintoasetoflower-dimensionalmatrices,andthentoquantizeandencodethesematricesusingconventionalvideocompressiontechniques.
Tofactorizethetensor,MSSusestheTuckerdecomposition,atechniquethatinvolvesorthogonalizingthetensoralongeachmode(i.e.eachdimensionofthetensor)andthenprojectingitontoalower-dimensionalsubspace.Theresultisasetofcorematrices,andasetoffactormatrices,whichdescribehowthetensorcanbereconstructedfromthecorematrices.
Oncethetensorhasbeenfactorized,thefactormatricescanbecompressedusingconventionalvideocompressiontechniquessuchasH.264orHEVC.However,becausethefactormatricesarelow-dimensional(typically2Dor3D),theycanbecompressedmuchmoreefficientlythantheoriginalhigh-dimensionaltensor.
ExperimentalResults
WeevaluatedtheperformanceoftheMSSmethodonasetofvideoswithv