PDF Graph spectral image processing gene cheung download

Page 1


Graph Spectral Image Processing Gene Cheung

Visit to download the full and correct content document: https://ebookmass.com/product/graph-spectral-image-processing-gene-cheung/

More products digital (pdf, epub, mobi) instant download maybe you interests ...

Digital Image Processing Using MATLAB Rafael C. Gonzalez

https://ebookmass.com/product/digital-image-processing-usingmatlab-rafael-c-gonzalez/

Feature extraction and image processing for computer vision Fourth Edition Aguado

https://ebookmass.com/product/feature-extraction-and-imageprocessing-for-computer-vision-fourth-edition-aguado/

Methods and techniques for fire detection : signal, image and video processing perspectives 1st Edition Çetin

https://ebookmass.com/product/methods-and-techniques-for-firedetection-signal-image-and-video-processing-perspectives-1stedition-cetin/

Digital Image Processing 3rd Edition Rafael C. Gonzalez And Richard E. Woods

https://ebookmass.com/product/digital-image-processing-3rdedition-rafael-c-gonzalez-and-richard-e-woods/

Design for Embedded Image Processing on FPGAs, 2nd Edition Donald G. Bailey

https://ebookmass.com/product/design-for-embedded-imageprocessing-on-fpgas-2nd-edition-donald-g-bailey/

Machine Learning Algorithms for Signal and Image Processing Suman Lata Tripathi

https://ebookmass.com/product/machine-learning-algorithms-forsignal-and-image-processing-suman-lata-tripathi/

Image Processing for Automated Diagnosis of Cardiac Diseases Rajeev Kumar Chauhan (Editor)

https://ebookmass.com/product/image-processing-for-automateddiagnosis-of-cardiac-diseases-rajeev-kumar-chauhan-editor/

Machine Intelligence, Big Data Analytics, and IoT in Image Processing Ashok Kumar

https://ebookmass.com/product/machine-intelligence-big-dataanalytics-and-iot-in-image-processing-ashok-kumar/

Machine vision inspection systems. Vol.1: image processing, concepts Malarvel M (Ed.)

https://ebookmass.com/product/machine-vision-inspection-systemsvol-1-image-processing-concepts-malarvel-m-ed/

Spectral Image Processing

Graph

SCIENCES

Image, Field Director – Laure Blanc-Feraud

Compression, Coding and Protection of Images and Videos, Subject Head – Christine Guillemot

Graph Spectral Image Processing

Apart from any fair dealing for the purposes of research or private study, or criticism or review, as permitted under the Copyright, Designs and Patents Act 1988, this publication may only be reproduced, stored or transmitted, in any form or by any means, with the prior permission in writing of the publishers, or in the case of reprographic reproduction in accordance with the terms and licenses issued by the CLA. Enquiries concerning reproduction outside these terms should be sent to the publishers at the undermentioned address:

ISTE Ltd

John Wiley & Sons, Inc.

27-37 St George’s Road 111 River Street London SW19 4EU Hoboken, NJ 07030

UK USA

www.iste.co.uk

www.wiley.com

© ISTE Ltd 2021

The rights of Gene Cheung and Enrico Magli to be identified as the author of this work have been asserted by them in accordance with the Copyright, Designs and Patents Act 1988.

Library of Congress Control Number: 2021932054

British Library Cataloguing-in-Publication Data

A CIP record for this book is available from the British Library

ISBN 978-1-78945-028-6

ERC code:

PE7 Systems and Communication Engineering

PE7_7 Signal processing

Chapter1.GraphSpectralFiltering .....................3 YuichiTANAKA

1.1.Introduction..................................3

1.2.Review:filteringoftime-domainsignals..................4

1.3.Filteringofgraphsignals..........................5

1.3.1.Vertexdomainfiltering........................6

1.3.2.Spectraldomainfiltering.......................8

1.3.3.Relationshipbetweengraphspectralfilteringandclassical filtering.................................10

1.4.Edge-preservingsmoothingofimagesasgraphspectralfilters.....11

1.4.1.Earlyworks...............................11

1.4.2.Edge-preservingsmoothing......................12

1.5.Multiplegraphfilters:graphfilterbanks..................15

1.5.1.Framework...............................16

1.5.2.Perfectreconstructioncondition...................17

1.6.Fastcomputation...............................20

1.6.1.Subdivision...............................20

1.6.2.Downsampling.............................21

1.6.3.PrecomputingGFT...........................22

1.6.4.Partialeigendecomposition......................22

1.6.5.Polynomialapproximation......................23

1.6.6.Krylovsubspacemethod........................26

1.7.Conclusion..................................26 1.8.References...................................26

Chapter2.GraphLearning ...........................31

XiaowenD ONG ,DorinaT HANOU ,MichaelR ABBAT andPascalF ROSSARD

2.1.Introduction..................................31

2.2.Literaturereview...............................33

2.2.1.Statisticalmodels............................33

2.2.2.Physicallymotivatedmodels.....................35

2.3.Graphlearning:asignalrepresentationperspective............36

2.3.1.Modelsbasedonsignalsmoothness.................38

2.3.2.Modelsbasedonspectralfilteringofgraphsignals.........43

2.3.3.Modelsbasedoncausaldependenciesongraphs..........48

2.3.4.Connectionswiththebroaderliterature...............50

2.4.Applicationsofgraphlearninginimageprocessing...........52

2.5.Concludingremarksandfuturedirections.................55

2.6.References...................................57

Chapter3.GraphNeuralNetworks

3.1.Introduction..................................63

3.2.Spectralgraph-convolutionallayers....................64

3.3.Spatialgraph-convolutionallayers.....................66

3.4.Concludingremarks.............................71 3.5.References...................................72

Chapter4.GraphSpectralImageandVideoCompression

HilmiE.E GILMEZ ,Yung-HsuanC HAO andAntonioO RTEGA

4.1.Introduction..................................75

4.1.1.Basicsofimageandvideocompression...............77

4.1.2.Literaturereview............................78

4.1.3.Outlineofthechapter.........................79

4.2.Graph-basedmodelsforimageandvideosignals.............79

4.2.1.Graph-basedmodelsforresidualsofpredictedsignals.......81

4.2.2.DCT/DSTsasGFTsandtheirrelationto1Dmodels........87

4.2.3.Interpretationofgraphweightsforpredictivetransformcoding..88

4.3.Graphspectralmethodsforcompression..................89

4.3.1.GL-GFTdesign.............................89

4.3.2.EA-GFTdesign.............................92

4.3.3.EmpiricalevaluationofGL-GFTandEA-GFT...........97

4.4.Conclusionandpotentialfuturework...................100

4.5.References...................................101

Chapter5.GraphSpectral3DImageCompression ...........105

5.1.Introductionto3Dimages..........................106

5.1.1.3Dimagedefinition..........................106

5.1.2.Pointcloudsandmeshes........................106

5.1.3.Omnidirectionalimages........................107

5.1.4.Lightfieldimages...........................109

5.1.5.Stereo/multi-viewimages.......................110

5.2.Graph-based3Dimagecoding:overview.................110

5.3.Graphconstruction..............................115

5.3.1.Geometry-basedapproaches.....................117

5.3.2.Jointgeometryandcolor-basedapproaches.............121

5.3.3.Separabletransforms..........................125

5.4.Concludingremarks.............................126

5.5.References...................................128

Chapter6.GraphSpectralImageRestoration ..............133 JiahaoPANG andJinZ ENG

6.1.Introduction..................................133

6.1.1.Asimpleimagedegradationmodel..................133

6.1.2.Restorationwithsignalpriors.....................135

6.1.3.Restorationviafiltering........................137

6.1.4.GSPforimagerestoration.......................140

6.2.Discrete-domainmethods..........................141

6.2.1.Non-localgraph-basedtransformfordepthimagedenoising...141

6.2.2.DoublystochasticgraphLaplacian..................142

6.2.3.Reweightedgraphtotalvariationprior................145

6.2.4.LefteigenvectorsofrandomwalkgraphLaplacian.........150

6.2.5.Graph-basedimagefiltering......................155

6.3.Continuous-domainmethods........................155

6.3.1.Continuous-domainanalysisofgraphLaplacianregularization..156

6.3.2.Low-dimensionalmanifoldmodelforimagerestoration......163

6.3.3.LDMMasgraphLaplacianregularization..............165

6.4.Learning-basedmethods...........................167

6.4.1.CNNwithGLR.............................169

6.4.2.CNNwithgraphwaveletfilter....................171

6.5.Concludingremarks.............................172

6.6.References...................................173

Chapter7.GraphSpectralPointCloudProcessing ...........181

WeiH U ,SihengC HEN andDongT IAN

7.1.Introduction..................................181

7.2.Graphandgraph-signalsinpointcloudprocessing............183

7.3.Graphspectralmethodologiesforpointcloudprocessing........185

7.3.1.Spectral-domaingraphfilteringforpointclouds..........185

7.3.2.Nodal-domaingraphfilteringforpointclouds...........188

7.3.3.Learning-basedgraphspectralmethodsforpointclouds......189

7.4.Low-levelpointcloudprocessing......................190

7.4.1.Pointclouddenoising.........................191

7.4.2.Pointcloudresampling........................193

7.4.3.Datasetsandevaluationmetrics....................198

7.5.High-levelpointcloudunderstanding...................199

7.5.1.Dataauto-encodingforpointclouds.................199

7.5.2.Transformationauto-encodingforpointclouds...........206

7.5.3.ApplicationsofGraphTERinpointclouds.............211

7.5.4.Datasetsandevaluationmetrics....................211

7.6.Summaryandfurtherreading........................213

7.7.References...................................214

Chapter8.GraphSpectralImageSegmentation

MichaelN G

8.1.Introduction..................................221

8.2.Pixelmembershipfunctions.........................222

8.2.1.Two-classproblems..........................222

8.2.2.Multiple-classproblems........................226

8.2.3.Multipleimages............................227

8.3.Matrixproperties...............................230

8.4.Graphcuts...................................232

8.4.1.TheMumford–Shahmodel......................234

8.4.2.Graphcutsminimization.......................235

8.5.Summary...................................237

8.6.References...................................237

Chapter9.GraphSpectralImageClassification

MinxiangY E ,VladimirS TANKOVIC ,LinaS TANKOVIC andGeneC HEUNG

9.1.Formulationofgraph-basedclassificationproblems...........243

9.1.1.Graphspectralclassifierswithnoiselesslabels...........243

9.1.2.Graphspectralclassifierswithnoisylabels.............246

9.2.Towardpracticalgraphclassifierimplementation.............247

9.2.1.Graphconstruction...........................247

9.2.2.Experimentalsetupandanalysis...................249

9.3.Featurelearningviadeepneuralnetwork.................255

9.3.1.Deepfeaturelearningforgraphconstruction............258

9.3.2.Iterativegraphconstruction......................260

9.3.3.Towardpracticalimplementationofdeepfeaturelearning.....262

9.3.4.Analysisoniterativegraphconstructionforrobustclassification.267

9.3.5.Graphspectrumvisualization.....................269

9.3.6.Classificationerrorratecomparisonusinginsufficienttraining data....................................270

9.3.7.Classificationerrorratecomparisonusingsufficienttraining datawithlabelnoise..........................270

9.4.Conclusion..................................271

9.5.References...................................272 Chapter10.GraphNeuralNetworksforImageProcessing

10.1.Introduction.................................277

10.2.Supervisedlearningproblems.......................278

10.2.1.Pointcloudclassification.......................278

10.2.2.Pointcloudsegmentation......................281 10.2.3.Imagedenoising............................283

10.3.Generativemodelsforpointclouds....................286

10.3.1.Pointcloudgeneration........................286

10.3.2.Shapecompletion...........................291

10.4.Concludingremarks............................294

10.5.References..................................294

IntroductiontoGraphSpectral ImageProcessing

GeneC HEUNG 1 andEnricoM AGLI 2

1 YorkUniversity,Toronto,Canada

2 PolitecnicodiTorino,Turin,Italy

I.1.Introduction

Imageprocessing isamatureresearchtopic.ThefirstspecificationofJoint PhotographicExpertsGroup(JPEG),nowthepredominantimagecodingstandardon theInternet,waspublishedin1992.MPEG1,thefirstdigitalvideocompression standardbyISO,wasstandardizedin1993.TheIEEEInternationalConferenceon ImageProcessing(ICIP),theflagshipimageprocessingconferenceheldannuallyfor theIEEESignalProcessingSociety(SPS),wasalsostartedin1993andhasbeenin existencefor27years,makingitolderthanmanyimageprocessingresearchersnow studyingingraduateschools!Giventhetopic’smaturity,itisalegitimatequestionto askifyetanotherbookonimageprocessingiswarranted.Asco-editorsofthisbook, weemphaticallyanswerthisquestionwitharesounding“Yes”.Wewillfirstdiscuss thefollowingrecenttechnologicaltrends,whichalsoserveasmotivationsforthe creationofthispublication.

1) SensingandDisplayTechnologies:Theadventofimagesensingtechnologies, suchasactivedepthsensorsanddisplaytechnologieslikehead-mounteddisplays (HMD),inthelastdecadealone,meansthatthenatureofadigitalimagehas drasticallychanged.Beyondhigherspatialresolutionandbit-depthperpixel,a modernimagingsensorcanalsoacquirescenedepth,hyper-spectralproperties,etc. Further,oftenacquiredimagedataisnotrepresentedasatraditional2Darrayofpixel information,butinanalternativeform,suchaslightfieldsand3Dpointclouds.This meansthattheprocessingtoolsmustflexiblyadapttoricherandevolvingimaging contentsandformats.

, coordinatedbyGeneC HEUNG andEnricoM AGLI .©ISTELtd2021

GraphSpectralImageProcessing

2) GraphSignalProcessing:Inthelasteightyears,wehavealsowitnessedthe birthofanewsignalprocessingtopic–called graphsignalprocessing (GSP)–that generalizestraditionalmathematicaltoolsliketransformsandwavelets,toprocess signalsresidingonirregulardatakernelsdescribedbygraphs(Shuman etal.2013). CentraltoGSPisthenotionof graphfrequencies:orthogonalcomponents,computed fromagraphvariationoperatorlikethegraphLaplacianmatrix,thatgeneralizethe notionofFouriermodestothegraphdomain,spanningagraphsignalspace.Because ofitsinherentpowerfulgenerality,onecaneasilyadoptordesignGSPtoolsfor differentimagingapplications,whereanodeinagraphrepresentsapixel,andthe graphconnectivityischosentoreflectinter-pixelsimilaritiesorcorrelations.Foran exampleoftheGSPtoolbeingusedforimagerestoration,seeFigureI.1foran illustrationofagraphspectralmethodcalledlefteigenvectorsoftherandomwalk graphLaplacian(LeRAG)forJPEGimagedequantization(Liu etal. 2017).GSPtools canalsoeasilyadapttotheaforementionedmodernimagingmodalities,suchaslight fieldimagesand3Dpointclouds,thatdonotresideonregular2Dgrids.

FigureI.1. VisualcomparisonofJPEGdequantizationmethodsforabutterflyat QF=5.ThecorrespondingPSNRvaluesarealsogiven.Foracolorversionofthis figure,seewww.iste.co.uk/cheung/graph.zip

3) DeepNeuralNetworks:Withoutadoubt,thesingularseismicparadigmshift indatascienceinthelastdecadeis deeplearning.Usinglayersofconvolutional filters,pointwisenonlinearitiesandpoolingfunctions,deepneuralnetwork(DNN) architectureslike convolutionalneuralnetworks (CNN)havedemonstratedsuperior performanceinawiderangeofimagingtasksfromdenoisingtoclassification,when alargevolumeoflabeleddataisavailablefortraining(Vemulapalli etal. 2016; Zhang etal. 2017).Whenlabeledtrainingdataisscarce,orwhentheunderlying datakernelisirregular(thuscomplicatingthetrainingofconvolutionalfiltersandthe selectionofpoolingoperators),howtobestdesignandconstructDNNforatargeted imageapplicationisachallengingproblem.Moreover,aCNNpurelytrainedfrom labeleddataoftenremainsa“blackbox”,i.e.thelearnedoperatorslikefilteringremain unexplainable.

Motivatedbythesetechnologicaltrends,wehavefocusedthisbookonthetheory andapplicationsofGSPtoolsforimageprocessing,coveringconventionalimages andvideos,newmodalitieslikelightfieldsand3Dpointclouds,andhybrid GSP/deeplearningapproaches.Differentfromothergraph-basedimageprocessing books(LezorayandGrady2012),weconcentrateon spectral processingtechniques withfrequencyinterpretationssuchasgraphFouriertransforms(GFT)andgraph wavelets,drawinginspirationfromthelonghistoryoffrequencyanalysistoolsin traditionalsignalprocessing. Graphfrequencyanalysis enablesthedefinitionof familiarsignalprocessingnotions,suchasgraphFouriermodes,bandlimitedness, andsignalsmoothness,usinggraphspectralmethodsthatcanbedesigned.

Specifically,thecontentofthisbookisstructuredintotwoparts:

1)ThefirstpartofthebookdiscussesthefundamentalGSPtheories.Chapter1, titled“GraphSpectralFiltering”byY.Tanaka,reviewsthebasicsofgraphfiltering suchasgraphtransformsandwavelets.Chapter2,titled“GraphLearning”byX. Dong,D.Thanou,M.RabbatandP.Frossard,reviewsrecenttechniquestolearn anunderlyinggraphstructuregivenasetofobservabledata.Chapter3,titled “GraphNeuralNetworks”byG.FracastoroandD.Valsesia,overviewsrecentworks generalizingDNNarchitecturestothegraphdatadomain,whereinputsignalsreside onirregulargraphstructures.

2)ThesecondpartofthebookreviewsdifferentimagingapplicationsofGSP. Chapters4and5,titled“GraphSpecralImageandVideoCompression”byH.E. Egilmez,Y.-H.ChaoandA.Ortegaand“GraphSpectral3DImageCompression” byT.Maugey,M.Rizkallah,N.M.Bidgoli,A.RoumyandC.Guillemot,focusonthe designandapplicationsofGSPtoolsforthecompressionoftraditionalimages/videos and3Dimages,respectively.Chapter6,titled“GraphSpectralImageRestoration”by J.PangandJ.Zeng,focusesonthegeneralrecoveryofcorruptedimages,e.g.image denoisinganddeblurring.Asanewimagingmodality,Chapter7,titled“Graph SpectralPointCloudProcessing”byW.Hu,S.ChenandD.Tian,focusesonthe processingof3Dpointcloudsforapplications,suchaslow-levelrestorationand high-levelunsupervisedfeaturelearning.Chapters8and9,titled“GraphSpectral ImageSegmentation”byM.Ngand“GraphSpectralImageClassification”byM.Ye, V.Stankovic,L.StankovicandG.Cheung,narrowthediscussionspecificallyto segmentationandclassification,respectively,twopopularresearchtopicsinthe computervisioncommunity.Finally,Chapter10,titled“GraphNeuralNetworksfor ImageProcessing”byG.FracastoroandD.Valsesia,reviewsthegrowingeffortsto employrecentGNNarchitecturesforconventionalimagingtaskssuchasdenoising.

Beforewejumpintothevariouschapters,webeginwiththebasicdefinitionsin GSPthatwillbeusedthroughoutthebook.Specifically,weformallydefineagraph, graphspectrum,variationoperatorsandgraphsignalsmoothnesspriorsinthe followingsections.

I.2.Graphdefinition

Agraph G (V , E , W ) containsaset V of N nodesandaset E of M edges.While directedgraphsarealsopossible,inthisbookwemorecommonlyassumean undirectedgraph,whereeachexistingedge (i,j ) ∈E isundirectedandcontainsan edgeweight wi,j ∈ R,whichistypicallypositive.Alargepositiveedgeweight wi,j wouldmeanthatsamplesatnodes i and j areexpectedtobesimilar/correlated.

Therearemanywaystocomputeappropriateedgeweights.Especiallycommon forimages,edgeweight wi,j canbecomputedusingaGaussiankernel,asdoneinthe bilateralfilter (TomasiandManduchi1998):

where li ∈ R2 isthelocationofpixel i onthe2Dimagegrid, xi ∈ R istheintensity ofpixel i,and σ 2 l and σ 2 x aretwoparameters.Hence, 0 ≤ wi,j ≤ 1.Largergeometric and/orphotometricdistancesbetweenpixels i and j wouldmeanasmallerweight wi,j .Edgeweightscanalternativelybedefinedbasedonlocalpixelpatches,features, etc.(Milanfar2013b).Toalargeextent,theappropriatedefinitionofedgeweightis applicationdependent,aswillbediscussedinvariousforthcomingchapters.

A graphsignal x on G isadiscretesignalofdimension N –onesample xi ∈ R foreachnode1 i in V .Assumingthatnodesareappropriatelylabeledfrom 1 to N ,we cansimplytreatagraphsignalasavector x ∈ RN .

I.3.Graphspectrum

Denoteby W ∈ RN ×N an adjacencymatrix,wherethe (i,j )thentryis Wi,j = wi,j .Next,denoteby D ∈ RN ×N adiagonal degreematrix,wherethe (i,i)thentryis Di,i = j Wi,j .A combinatorialgraphLaplacianmatrix L is L = D W (Shuman etal.2013).Because L isrealandsymmetric,onecanshow, viathespectraltheorem,thatitcanbeeigen-decomposedinto:

L = UΛU [I.2] where Λ isadiagonalmatrixcontainingrealeigenvalues λk alongthediagonal,and U isaneigen-matrixcomposedoforthogonaleigenvectors ui ascolumns.Ifalledge

1Ifagraphnoderepresentsapixelinanimage,eachpixelwouldtypicallyhavethreecolor components:red,greenandblue.Forsimplicity,onecantreateachcolorcomponentseparately asadifferentgraphsignal.

weights wi,j arerestrictedtobepositive,thengraphLaplacian L canbeproventobe positivesemi-definite (PSD)(Chung1997)2,meaningthat λk ≥ 0, ∀k and x Lx ≥ 0, ∀x.Non-negativeeigenvalues λk canbeinterpretedas graph frequencies,andeigenvectors U canbeinterpretedascorrespondinggraphFourier modes.Togethertheydefinethe graphspectrum forgraph G .

Thesetofeigenvectors U for L collectivelyformtheGFT(Shuman etal.2013), whichcanbeusedtodecomposeagraphsignal x intoitsfrequencycomponentsvia α = U x.Infact,onecaninterpretGFTasageneralizationofknowndiscrete transformslikethe DiscreteCosineTransform (DCT)(seeShuman etal.2013for details).

Notethatifthemultiplicity mk ofaneigenvalue λk islargerthan1,thenthe setofeigenvectorsthatspanthecorrespondingeigen-subspaceofdimension mk is non-unique.Inthiscase,itisnecessarytospecifythegraphspectrumasthecollection ofeigenvectors U themselves.

Ifwealsoconsidernegativeedgeweights wi,j thatreflectinter-pixel dissimilarity/anti-correlation,thengraphLaplacian L canbeindefinite.Wewill discussafewrecentworks(Su etal. 2017;Cheung etal. 2018)thatemploynegative edgesinlaterchapters.

I.4.Graphvariationoperators

Closelyrelatedtothecombinatorialgraph,Laplacian L,areothervariantsof Laplacianoperators,eachwiththeirownuniquespectralproperties.A normalized graphLaplacian Ln = D 1/2 LD 1/2 isasymmetricnormalizedvariantof L.In contrast,a randomwalkgraphLaplacian Lr = D 1 L isanasymmetricnormalized variantof L.A generalizedgraphLaplacian Lg = L + diag(D) isagraphLaplacian withself-loops di,i atnodes i –calledthe loopygraphLaplacian inDörflerand Bullo(2013)–resultinginageneralsymmetricmatrixwithnon-positive off-diagonalentriesforapositivegraph(Biyikoglu etal. 2005). Eigen-decompositioncanalsobeperformedontheseoperatorstoacquireasetof graphfrequenciesandgraphFouriermodes.Forexample,normalizedvariants Ln and Lr (whicharesimilaritytransformsofeachother)sharethesameeigenvalues between 0 and 2.While L and Ln arebothsymmetric, Ln doesnothavetheconstant vectorasaneigenvector.Asymmetric Lr canbesymmetrizedvialeftandright diagonalmatrixmultiplications(Milanfar2013a).Differentvariationoperatorswill beusedthroughoutthebookfordifferentapplications.

2Onecanprovethatagraph G withpositiveedgeweightshasPSDgraphLaplacian L viathe Gershgorincircletheorem:eachGershgorindisccorrespondingtoarowin L islocatedinthe non-negativehalf-space,andsincealleigenvaluesresideinsidetheunionofalldiscs,theyare non-negative.

I.5.Graphsignalsmoothnesspriors

Traditionally,forgraph G withpositiveedgeweights,signal x isconsidered smooth ifeachsample xi onnode i issimilartosamples xj onneighboringnodes j withlarge wi,j .Inthegraphfrequencydomain,itmeansthat x mostlycontainslow graphfrequencycomponents,i.e.coefficients α = U x arezeros(ormostlyzeros) forhighfrequencies.Thesmoothestsignalistheconstantvector–thefirst eigenvector u1 for L,correspondingtothesmallesteigenvalue λ1 =0

Mathematically,wecandeclarethatasignal x issmoothifits graphLaplacian regularizer (GLR) x Lx issmall(PangandCheung2017).GLRcanbeexpressed as:

Because L isPSD, x Lx islowerboundedby 0 andachievedwhen x = cu1 for somescalarconstant c.OnecanalsodefineGLRusingthenormalizedgraph Laplacian Ln insteadof L,resultingin x Ln x.Thecaveatsisthattheconstant vector u1 –typicallythemostcommonsignalinimaging–isnolongerthefirst eigenvector,andthus u1 Ln u1 =0

InChen etal. (2015),theadjacencymatrix W isinterpretedasashiftoperator, andthus,graphsignalsmoothnessisinsteaddefinedasthedifferencebetweenasignal x anditsshiftedversion Wx.Specifically, graphtotalvariation (GTV)basedon lp -normis:

where λmax istheeigenvalueof W withthelargestmagnitude(alsocalledthe spectralradius),and p isachoseninteger.Asavarianttoequation[I.4],aquadratic smoothnesspriorisdefinedinRomano etal.(2017),usingarow-stochasticversion Wn = D 1 W oftheadjacencymatrix W :

Toavoidconfusion,wewillcallequation[I.5]the graphshiftvariation (GSV) prior.GSViseasiertouseinpracticethanGTV,sincethecomputationof λmax is requiredforGTV.NotethatGSV,asdefinedinequation[I.5],canalsobeusedfor signalsondirectedgraphs.

I.6.References

Biyikoglu,T.,Leydold,J.,Stadler,P.F.(2005).Nodaldomaintheoremsandbipartitesubgraphs. ElectronicJournalofLinearAlgebra,13,344–351.

Chen,S.,Sandryhaila,A.,Moura,J.,Kovacevic,J.(2015).Signalrecoveryongraphs:Variation minimization. IEEETransactionsonSignalProcessing,63(17),4609–4624.

Cheung,G.,Su,W.-T.,Mao,Y.,Lin,C.-W.(2018).Robustsemisupervisedgraphclassifier learningwithnegativeedgeweights. IEEETransactionsonSignalandInformation ProcessingoverNetworks,4(4),712–726.

Chung,F.(1997).Spectralgraphtheory. CBMSRegionalConferenceSeriesinMathematics, 92.

Dörfler,F.andBullo,F.(2013).Kronreductionofgraphswithapplicationstoelectrical networks. IEEETransactionsonCircuitsandSystemsI:RegularPapers,60(1),150–163.

Lezoray,O.andGrady,L.(2012). ImageProcessingandAnalysiswithGraphs:Theoryand Practice,CRCPress,BocaRaton,Florida.

Liu,X.,Cheung,G.,Wu,X.,Zhao,D.(2017).RandomwalkgraphLaplacianbasedsmoothness priorforsoftdecodingofJPEGimages. IEEETransactionsonImageProcessing,26(2), 509–524.

Milanfar,P.(2013a).Symmetrizingsmoothingfilters. SIAMJournalonImagingSciences, 6(1),263–284.

Milanfar,P.(2013b).Atourofmodernimagefiltering. IEEESignalProcessingMagazine, 30(1),106–128.

Pang,J.andCheung,G.(2017).GraphLaplacianregularizationforimagedenoising:Analysis inthecontinuousdomain. IEEETransactionsonImageProcessing,26(4),1770–1785.

Romano,Y.,Elad,M.,Milanfar,P.(2017).Thelittleenginethatcould:Regularizationby denoising(RED). SIAMJournalonImagingSciences,10(4),1804–1844.

Shuman,D.I.,Narang,S.K.,Frossard,P.,Ortega,A.,Vandergheynst,P.(2013),Theemerging fieldofsignalprocessingongraphs:Extendinghigh-dimensionaldataanalysistonetworks andotherirregulardomains. IEEESignalProcessingMagazine,30(3),83–98.

Su,W.-T.,Cheung,G.,Lin,C.-W.(2017).GraphFouriertransformwithnegativeedgesfor depthimagecoding. IEEEInternationalConferenceonImageProcessing,Beijing.

Tomasi,C.andManduchi,R.(1998),Bilateralfilteringforgrayandcolorimages. IEEE InternationalConferenceonComputerVision,839–846.

Vemulapalli,R.,Tuzel,O.,Liu,M.-Y.(2016).DeepGaussianconditionalrandomfield network:Amodel-baseddeepnetworkfordiscriminativedenoising. Proceedingsofthe IEEEConferenceonComputerVisionandPatternRecognition,4801–4809.

Zhang,K.,Zuo,W.,Chen,Y.,Meng,D.,Zhang,L.(2017).BeyondaGaussiandenoiser: ResiduallearningofdeepCNNforimagedenoising. IEEETransactionsonImage Processing,26(7),3142–3155.

PART 1 FundamentalsofGraph SignalProcessing

1

GraphSpectralFiltering

TokyoUniversityofAgricultureandTechnology,Japan

1.1.Introduction

Thefilteringoftime-andspatial-domainsignalsisoneofthefundamental techniquesforimageprocessingandhasbeenstudiedextensivelytodate.GSPcan treatsignalswithirregularstructuresthataremathematicallyrepresentedasgraphs. Theoriesandmethodologiesforthefilteringofgraphsignalsarestudiedusing spectralgraphtheory.Inimageprocessing,graphsarestrongtoolsforrepresenting structuresformedbypixels,likeedgesandtextures.

Thefilteringofgraphsignalsisnotonlyanextensionofthatforstandardtime-and spatial-domainsignals,butitalsohasitsowninterestingproperties.Forexample,GSP canrepresenttraditionalpixel-dependentimagefilteringmethodsasgraphspectral domainfilters.Furthermore,theoryanddesignmethodsforwaveletsandfilterbanks, whicharestudiedextensivelyinsignalandimageprocessing,arealsoupdatedtotreat graphsignals.

Inthischapter,thespectral-domainfilteringofgraphsignalsisintroduced.In section1.2,thefilteringoftime-domainsignalsisbrieflydescribedasastarting point.Thefilteringofgraphsignals,bothinthevertexandspectraldomains,is detailedinsection1.3,inadditiontoitsrelationshipwithclassicalfiltering. Edge-preservingimagesmoothingisrepresentedasagraphfilterinsection1.4. Furthermore,aframeworkoffilteringbymultiplegraphfilters,i.e.graphwavelets andfilterbanks,ispresentedinsection1.5.Eventually,section1.6introducesseveral fastcomputationmethodsofgraphfiltering.Finally,theconcludingremarksofthis chapterarediscussedinsection1.7.

1.2.Review:filteringoftime-domainsignals

Westartbyreviewingthefilteringindiscrete-timelineartime-invariant(LTI) systems,whichhasbeenextensivelystudiedinliterature.Supposethata one-dimensionaldiscrete-timesignal xn isobtainedbysamplingitscontinuous-time counterpart x(t),withafixedsamplingperiod T ,i.e. xn = x(nT ).A two-dimensionalimagesignalcanbesimilarlyobtainedbyperformingsamplingin boththehorizontalandverticaldirections.Inthiscase,thespatialsamplingperiod usuallycorrespondstothespacingbetweenanarrayofphotosensors.

Supposethatanimpulseresponseofafilter hn isgiven apriori.Thediscrete-time filteredsignal yn intheLTIsystemiscalculatedfrom xn and hn byconvolutionas follows:

Thisequationisbasedonthe shift ofthesignalorimpulseresponse.InLTI systems,we(implicitly)assumethattheshiftofadiscrete-timesignaliswell defined,i.e. xn k isuniqueandtimeinvariant.Therefore,equation[1.1]is equivalentlyrepresentedas

Inequation[1.2],theimpulseresponse hk isinvariantfor n,i.e.thesamefilteris usedfordifferentvaluesof n.Instead,wecanusedifferentfiltersfordifferentvalues of n toyield yn ,whoseimpulseresponse hk [n] isoftendefinedinasignal-dependent manner,i.e. hk [n] = hk [m] for m = n.Itisformulatedas

1Here,weassumeboth x and y arefinitelengthsignalsandtheirboundariesareextendedor filteredbyaboundaryfiltertoensurethattheequationisvalid.

anditsmatrixformrepresentationis

Famousimageprocessingfiltersinthiscategoryincludethebilateralfilter (TomasiandManduchi1998;Barash2002;DurandandDorsey2002;Fleishman etal. 2003),anisotropicdiffusion(Weickert1998;Desbrun etal. 1999),adaptive directionalwavelets(ChangandGirod2007;Ding etal. 2007;Tanaka etal. 2010) andtheirvariants.

Itiswellknownthatconvolutioninthetimedomainequation[1.1]hasan equivalentexpressioninthefrequency(i.e.Fourier)domainasfollows:

Here, ˆ x(ω ) isthediscrete-timeFouriertransform(DTFT)of xn .Weutilizethe factthatconvolutioninthetimedomainisidenticaltomultiplicationinthefrequency domain.Notethatthefixedfilterhasacorrespondingfixedfrequencyresponse,and thus,wecanintuitivelyunderstandthefiltercharacteristicsfromthefrequency response.Incontrast,thefrequencyresponseofasignal-dependentfilterisnot alwaysclearingeneral.Fortunately,thisdrawbackcanbepartiallysolvedwitha graphspectraldomainperspective,whichisdescribedfurther.

1.3.Filteringofgraphsignals

Inthischapter,weconsider linear graphfilters.Readerscanfindnonlineargraph filters,likeoneusedindeeplearning,inthefollowingchapters,specifically Chapter10.

Letusdenoteagraphfilteras H ∈ RN ×N ,whereitselementsaretypicallyderived from G and x.AsintheLTIsystem,thefilteredsignalisrepresentedas

Therepresentationofitselement yn issimilartothatobservedinequation[1.3], i.e.

where [ ]n,k isthe n,k -elementinthematrix.Similartodiscrete-timesignals,graph signalfilteringmaybedefinedinthevertexandgraphfrequencydomains.Theseare describedinthefollowing.

1.3.1. Vertexdomainfiltering

Vertexdomainfilteringisananalogoffilteringinthetimedomain.However, GSPsystemsarenotshift-invariant:Thismeansnodeindicesdonothaveany physicalmeaning,ingeneral.Therefore,theshiftofgraphsignalsbasedonthe indicesofnodes,similartothatusedfordiscrete-timesignals,wouldbe inappropriate.Moreover,theunderlyinggraphwillexhibitahighlyirregular connectivity,i.e.thedegreeineachnodewillvarysignificantly.Forexample,thestar graphshowninFigure1.1hasonecenternodeand N 1 surroundingnodes.Itis clearthat N 1 edgesareconnectedtothecenternode,i.e.thecenternodehas degree N 1,whereasallofthesurroundingnodeshavedegree 1.Inanimage processingperspective,suchanirregularitycomesfromtheedgeandtextureregions. AnexampleisprovidedontherightsideofFigure1.1.Supposethatweconstructa graphbasedonpixelintensity,i.e.pixelsarenodes,andtheyareconnectedbyedges withhigherweightswhentheirpixelvaluesarecloser.Inthissituation,pixels along edge/texturedirectionswillbeconnectedtoeachotherstronglywithalargedegree, whereasthose across edge/texturedirectionsmayhaveweakeredges,ormayevenbe disconnected.Filteringbasedonsuchagraphwillthereforereflectstructuresinthe vertex(i.e.pixel)domain.

Figure1.1. Left:Stargraphwith N =7.Right:TexturedregionofBarbara

Vertexdomainfilteringcanbedefinedmoreformallyasfollows.Let Nn,p beaset of p-hopneighborhoodnodesofthe nthnode.Clearly |Nn,p | variesaccordingto n

Vertexdomainfilteringmaybetypicallydefinedasalocallinearcombinationofthe neighborhoodsamples yn := k ∈Nn,p [H]n,k xk

Since Nn,p variesaccordingto n, [H]n,k shouldbeappropriatelydeterminedfor all n.Thematrixformofequation[1.9]mayberepresentedas

y =(diag([H]0,0 ,..., [H]N 1,N 1 )+ h(W ))x,

where h(W ) isamatrixcontainingfiltercoefficients h[n,k ](n = k ) asafunctionof theadjacencymatrix W ,inwhich [h(W )]n,k =0 if k ∈Nn,p .

Thevertexdomainfilteringinequations[1.9]and[1.10]requiresthe determinationof n |Nn,p | filtercoefficients,ingeneral;moreover,itsometimes needsincreasedcomputationalcomplexity.Typically, [H]n,k maybeparameterized inthefollowingform:

where hp isarealvalueand Wp ∈ RN ×N isamaskedadjacencymatrixthatonly contains p-hopneighborhoodelementsof W .Itisformulatedas

Thenumberofparametersrequiredinequation[1.12]is P ,whichissignificantly smallerthanthatrequiredinequation[1.10].

Onemayfindasimilaritybetweenthetimedomainfilteringinequation[1.2]and theparameterizedvertexdomainfilteringinequation[1.11].Infact,iftheunderlying graphisacyclegraph,equation[1.11]coincideswithequation[1.2]withaproper definitionof Wp .However,theydonotcoincideingeneralcases:Itiseasily confirmedthatthesumofeachrowofthefiltercoefficientmatrixinequation[1.11] isnotconstantduetotheirregularnatureofthegraph,whereas k hk isaconstant intime-domainfiltering.Therefore,theparametersofequation[1.11]shouldbe determinedcarefully.

1.3.2. Spectraldomainfiltering

Thevertexdomainfilteringintroducedaboveintuitivelyparallelstime-domain filtering.However,ithasamajordrawbackinafrequencyperspective.Asmentioned insection1.2,time-domainfilteringandfrequencydomainfilteringareidenticalup totheDTFT.Unfortunately,ingeneral,suchasimplerelationshipdoesnotholdin GSP.Asaresult,thenaïveimplementationofthevertexdomainfiltering equation[1.10]doesnotalwayshaveadiagonalresponseinthegraphfrequency domain.Inotherwords,thefiltercoefficientmatrix H isnotalwaysdiagonalizable bytheGFTmatrix U,i.e. U HU isnotdiagonalingeneral.Therefore,thegraph frequencyresponseof H isnotalwaysclearwhenfilteringisperformedinthevertex domain.Thisisacleardifferencebetweenthefilteringofdiscrete-timesignalsand thatofthegraphsignals.

Fromtheabovedescription,wecancomeupwithanotherpossibilityforthe filteringofgraphsignals:graphsignalfilteringdefinedinthegraphfrequencydomain. ThisisananalogoffilteringintheFourierdomaininequation[1.5].Thisspectral domaindefinitionofgraphsignalfilteringhasmanydesirablepropertieslistedas follows:

–diagonalgraphfrequencyresponse;

–fastcomputation;

–interpretabilityofpixel-dependentimagefilteringasgraphspectralfiltering. Thesepropertiesaredescribedfurther.

Asshowninequation[1.5],theconvolutionof hn and xn inthetimedomain isamultiplicationof ˆ h(ω ) and ˆ x(ω ) intheFourierdomain.Filteringinthegraph frequencydomainutilizessuchananalogtodefinegeneralizedconvolution(Shuman etal. 2016b):

where ˆ xi = ui , x isthe ithGFTcoefficientof x andtheGFTbasis ui isgivenbythe eigendecompositionofthechosengraphoperatorequation[I.2].Furthermore, ˆ h(λi ) isthegraphfrequencyresponseofthegraphfilter.Thefilteredsignalinthevertex domain, y [n],canbeeasilyobtainedbytransforming ˆ y backto y

where [ui ]n isthe nthelementof ui .Thisisequivalentlywritteninthematrixform as

isaprojectionmatrixinwhich σ (λ) isasetofindicesforrepeatedeigenvalues,i.e.a setofindicessuchthat Luk = λuk .

Forsimplicity,letusassumethatalleigenvaluesaredistinct.UnderagivenGFT basis U,graphfrequencydomainfilteringinequation[1.13]isrealizedbyspecifying N graphfrequencyresponsesin ˆ h(λi ).Sincethisisadiagonalmatrix,asshownin equation[1.14],itsfrequencycharacteristicbecomesconsiderablyclearincontrastto thatobservedinvertexdomainfiltering.Notethatthenaïverealizationof equation[1.13]requiresspecificvaluesof λi ,i.e.graphfrequencyvalues.Therefore, theeigenvaluesofthegraphoperatormustbegivenpriortothefiltering.Instead,in thiscase,wecanparameterizeacontinuousspectralresponse ˆ h(λ) fortherange λ ∈ [λmin ,λmax ].Thisgraph-independentdesignprocedurehasbeenwidely implementedinmanyspectralgraphfilters,sincetheeigenvaluesoftenvary significantlyindifferentgraphs.

FortheclassicalFourierdomainfiltering,itisenoughtoconsiderthefrequency range ω ∈ [ π,π ] (oranarbitrary 2π interval).However,graphfrequencyvaries accordingtoanunderlyinggraphand/orthechosengraphoperator.Forexample, symmetricnormalizedgraphLaplacianshaveeigenvalueswithin [0, 2],whereas combinatorialgraphLaplaciansdonothavesuchagraph-independentmaximum bound.ThesimplemaximumboundofcombinatorialgraphLaplacianis,for example,givenas(AndersonJrandMorley1985)

where du isthedegreeofthevertex u.Severalotherimprovementsonthebound arealsofoundinliterature.AlthoughthegraphLaplaciansmentionedabovehave aboundofthelargesteigenvalue,suchboundsarenotapplicabletotheadjacency matrix.Consideringthis,appropriatecareofthegraphfrequencyrangemustbetaken whiledesigninggraphfilters.

Asmentioned,graphfrequencydomainfilteringisananalogofFourierdomain filtering.However,thisdoesnotmeanwealwaysobtainavertexdomainexpression ofthissimilartoequation[1.9].Hence,weneedtocomputetheGFToftheinput signal,whichraisesacomputationalissuedescribedasfollows.FortheGFT,the eigenvectormatrix U hastobecalculatedfromthegraphoperator.The eigendecompositionrequires O (N 3 ) complexityforadensematrix2.This

2Whilethecomputationcostforeigendecompositionofasparsematrixisgenerallylowerthan O (N 3 ),itstillrequiresahighcomputationalcomplexity,especiallyforlargegraphs.

calculationoftenbecomesincreasinglycomplex,especiallyforbigdataapplications, includingimageprocessing.

Typically,graphspectralimageprocessingvectorizesimagepixels.Letusassume thatwehaveagrayscaleimageofsize W × H pixels.Itsvectorizedversionis x ∈ RWH anditscorrespondinggraphoperatorwouldbe RWH ×WH .Forexample, 4Kultra-high-definitionresolutioncorrespondsto W =3, 840 and H =2, 160, whichleadsto WH> 8 × 106 :thisistoolargetoperformeigendecomposition, evenforarecenthigh-speccomputer.Insection1.6,severalfastcomputation methodsofgraphspectralfilteringwillbediscussedtoalleviatethisproblem.

1.3.3. Relationshipbetweengraphspectralfilteringandclassical filtering

Filteringinthegraphfrequencydomainseemstobeanintuitiveextensionof Fourierdomainfilteringintothegraphsetting.Infact,itcoincideswithtime-domain filteringinaspecialcase,beyondtheintuition.

Supposethattheunderlyinggraphisacyclegraphwithlength N ,anditsgraph Laplacian Lcycle isassumedasfollows:

whereitsblankelementsarezero.Itiswellknownthattheeigenvectormatrixof Lcycle istheDFT(Strang1999),i.e.

inwhich

Inotherwords,whenweconsideracyclegraphandassumeitsassociatedgraph Laplacianis Lcycle ,itsGFTistheDFT.Therefore,graphspectralfilteringin equation[1.13]isidenticaltothetime-domainfiltering.Notethat,while U isthe DFT,theintervalofitseigenvaluesisnotequalto 2πk/N .Specificallly,the k th eigenvalueof Lcycle is λk =2 2cos(2πk/N ).

1.4.Edge-preservingsmoothingofimagesasgraphspectralfilters

Thisbook(especiallythischapter)focusesongraphspectraldomainoperations forimageprocessing.Here,wedescribeinterconnectionsbetweenwell-studiededge preservingfiltersandtheirGSP-basedrepresentations.Aspreviouslymentionedin thissection,pixel-dependentfiltersdonothavefrequencydomainexpressionsina classicalsense.Thisisbecausetheimpulseresponsesvaryfordifferentpixelindex values n.Inthefollowing,weshowthatsuchapixel-dependentfiltercanbeviewed asagraphspectralfilter,i.e.itpresentsadiagonalgraphfrequencyresponse. Roughlyspeaking,GSP-basedimageprocessingconsidersthe pixelstructure andthe filterkernel independently.Therefore,thepixel-dependentprocessingcanbe performedwithafixedfilterkernel,owingtotheunderlyinggraph.

1.4.1. Earlyworks

LetusbeginwiththehistorybeforetheGSPera.Inthemid-1990s,Taubin proposedseminalworksonsmoothingusinggraphspectralanalysisfor3Dmesh processing(Taubin1995;Taubin etal. 1996)3.Hedeterminedtheedgeweightsof polygonmeshesusingtheEuclidean(geometric)distancebetweennodes.Assuming pi ∈ R3 asa3-Dcoordinateofthe ithnode,theedgeweightisthendefinedas

where η isthenormalizingfactorand φ(pi , pj ) isanon-negativefunction,which assignsalargeweightif pi and pj areclose.Thetypicalchoiceof φ(pi , pj ) willbe pi pj 1 .

Thematrix W issymmetric.Ifwechoose φ(pi , pi )=0,itsdiagonalelements wouldbecomezero,andasaresult, W couldbeviewedasanormalizedadjacency matrix.Thecoordinatesarethensmoothedbyagraphlow-passfilter,after computingtheGFTbasis U.Similarapproachestothismethodhavebeenusedin severalcomputergraphics/visiontasks(Zhang etal. 2010;ValletandLévy2008; Desbrun etal. 1999;Fleishman etal. 2003;KimandRossignac2005).

Forimagesmoothing,filteringwithaheatkernelrepresentedinthegraph frequencydomainhasalsobeenproposedbyZhangandHancock(2008).Inthis work,theedgeweightsofthepixelgrapharecomputedaccordingtophotometric distance,i.e.largeweightsareassignedtotheedgeswhoseendshavesimilarpixel

3Theterm“graphsignal”wasfirstintroducedinTaubin etal. (1996),tothebestofour knowledge.

valuesandviceversa.Additionally,thegraphspectralfilterisdefinedasasolution fortheheatequationonthegraph,andisexpressedasfollows:

where t> 0 isanarbitraryparameterthathelpscontrolthespreadingspeedcausedby diffusion.NotethatthismethodstillneedseigendecompositionofthegraphLaplacian ifwedecidetoimplementequation[1.21]naïvely.Instead,(ZhangandHancock2008) representequation[1.21]usingtheTaylorseriesaroundtheoriginasfollows:

Bytruncatingtheaboveequationwithanarbitraryorder K ,wecanapproximate theheatkernelasafinite-orderpolynomial(Hammond etal. 2011;Shuman etal. 2013).InZhangandHancock(2008),theKrylovsubspacemethodisused,along withequation[1.22]toapproximatethegraphfilter.Thepolynomialmethodforgraph spectralsmoothingisdetailedinsection1.6.5.

Figure1.2depictstheapproximationerroroftheheatkernelusingtheTaylor series.Clearly,itsapproximationaccuracygetssignificantlyworsewhen λ isaway from 0.Sincethemaximumeigenvalue λmax highlydependsonthegraphused,itis bettertousedifferentapproximationmethodsliketheChebyshevapproximation, whichisintroducedinsection1.6.

1.4.2. Edge-preservingsmoothing

Edge-preservingimagesmoothingiswidelyusedforvarioustasks,aswellasfor imagerestoration(NagaoandMatsuyama1979;Pomalaza-RaezandMcGillem 1984;Weickert1998;TomasiandManduchi1998;Barash2002;DurandandDorsey 2002;Farbman etal. 2008;Xu etal. 2011;He etal.2013).Imagerestorationaimsto approximateanunknownground-truthimagefromitsdegradedversion(s).In contrast,edge-preservingsmoothingistypicallyusedtoyieldauser-desiredimage fromtheoriginalone.Theresultingimageisnotnecessarilyclosetotheoriginalone.

Inthegraphsetting,weneedtodefinepixel-wiseorpatch-wiserelationshipsasa distancebetweenpixelsorpatches,anditisusedtoconstructagraph.Thefollowing distancesareoftenconsidered(Milanfar2013b),where i and j arepixelorpatch indicesand φ(·) issomenonnegativefunction:

1)Geometricdistance: dg (i,j )= φ( pi pj ),where pi isthe ithpixel coordinate.

2)Photometricdistance: dp (i,j )= φ( [x]i [x]j ),where [x]i isthepixelvalue (oftenthreedimensional)ofthe ithpixel/patch.

3)Saliencydistance: ds (i,j )= φ( si sj ),where si isthe ithsaliencyvalue.

4)Combinationsoftheabove.

Taylor series (5th order)

Taylor series (10th order)

Chebyshev polynomial (5th order)

Chebyshev polynomial (10th order)

Figure1.2. Comparisonofapproximationerrorsin ˆ h(λ)= e λ .Foracolorversionof thisfigure,seewww.iste.co.uk/cheung/graph.zip

Saliencyoftheimage/region/pixelisdesignedtosimulateperceptualbehavior (Itti etal. 1998;Harel etal. 2006).Apopularchoiceof φ(·) istheGaussianweight

where σ controlsthespreadofthefilterkernel.

Supposethatthefiltercoefficientsaredeterminedbasedontheabovefeatures,and thattheyaresymmetric,i.e.theoutputpixelvalue yi isrepresentedas

Here, dk ( , ) isoneofthedistancemetricsmentionedearlierand K isthe numberoffeaturesweconsidered.Thescalingfactor Di normalizesthefilterweights as Di = j Wi,j .Forexample,thebilateralfilteris K =2 for dg ( , ) and dp ( , ).

TheFourierdomainrepresentationofsuchpixel-dependentfilterscannotbe calculatedinaclassicalsensebecauseitisnolongershift-invariant:thefiltermatrix W cannotbediagonalizedbytheDFTmatrix.Incontrast,GSPprovidesa frequency-likenotioninthegraphfrequencydomain.Ingeneral,theweightmatrix W inequation[1.24]canberegardedasanadjacencymatrixbecauseall dk ( , ) are assumedtobedistancesbetweenpixels.Supposethatthereisnoself-loopin W ,for simplicity.Ingeneral,thesmoothedimageinequation[1.24]isrepresentedinthe followingmatrixform:

= D 1 Wx

where D = diag(D0 ,...,DN 1 ).Thiscanberewrittenbyusingtherelationship L = D W as(Gadde etal.2013):

where x := D1/2 x isadegree-normalizedsignal.Letusdenotethe eigendecompositionof Ln as Ln := UΛU .Theabovefilteringinequation[1.29] isfurtherrewrittenas:

= U(I Λ)U x

= U ˆ h(Λ)U x [1.31] = h(Ln )x, [1.32]

where y := D1/2 y andthegraphspectralfilterisdefinedas ˆ h(λ):=1 λ.Moreover, λ ∈ [0, 2] forthesymmetricnormalizedgraphLaplacian;therefore,itactsasalinear decaylow-passfilterinthegraphfrequencydomain.

Thisgraphspectralrepresentationofapixel-dependentfiltersuggeststhat the pixel-dependentfilter W implicitlyandsimultaneouslydesignstheunderlyinggraph (andtherefore,theGFTbasis)andthespectralresponseofthegraphfilter.Inother words,theGSPexpressionofthepixel-dependentfilterisfreetodesignthespectral response ˆ h(λ),apartfromthelineardecayone,oncewedetermine W .Forexample, letusconsiderthefollowingspectralresponse:

h(λ)= 1 1+ η ˆ hHPF (λ) ,

where ˆ hHPF (λ) isanarbitrarygraphhigh-passfilterand η> 0 isaparameter.Inthis case, ˆ h(λ) worksasagraphlow-passfilteranditsspectralshapeiscontrolledby

ˆ hHPF (λ).Infact,Gadde etal.(2013)showthatequation[1.33]istheoptimalsolution forthefollowingsignalrestorationproblem:

z = x + n withadditivenoise n and HHPF = U ˆ hHPF (Λ)U .

Imagefilteringsometimesneedsnumerousiterationstosmoothoutthedetails,in caseoftexturedand/ornoisyimages.Therefore,toboostupthesmoothingeffect,the trilateralfiltermethod(ChoudhuryandTumblin2003)firstsmoothsthegradientsof theimage,andsubsequently,thesmoothedgradientisutilizedtosmooththe intensities.Itscounterpartinthegraphspectraldomainisalsoproposedin Onuki etal. (2016)withtheparameteroptimizationmethodfor ρ inequation[1.33], whichminimizesMSEafterdenoisingit.

Figure1.3. Imagedenoisingexampleusingbilateralfilters.Fromlefttoright:Original, noisy(PSNR: 20 02 dB),bilateralfilterinthepixeldomain(PSNR: 26 23 dB),and bilateralfilterinthegraphfrequencydomain(PSNR: 27.14 dB).Bothbilateralfilters usethesameparameters

Figure1.3depictsanexampleofimagedenoisingbythebilateralfilterinthe graphfrequencydomain(Gadde etal.2013).Theimageisdegradedbyadditivewhite Gaussiannoise.Thebilateralfilterinthegraphfrequencydomainusesthespectral filterparameterizedinequation[1.33],with ˆ hHPF = λ and η =5.Itisclearthatthe graphspectralversionefficientlyremovesnoisewhilepreservingimageedges.

1.5.Multiplegraphfilters:graphfilterbanks

Intheprevioussections,weonlyconsideredthecasewhereasinglegraph spectralfilterwasapplied.Severalimageprocessingapplications,suchas

Original Noisy
Bilateral filter (pixel domain)
Bilateral filter (graph freq. domain)

Turn static files into dynamic content formats.

Create a flipbook
Issuu converts static files into: digital portfolios, online yearbooks, online catalogs, digital photo albums and more. Sign up and create your flipbook.
PDF Graph spectral image processing gene cheung download by Education Libraries - Issuu