Orthogonal filters and the implications of wrapping on discrete wavelet transforms

Citation
Orthogonal filters and the implications of wrapping on discrete wavelet transforms

Material Information

Title:
Orthogonal filters and the implications of wrapping on discrete wavelet transforms
Creator:
Bleiler, Sarah K
Place of Publication:
[Tampa, Fla]
Publisher:
University of South Florida
Publication Date:
Language:
English

Subjects

Subjects / Keywords:
Filter
Matrix transform
Fourier series
Smoothness
Zero moments
Dissertations, Academic -- Mathematics -- Masters -- USF ( lcsh )
Genre:
non-fiction ( marcgt )

Notes

Abstract:
ABSTRACT: Discrete wavelet transforms have many applications, including those in image compression and edge detection. Transforms constructed using orthogonal filters are extremely useful in that they can easily be inverted as well as coded. We review the major properties of three well-known orthogonal filters, namely, the Haar, Daubechies, and Coiflet filters. Subsequently, we analyze the Fourier series that corresponds to each of those filters and recall some important results about the smoothness of the modulus of those Fourier series. We consider a specialized case in which the length of the discrete wavelet transform is not much longer than the length of the filter used in its construction. For this case, we prove the existence of additional degrees of freedom in the system of equations used in the construction of the aforementioned orthogonal filters. We suggest a modified Coiflet filter which takes advantage of the extra degrees of freedom by imposing further conditions on the derivative of the Fourier series.
Thesis:
Thesis (M.A.)--University of South Florida, 2008.
Bibliography:
Includes bibliographical references.
System Details:
Mode of access: World Wide Web.
System Details:
System requirements: World Wide Web browser and PDF reader.
General Note:
Title from PDF of title page.
General Note:
Document formatted into pages; contains 52 pages.
Statement of Responsibility:
by Sarah K. Bleiler.

Record Information

Source Institution:
University of South Florida Library
Holding Location:
University of South Florida
Rights Management:
All applicable rights reserved by the source institution and holding location.
Resource Identifier:
002004830 ( ALEPH )
352829605 ( OCLC )
E14-SFE0002676 ( USFLDC DOI )
e14.2676 ( USFLDC Handle )

Postcard Information

Format:
Book

Downloads

This item has the following downloads:


Full Text
xml version 1.0 encoding UTF-8 standalone no
record xmlns http:www.loc.govMARC21slim xmlns:xsi http:www.w3.org2001XMLSchema-instance xsi:schemaLocation http:www.loc.govstandardsmarcxmlschemaMARC21slim.xsd
leader nam Ka
controlfield tag 001 002004830
003 fts
005 20090527084359.0
006 m||||e|||d||||||||
007 cr mnu|||uuuuu
008 090527s2008 flu s 000 0 eng d
datafield ind1 8 ind2 024
subfield code a E14-SFE0002676
035
(OCoLC)352829605
040
FHM
c FHM
049
FHMM
090
QA36 (Online)
1 100
Bleiler, Sarah K.
0 245
Orthogonal filters and the implications of wrapping on discrete wavelet transforms
h [electronic resource] /
by Sarah K. Bleiler.
260
[Tampa, Fla] :
b University of South Florida,
2008.
500
Title from PDF of title page.
Document formatted into pages; contains 52 pages.
502
Thesis (M.A.)--University of South Florida, 2008.
504
Includes bibliographical references.
516
Text (Electronic thesis) in PDF format.
3 520
ABSTRACT: Discrete wavelet transforms have many applications, including those in image compression and edge detection. Transforms constructed using orthogonal filters are extremely useful in that they can easily be inverted as well as coded. We review the major properties of three well-known orthogonal filters, namely, the Haar, Daubechies, and Coiflet filters. Subsequently, we analyze the Fourier series that corresponds to each of those filters and recall some important results about the smoothness of the modulus of those Fourier series. We consider a specialized case in which the length of the discrete wavelet transform is not much longer than the length of the filter used in its construction. For this case, we prove the existence of additional degrees of freedom in the system of equations used in the construction of the aforementioned orthogonal filters. We suggest a modified Coiflet filter which takes advantage of the extra degrees of freedom by imposing further conditions on the derivative of the Fourier series.
538
Mode of access: World Wide Web.
System requirements: World Wide Web browser and PDF reader.
590
Advisor: Catherine Beneteau, Ph.D.
653
Filter
Matrix transform
Fourier series
Smoothness
Zero moments
690
Dissertations, Academic
z USF
x Mathematics
Masters.
773
t USF Electronic Theses and Dissertations.
4 856
u http://digital.lib.usf.edu/?e14.2676



PAGE 1

by SarahK.Bleiler Adissertationsubmittedinpartialfulllment oftherequirementsforthedegreeof MasterofArts DepartmentofMathematics CollegeofArtsandSciences UniversityofSouthFlorida MajorProfessor:CatherineBeneteau,Ph.D. ThomasBieske,Ph.D. DmitryKhavinson,Ph.D. DateofApproval: November18,2008 Keywords:lter,matrixtransform,Fourierseries,smoothness,zeromoments cCopyright2008,SarahK.Bleiler

PAGE 2

Iwouldliketothankmyadvisorandmentor,Dr.CatherineBeneteau,forallofherhelpandsupportthroughoutmytimeinthemathematicsdepartmentatUniversityofSouthFlorida.Herenthusiasm,guidance,andadvicewereinvaluableinthewritingofthisthesisandinmyacademicjourneythusfar. IwouldalsoliketothankbothDr.ThomasBieskeandDr.DmitryKhavinsonforalltheiradviceandsuggestionsduringtheformulationofthisthesis.Dr.Bieskewasparticularlywonderfulinassistingmewiththetechnicalformattingofthispaper,andDr.Khavinson'sreviewandsuggestionswereinvaluableinkeepingthesubjectmatterfocusedandinformative. Specialthankstomymotherandfather,withoutwhomIwouldneverbeinthegreatplaceIamtoday.Yourloveandsupporthavebeenmygreatestmotivation. Lastly,Iwanttothankmybestfriendinthisworld,Ray.Withoutyou,Isurelywouldneverhavemadeittothispoint.Iloveyou.

PAGE 3

ListofFiguresi Abstractiii 1Introduction1 2OrthogonalFilters14 2.1HaarFilter...................................14 2.2DaubechiesFilters...............................15 2.3CoietFilters..................................19 3Smoothness23 3.1SmoothnessofjH(w)j2.............................23 3.2SmoothnessofjH(w)j.............................25 4Wrapping33 4.1Non-wrappingRows..............................35 4.2WrappingRows.................................36 4.2.1Non-wrappingPortion.........................36 4.2.2WrappingPortion............................37 4.3EectsofWrappingonZeroOrthogonalityConditions...........39 4.4ANewFilterConstruction...........................43 5Conclusion48 5.1FutureResearchandStudy..........................48 5.2MathematicaProgramforConstructionofDaubechiesFilters.......49 i

PAGE 4

1.1jH(w)j(fromExample1.0.12)on[0;],hisalowpasslter.........8 1.2jG(w)j(fromExample1.0.12)on[0;],gisahighpasslter........9 1.3ImagebeforeandaftertheapplicationoftheHaarwavelettransform...11 3.1jH(w)jon[0;]forDaubechieslength4lter................30 3.2jH(w)j2on[0;]forDaubechieslength4lter................30 3.3jH(w)jandjH(w)j2on[0;4]forDaubechieslength4lter.........31 3.4jH(w)jforDaubechiesandCoietlters(length6).............31 3.5jH(w)jforDaubechiesandCoietlters(length12).............32 4.1Lowpassportionofa1414DWT(thetophalfofthematrix)......34 4.2Lowpassportionofa1010DWT(thetophalfofthematrix)withc>L.37 4.3Length12modiedlters...........................45 4.4Length18modiedlters...........................46 4.5ModuluscomparisonofDaubechieslength12,Coietlength12,andmodi-edlength12(K=2,c=3)lters......................47 4.6ModuluscomparisonofDaubechieslength18,Coietlength18,andmodi-edlength18(K=3,c=5)lters......................47ii

PAGE 5

SarahK.Bleiler ABSTRACT Discretewavelettransformshavemanyapplications,includingthoseinimagecompres-sionandedgedetection.Transformsconstructedusingorthogonalltersareextremelyusefulinthattheycaneasilybeinvertedaswellascoded.Wereviewthemajorpropertiesofthreewell-knownorthogonallters,namely,theHaar,Daubechies,andCoietlters.Subsequently,weanalyzetheFourierseriesthatcorrespondstoeachofthoseltersandrecallsomeimportantresultsaboutthesmoothnessofthemodulusofthoseFourierseries.Weconsideraspecializedcaseinwhichthelengthofthediscretewavelettransformisnotmuchlongerthanthelengthofthelterusedinitsconstruction.Forthiscase,weprovetheexistenceofadditionaldegreesoffreedominthesystemofequationsusedintheconstructionoftheaforementionedorthogonallters.WesuggestamodiedCoietlterwhichtakesadvantageoftheextradegreesoffreedombyimposingfurtherconditionsonthederivativeoftheFourierseries. iii

PAGE 6

Discretewavelettransforms(DWTs)arematricesusedtotransformsetsofdiscretedataintodatathatiseasiertoworkwithandmanipulate.Thistransformation,orpro-cessing,ofdataisdonethroughmatrixmultiplication.OneexampleofsuchprocessingistheapplicationofaDWTtoanimage.Forinstance,wecanuseDWTstoaidintheprocessofimagecompression.Todothis,wewouldneedtochooseanappropriateDWT,onewhichisknowninpracticetobeusefulinimagecompression,andthenapplyittothematrixcontainingthepixelintensitiesoftheimagewewishtocompress.Theresulting(transformed)imageshouldthenbebettersuitedtotheprocessofimagecompression.AnothersuchimageprocessingdomainforwhichDWTshaveprovenusefulisedgedetec-tion.Wewillexploretheseimageprocessingapplicationsinmoredetailattheendofthissection(seeFigure1.3). Wetakesometimetodiscussthehistoryofwaveletsandwhytheycameabout.Fourieranalysishasbeenusedforcenturiesasatoolforanalyzingandtransformingsignalsandimages.Fourierseriesarebuiltusingsineandcosinefunctions,whichareperiodicwavesthatcontinueinbothdirectionsforever.Therefore,whenprocessingdatawhichisnotitselfperiodicand/ortime-independent,butratheristransientandcontainsabruptchanges,Fouriertransformsarenotveryecient.Waveletswereconstructedasthetoolneededtosuccessfullyworkwithsuchtransientdata.In1946,GaborexaminedFourierserieswhichweretakenonaniteinterval,suchthatbothtimeandfrequencycouldbeconsidered[11].1

PAGE 7

Wenowoutlinethestructureofthisthesis.FirstnotethatthebuildingblocksofDWTsarecalledlters(essentiallytherowsoftheDWT),andtheywillbeourmainmathematicalobjectsofdiscussion.Wewillbegin,inChapters1and2,byintroducingsometerminologyandthebasicsbehindtheconstructionofsomeofthemostwell-knownlters,namelytheHaar,Daubechies,andCoietlters.Thenwewillseehowtheycanbeusedinapplicationssuchasimagecompressionandedgedetection.WewillclassifytheconstructionoftheseltersintermsoftheircorrespondingFourierseriesanddiscusswhythisisuseful.InChapter3,wewillconsiderthe\smoothness"ofthegraphofthemodulusoftheFourierseriesforsomegenerallters.Thiswillhelpustobetterunderstandtheimplicationsofeachlterconstructionaswellastoapproximatehowclosealteristobeing\ideal."WenoticethatforanyDWTcomposedofltersoflengthgreaterthantwo,therewillalwaysbewhatiscalled\wrapping"throughoutthematrix.OurmainquestionforthisthesisistondoutwhatimplicationstheprocessofwrappinghasontheconstructionconditionsthatwehighlightinChapters1and2.Therefore,inChapter4ourfocuswillbeonexploringwrapping,identifyinginterestingoutcomesforDWTscontainingsignicantamountsofwrapping,andconsideringanewlterconstructionthatresultsfromthoseoutcomes.Thisnewlterproducesacorrespondinggraphwhichis\smoother",thusimplyingthatthelteriscloserto\ideal." Webeginwiththedenitionofalter.Denition1.0.1

PAGE 8

2;1 2;0;0;:::),wherehk=1 2fork=0;1,andhk=0otherwise.Ifxisanysequence,thenthecomponentsoftheconvolutionproductcanbewrittenasyn=1 2xn+1 2xn1.Therefore,the\processing"ofdatainthiscasecanbethoughtofasasimpleaveragingofconsecutiveentriesoftheinputdatasequencex.Someltersareusefulforidentifyinghomogeneous(i.e.,locallyconstant)portionsoftheinputdata,whileothersareusefulforidentifyinglargedierencesintheinputdata.Therefore,thelteronechoosesdependsgreatlyontheapplication,andthetypeofprocessingthatisdesired. Throughoutthisthesis,wewillonlybeconcernedwithnitelengthlters.Suchaltercanbewrittenash=(hl;hl+1;:::;hL1;hL),wherehl6=0,hL6=0.Thismeansthateachcomponentoftheconvolutionproductwillbeaconvergentseries,representedbyyn=LXk=lhkxnk: Inordertogainabetterunderstandingofhowdierentlterswillprocessdata,itisoftenusefultoconsidertheircorrespondingFourierseries.Denition1.0.4 Itisbenecialtoconsideralterinthe\Fourierdomain"forseveraldierentreasons,suchasthefollowing:

PAGE 9

LetusinvestigatetheimportanceofthemodulusjH(w)jthatisassociatedwithalterh.HowcanjH(w)jbeusedtopredictalter'sproperties?Proposition1.0.5 Proof.H(w)=LXk=lhkeik(w): H(w)=LPk=lhkeikw(sincehkisreal). SoH(w)= ByEuler'sformula,weknoweikw=coskw+isinkw.BecauseH(w)=LPk=lhkeikw=LPk=lhk(coskw+isinkw),itiseasytoseethatH(w)canbewrittenasapolynomialintermsofcoskwandsinkw,andhencejH(w)jis2-periodic.GiventhefactthatjH(w)jisbothevenand2-periodic,wecangatherallimportantinformationofthegraphbyconsideringonlytheportiondenedon[0;]. ByexaminingjH(w)jonthatinterval,wecangetagoodideaofthenatureoftheprocessingthatthecorrespondinglterhwillperform.Forinstance,ifthevalueofjH(w)jislargeatw=0,thenwhenweprocessthedatausinglterh,ifthereareportionsofthedatathatarelargelyhomogeneous,thelterwillpreservethoseportionsofthedata.Asanexample,considerthelterh=(h0;h1;h2)=(1 4;1 2;1 4).ThecorrespondingFourierseriesisH(w)=1 4+1 2eiw+1 4e2iw,andthusjH(0)j=1.Ifweusethisltertoprocessthesignalx=(:::;1;1;1;:::),thenhx=(:::;1;1;1;:::),thuspreservingthehomogeneityofthesignal.4

PAGE 10

4;1 2;1 4)asabove,notingthatjH()j=0.Ifweusethisltertoprocessthesignalx=(:::;1;1;1;1;:::),thenhx=(:::;0;0;0;:::),thusannihilatingtheoscillationofthesignal. Wecharacterizethistypeoflter(i.e.,onewhichpreserveshomogeneitywhilesimulta-neouslyannihilatinglargeamountsofoscillation)asalowpasslter.Example1.0.12willhelptoclarifythisnotionevenfurther.Wegivetheprecisedenitionsforlowpass(andhighpass)ltersbelow(see[20],p.143,147):Denition1.0.6 Forexample,ifhisaltersuchthatjH(w)j=p Thedenitionoftheidealhighpasslterisanalogoustothedenitiongivenabove. Ingeneral,weconsiderasimplicationofthedenitionsoflowpassandhighpassltersgivenabovebyrequiringonlythefollowingconditions:5

PAGE 11

IfhisalowpasslterthenwemusthavejH(0)j=p IfgisahighpasslterthenwemusthavejG(0)j=0andjG()j=p Thevaluep

PAGE 12

ADWTisusedtoprocessdatainmuchthesamewayaswasdescribedforasinglelter.Lookingcloser,ifxisavectorthatrepresents,say,asignal,theninmultiplyingxbyWN,weareessentiallycomputingeveryothercomponentoftheconvolutionproduct.FormoreinformationontherationalebehindtheconstructionofaDWTanditsrelationtoconvolution,see[20]. WeworkthroughanexamplebelowtoshowaspeciccaseofaDWT.Example1.0.12 2;p 2)andg=(p 2;p 2).Wecaneasilycheckthathisalowpasslter(usingthesimplieddenition):H(w)=p 2+p 2eiw=p 2(1+eiw)=p 2(1+cosw+isinw): 2(1+1)=p 2(1+(1))=0.Similarly,gisahighpasslter.ThegraphsofjH(w)jandjG(w)jaregiveninFigures1.1and1.2. Nowsupposewewanttoprocessthevectorv=(1;1;1;1;1;1;1;1)usingaDWTconstructedwithlowpasslterh=(p 2;p 2)andhighpasslterg=(p 2;p 2).Thevectorvisoflength8sotheDWTwewillusetoprocessthedatainvshouldbeamatrixof7

PAGE 13

Figure1.2:jG(w)j(fromExample1.0.12)on[0;],gisahighpasslter8

PAGE 14

2p 200000000p 2p 200000000p 2p 200000000p 2p 2p 2p 200000000p 2p 200000000p 2p 200000000p 2p 21CCCCCCCCCCCCCCCCCA 2p 200000000p 2p 200000000p 2p 200000000p 2p 2p 2p 200000000p 2p 200000000p 2p 200000000p 2p 21CCCCCCCCCCCCCCCCCA0BBBBBBBBBBBBBBBBB@111111111CCCCCCCCCCCCCCCCCA=0BBBBBBBBBBBBBBBBB@p

PAGE 15

Wecouldeasilycheckthatthehighpassportionofthematrixactsinexactlytheoppositemanner,preservingdierencesandannihilatingsimilarities.Remark1.0.13 2;p 2)andhighpassvectorg=(p 2;p 2).WewilllearninthenextchapterthatthenameofthisparticularDWTistheHaarwavelettransform.

PAGE 16

Weseeimmediatelyhowthisprocesscanaidinbothedgedetectionandimagecompres-sion.Asstatedabove,afterweperformthetransform,anapproximationoftheoriginalimageappearsinthetopleftcornerwhiletheotherthreeblocksrepresentthe\dierences"amongpixelintensitiesinthatimage.Ifthereisalargedierenceinvaluebetweentwoconsecutivepixelsintheoriginalimage,thentheresultingvalueinthe\dierence"blockswillalsobelarge.Thepixelintensitiesforatypical8-bitgrayscaleimagerangefrom0to255where0isblackand255iswhite.Thereforetheedgesofanimageareeasilydetectedbylookingforthelargestvalues(i.e.,thewhitestpixels)intheresultingthree\dierence"blocks. Also,notethattheresultingtransformedmatrixhassignicantlylessdetailthantheoriginalimage.Infact,manyofthepixelvalueshavebeenconvertedto0(black)orcloseto0.Thisislikelytooccurwithmostnaturalimagesbecausetheyusuallyhavelargehomogeneousregions,andthereforethedierencesinthoseareaswillbeminimal.Performingwhatiscalledlossycompression,wecanconvertto0allofthosepixelintensitieswhicharecloseto0.Thisway,wesignicantlyreducetheamountofstoragespaceneededinordertostoretheimageorsenditasaleovertheinternet.Wecallthislossycompressionbecausewhenwerestoretheimage(byperforminganinversetransform),someoftheoriginalinformationwillbelostduetotheslightconversionofthosepixelintensities.11

PAGE 17

ThesignicantbenetofdealingwithorthogonalDWTsisthatwecaneasilyinvertthetransformprocesssincetheinverseofWNissimplyitstranspose.Inaddition,insteadofcomputingtheentirematrixmultiplication,wecantakeadvantageofthesparsenessofWNandwriteanalgorithmwhichecientlycomputestheDWT.HavingorthogonalityinourDWTsmakesiteasytowriteanearlyidenticalalgorithmfortheinversetransformprocess(see[20]foranexample). Inthenextchapter,weshowhowtoconstructsomespecicnitelengthlterswhichcometogethertoformanorthogonalDWT. 12

PAGE 18

2;p 2)andg=(p 2;p 2)whichwediscussedinChapter1areactuallythosecoecients.ThesetwoltersformthebuildingblocksfortheHaarwavelettransform.Recallthathisalowpasslterandgisahighpasslter.Therefore,wecanplacetheselters(andtheir2-translates)intheirrespectivepositionsofthelowpassandhighpassportionsofaDWT.The(lengthN)HaarwavelettransformWNiswrittenasfollows:0BBBBBBBBBBBBBBBBB@p 2p 2000000N 2p 20000rows0000......00000000p 2p 2p 2p 2000000N 2p 20000rows0000......00000000p 2p 21CCCCCCCCCCCCCCCCCA13

PAGE 19

NextwecheckfororthogonalityoftheHaarwavelettransform.Remember,WNisorthogonalifandonlyifWNWTN=IN.Inotherwords,thedotproductofrowiwithrowjmustbe1ifi=jand0ifi6=j.Itisenoughtoconsiderthedotproductofonlyrow1withallotherrowsofWN.Giventhestructureofh,g,andWNitiseasytoseethatallotherdotproductswillgivethesamesetofresults. 2p 2+p 2p 2=1 2+1 2=1. 2p 2+p 2(p 2)=0. Therefore,wehaveshownthattheHaarwavelettransformisorthogonal.2.2DaubechiesFilters Supposewewanttoconstructltershandgofarbitraryevenlengthsuchthathislowpass,gishighpass,andtheresultingDWTisorthogonal.Inthissectionweassume14

PAGE 20

2(correspondingly,LPk=0g2k=1andLPk=2mgkgk2m=0form=1;2;:::;L1 2),thenthesetoflowpass(highpass)rowsformanorthonormalset.Remark2.2.1 2.Note2.2.3 jH(0)j=p jH()j=0(lowpass).15

PAGE 21

jH(w)j2+jH(w+)j2=2(orthonormality). IfwesolvethissystemforalterwithL=1(i.e.,h=(h0;h1)),thenoneofoursolutionswillbetheHaarlowpasslter.IfwesolvethesystemaboveforanygreateroddvalueofL,thenweobtainaninnitenumberofsolutions[8].Therefore,weneedtoaddmoreconditionstooursystemofequations.Theequationswewilladdcanbeconsidered\smoothness"conditions.TheycomefromtakingderivativesoftheFourierseriesatw=andhenceatten(orsmooth)thegraphofjH(w)jatthatpoint.IfL=3,weonlyneedtoaddoneofthesederivativeconditionstothesystem,thatis,werequireH0()=0.ForL=5,weneedtwoderivativeconditions,namelyH0()=0andH00()=0.Ingeneral,forlterh=(h0;h1;:::;hL)oflengthL+1,weneedthederivativeconditionsH(m)()=0form=1;2;:::;L1 2,[8].Infact,Daubechiesshowsthatthisisthemaximalnumberofderivativesthatcanbetaken.Remark2.2.4WewillseeinChapter3howthesederivativeconditionsatplayanimportantrolenotonlyinthesmoothnessofjH(w)jatw=,butalsoatw=0. 2derivativeconditionsdescribedabove.Therefore,thesystemofequationsweneedtosolveinordertogetanitenumberofsolutionsforpossibleDaubechiesltersisthefollowing: jH(0)j=p jH()j=0(lowpass). jH(w)j2+jH(w+)j2=2(orthonormality). 2(derivativeconditions). Onceweobtainanitesetofsolutionstothissystem,wehavetonarrowitdowntooneparticularsolutionthatwewillcalltheDaubechieslter.Daubechies[8]explainsherrationaleforpickingtheDaubechieslterfromthisnitesetofsolutions.Fortheinterestedreader,wehaveincludedaMathematica[22]programwhichoutlinestheprocessfornding16

PAGE 22

SofarinthissectionwehavedescribedhowtoconstructtheDaubechieslowpasslter.Onceweconstructthislter,theonlyremainingproblemistoconstructthecorrespondinghighpasslterg=(g0;g1;:::;gL)thatcompletestheorthogonalDWT.Itiseasytocheckthatifweletgk=(1)khLk,thengishighpass(i.e.,satisesG(0)=0;jG()j=p 2.TheseconditionsgiveorthonormalitybetweenrowsinthehighpassandthelowpassportionsofaDWT.Thefollowingpropositiongivesanequivalentconditionfororthonormality(betweenlowpassandhighpassportions)intermsoftheFourierseriesH(w)andG(w)(see[20],p.292):Proposition2.2.5 2.Note2.2.6

PAGE 24

Proof.ByProposition2.2.2,weknowthatjH(w)j2+jH(w+)j2=2impliesthatthelowpassrows(ofanyDWTconstructedusingh)formanorthonormalset.SupposethatlandLhavethesameparity.IflandLarebothodd,thenhhasLl+1(i.e.,anoddnumberof)entries.ThesameistrueiflandLarebotheven. FortheorthonormalityofthelowpassrowstoholdforallN(i.e.,foralltransformmatricesWN),everyrowtranslatedbytwoentriesanddottedwithrow1mustgivezero.However,sincethereareanoddnumberofentriesinh,whenwetranslateforthe(Ll WewillseelaterthatCoietlterscanbedenedbytheFourierseriesH(w)=p HerewepresentaformulafortheresultingstartingandstoppingindicesoftheltercorrespondingtosuchaFourierseries(see[9]or[20],p.307):Proposition2.3.2 2)19

PAGE 25

See[3],Section6.9,foradiscussionofsomeofthebenetsandapplicationsofusingalterthathasabalanceofderivativeconditionsbetweenw=0andw=.Remark2.3.3 So,inreview,thecharacteristicswhichdeneaCoietlterareasfollows:1. Aswehaveseen,theFourierseriescorrespondingtotheCoietlterisalongandcompli-catedone.WecansimplifytheprocessofsolvingforCoietltersbyusingthestarting20

PAGE 27

WeareinterestedinthesmoothnessofjH(w)jbecauseitgivesusagoodideaofhowcloseourlterisincomparisontotheideallter(seeChapter1).Thatis,thesmootherthegraphofjH(w)jnearw=andw=0,thecloseritsrepresentativelteristobeingideal.Thisisimportantforimageprocessingapplicationssuchasedgedetection.Forinstance,ahighpasslterwhichisclosetoidealwilldoagoodjobofpreservingoscillatory(i.e.,dierent)andannihilatinghomogeneous(i.e.,similar)portionsofthedata.Therefore,theedges(largestdierences)willbeclearlyvisibleinthetransformedimage. In[9],Daubechiesconsidersthesmoothnessresultsintroducedbelow.Wenotethattheresultsaswrittenhereareourownproofs.Note3.0.5

PAGE 28

BytheLeibnizproductrule,(d2m+1 Fork=0;:::;m,ourassumptionconditionsH(j)()=0forj=0;1;:::;mimplythatthecorrespondingtermsinthesummation(3.1.1)arezero.Wenowonlyneedtoconsiderthetermsofthesummationwhichcorrespondtok=m+1;:::;2m+1,namely,2m+1m+1H(m+1)(w) Therefore,(d2m+1 RecallProposition2.2.2whichstatesthatifwehaveareal-valuedlterh=(hl;hl+1;:::;hL1;hL)whoseFourierseriessatisesjH(w)j2+jH(w+)j2=2,thentheorthonormalityconditionsLXk=0h2k=1andLXk=2mhkhk2m=0form=1;2;:::;Ll1 2aresatised.Wewillneedtoassumethatorthonormalityforthenextfewresults.Lemma3.1.2

PAGE 29

Noticethat(d2m+1 WeknowfromLemma3.1.1thatthesecondtermin(3.1.2)mustbe0.Therefore,wecanconcludethat(d2m+1 Proof.(Inductiononm) dwjH(w)j)jw=0=0. WeknowfromLemma3.1.2that(d dwjH(w)j2)jw=0=0. Wecanwritethisas(2jH(w)jd dwjH(w)j)jw=0=0andsincejH(0)j=p dwjH(w)j)jw=0=0.

PAGE 30

Wewillonlyshow(d2m+1 Considerd2m+1 dwjH(w)j2=d2m dwjH(w)j=2d2m dwjH(w)j: dwjH(w)j(2mk)=22mXk=02mk(jH(w)j)(k)(jH(w)j)(2mk+1): Recall,byLemma3.1.2,weknowthatd2m+1 25

PAGE 31

dwjH(w)j)=0.(Itistrivialthatlimw!jH(w)j=0.) Considerd dwjH(w)j2.Wecanwritethistwodierentways,asfollows: dwjH(w)j2=2jH(w)jd dwjH(w)j. dwjH(w)j2=H(w) Therefore,limw!(2jH(w)jd dwjH(w)j)=limw!(H(w) Becausewearetakingalimitasw!,wecandividethroughby2jH(w)jtogetlimw!d dwjH(w)j=limw!H(w) 2jH(w)j 2jH(w)jH0(w): 2jH(w)jand 2jH(w)jareequalto1 2. Thereforewehavelimw!(d dwjH(w)j)=limw!(1 2 2H0(w))=0+0=0. Considerdm dwjH(w)j)=2dm1 dwjH(w)j). UsingtheLeibnizproductrule,wecanrewritetheseasfollows: dwjH(w)j)(m1k)=2m1Pk=0m1k(jH(w)j)(k)(jH(w)j)(mk).

PAGE 33

2jH(w)j( 2jH(w)j!: 4p 4p 4p 4p TheDaubechiesltersatisesalltheconditionsinLemma3.1.1,Lemma3.1.2,Theorem3.2.1,andTheorem3.2.3.Therefore,inthegures,weseethatthe\smoothness"ofjH(w)j2isthesameatw=0asitisatw=.Evenmoreinteresting,thegraphofjH(w)jisabouttwiceas\smooth"atw=0asitisatw=,eventhoughthederivativeconditionsaretakenatw=. Nowlet'sconsiderwhathappensifwehavethederivativeconditionsH(j)(0)=0forj=1;2;:::;m,alongwithorthonormality,asisthecaseforhalfthederivativeconditionsimposedonCoietlters.WhatimplicationsdotheseconditionshaveonthesmoothnessofjH(w)jandjH(w)j2?Itiseasytoshow(similartotheargumentsinLemma3.1.2andTheorem3.2.1)thatwiththeseconditions,weonlyget(dj

PAGE 34

Figure3.2:jH(w)j2on[0;]forDaubechieslength4lter29

PAGE 35

WhencomparingDaubechiesltersandCoietltersofthesamelength,wewouldexpectthegraphofthemodulusfortheDaubechiesltertobesmootheratbothw=0andw=thanthegraphofthemodulusoftheCoietlterwouldbenearthosepoints.Figure3.4andFigure3.5comparethemodulusoftheFourierseriescorrespondingtotheDaubechiesandCoietltersoflength6andlength12,respectively.WewillcomebacktothiscomparisonattheendofChapter4whenwesuggestanewlterconstructionthatblendsthecharacteristicsofboththeDaubechiesandCoietlterconstructions.

PAGE 37

Wenowhaveathoroughunderstandingofthelterconstructionsforsomeofthemostwell-knownlters,namelytheDaubechiesandCoietlters.Hence,wethoughtitwouldbeinterestingtoconsiderwhywrapping,whichoccursinthemajorityofDWTs,doesnotseemtoaecttheconstructionconditionsoftheselters.Morespecically,sincetheselterswereconstructedinawaywhichassuredorthogonalityoftheircorrespondingDWTs,wequestionifandhowwrappingaectstheorthogonalityconditionsastheywereoriginallystatedinChapter2. SupposethatweconsiderthelowpassportionofNNtransformmatriceswhichareformedusingltersoflengthL+1.Inthissection,weassumethatNandL+1arebotheven.Forsimplicity,wewilluseltersoftheformh=(h0;h1;:::;hL)(sameasDaubechieslters),butwenotethateverythingthatfollowsinthischaptercaneasilybeextendedtoevenlengthlterswithanyindices(aswithCoietlterswhoseindicesarebothnegativeandpositive).Figure4.1givesanexampleofthelowpassportionofaDWToflengthN=14withlterlengthL+1=9+1=10.Takenotethatinrow1thereare4zerosfollowingtheltercoecientsofh.Ingeneral(i.e.,forlterlengthL+1andmatrixlengthN),wesaythattherearec1zerosplacedintheremainingentriesofrow1sothatN=L+c=(L+1)+(c1).Denition4.0.5 2orfewertimes(i.e.,c1orfewerspacestotheright)fromtheiroriginalpositioninrow1,wherec=NL,andL+1isthelengthofthelter.32

PAGE 38

2ormoretimes(i.e.,c+1ormorespacestotheright)fromtheiroriginalpositioninrow1,wherec=NL,andL+1isthelengthofthelter. Inotherwords,awrappingrowhasatleasttwononzeroentrieswrappedaroundtotherstcolumnsofthetransformmatrix.Note4.0.7 Itisclearfromthedenitionofnon-wrappingrowsthattherearec1 2non-wrappingrowsinthelowpassportionofanytransformmatrix.InthefollowingLemma,weshowthatthenumberofwrappingrowsinthelowpassportionofaDWTdependssolelyonthelengthofthelter.Lemma4.0.8 2. Proof.ThelowpassportionofaDWThasN 2)andanother1(sincewedonotincludeR1).Therefore,there33

PAGE 39

21=L+c 22 2=L1 2wrappingrowsinthelowpassportionofthetransformmatrix. WerecallthezeroorthogonalityconditionswhichcharacterizethesolutionofDaubechiesltersoflengthL+1:LXk=2mhkhk2m=0form=1;2;:::;L1 2:(4.0.1) Wecanthinkoftheseconditionsasdotproductsoftheoriginallterwithallpossible2-translatesofthelter.Note4.0.9 2non-wrappingrows(ofatransformmatrix)withR1arethesameastherstc1 2standardzeroorthogonalityconditions.Inotherwords,wecangeneralizethedotproductofRNWnwithR1asfollows:RNWnR1=LXk=2nhkhk2nforn=1;2;:::;c1 2:(4.1.2)Example4.1.1 2=4 2=2non-wrappingrowsinthis34

PAGE 40

NextwewillgeneralizeaformulaforthedotproductofthewrappingrowswithR1.4.2WrappingRows 2withR1,wemustconsiderboththewrappingportionandthenon-wrappingportionofeachdotproduct.Denition4.2.1 2times(eachtimeshiftingtwospaces)sothateachlterelementhasbeenshiftedtotherightc+1spacesfromitsoriginalposition(asinR1).Therefore,thenon-wrappingportionofthedotproductR1RW1isreallyjustthe(c+1 2)thstandardzeroorthogonalitycondition.Itisclearthatthisextendsforthenon-wrappingportionofeachofthedotproductsR1RW2;R1RW3;:::;R1RWL1 2.Thatis,thenon-wrappingportionofR1RWnforn=1;2;:::;L1 2isgivenbyLXk=2mhkhk2mform=n+c1 2(4.2.3) (i.e.,form=c+1 2;c+3 2;:::;c+L2 2):35

PAGE 41

2;:::;c+L2 2,therewillbenoresultingorthogonalityconditions.Inotherwords,thenon-wrappingportionsofR1RWLc+2 2;:::;R1RWL1 2areallzero. 2.Todothis,wesubstitutem=n+c1 2togetLXk=2mhkhk2m=LXk=2(n+c1 2)hkhk2(n+c1 2)forn=1;2;:::;L1 2:(4.2.4)4.2.2WrappingPortion 2.Next,weconsiderthewrappingportionofthedotproductofR1witheachofthewrappingrowsRW1;RW2;:::;RWL1 2.WeobservethatRW1haspreciselytwolterelements(h1andh0)whicharewrappedaroundtotherstcolumnsoftheDWT,RW2hasfourlterelements(h3;h2;h1,andh0)whicharewrappedaround,andsooninincrementsoftwountilRWL1 2hasL1lterelements(hL2;:::;h0)wrappedaround.Therefore,theorthogonalityconditionthatweobtainfromthewrappingportionofR1RW1ish1hL+h0hL1,fromthewrappingportionofR1RW2ish3hL+h2hL1+h1hL2+h0hL3,andsoonuntilthewrappingportionofR1RWL1 2whichishL2hL+hL3hL1++h1h3+h0h2.36

PAGE 42

2)tothosethatwegetbyshiftingofrowsinthestandardzeroorthogonalityconditions,exceptinreverseorder.Thatis,thewrappingportionofR1RWnisgivenbythefollowing:LXk=2(L+1 2n)hkhk2(L+1 2n)forn=1;2;:::;L1 2:(4.2.5) Now,bycombiningtheresultsin(4.2.4)and(4.2.5),thecompletegeneralizationofR1RWncanbewrittenasfollows:LXk=2(L+1 2n)hkhk2(L+1 2n)+LXk=2(n+c1 2)hkhk2(n+c1 2)forn=1;2;:::;L1 2:(4.2.6) InordertohaveorthogonalityamongalltherowsofthelowpassportionofaDWT,alloftheexpressionsin(4.2.6)mustbesettozero.(Remember,wealsoneedtheconditionR1R1=1tohaveanorthonormalset,butwewillconcernourselvesonlywiththezeroorthogonalityconditionsatthispoint.)Example4.2.4 2=8 2=4wrappingrowsinthismatrix(seeLemma4.0.8),namelyrow4,row5,row6,androw7.ThedotproductsofthesefourrowswithR1areh9h1+h8h0+h3h9+h2h8+h1h7+h0h6;h9h3+h8h2+h7h1+h6h0+h1h9+h0h8;h9h5+h8h4+h7h3+h6h2+h5h1+h4h0;andh9h7+h8h6+h7h5+h6h4+h5h3+h4h2+h3h1+h2h0;respectively.Thesearethesameconditionsasin(4.2.6). Thecompletegeneralizationformulagivenin(4.2.6)showsthatthewrappingportionandthenon-wrappingportionareeachequivalenttooneofthestandardzeroorthogonality37

PAGE 43

2ofthestandardorthogonalityconditionsinordertoobtainorthogonalityinthelowpassportionofaDWT?Thatis,weknowthestandardzeroorthogonalityconditionsaresucientforanorthogonalDWTcontainingwrapping,butaretheynecessary?4.3EectsofWrappingonZeroOrthogonalityConditions 2(m1)holdsform=1;2;:::;c1 2. Proof.RecallfromNote4.2.3thatthenon-wrappingportionsofR1RWLc+2 2;:::;R1RWL1 2areallzero.SoforthelastL1 2Lc+2 2+1=c1 2wrappingrows,weobtainnoorthogonalityconditionsstemmingfromthenon-wrappingportionofthedotproduct.Therefore,weonlyneedtocomparetheorthogonalityconditionsfromthec1 2non-wrappingrowstotheorthogonalityconditionscomingfromthewrappingportionofthelastc1 2wrappingrows.38

PAGE 44

2non-wrappingrowsaregivenbyLXk=2mhkhk2mform=1;2;:::;c1 2:(4.3.7) Theconditionsfromthelastc1 2wrappingrowsaregivenbyLXk=2(L+1 2n)hkhk2(L+1 2n)forn=L1 2(m1)wherem=1;2;:::;c1 2:(4.3.8) Wesubstitutethevalueforninto(4.3.8)togetLXk=2(L+1 2n)hkhk2(L+1 2n)=LXk=2(L+1 2(L1 2(m1)))hkhk2(L+1 2(L1 2(m1)))=LXk=2mhkhk2mform=1;2;:::;c1 2: 2c1 2=Lc Proof.1. 2m)hkhk2(L+1 2m)+LXk=2(m+c1 2)hkhk2(m+c1 2)=LXk=L+12mhkhk(L+12m)+LXk=2m+c1hkhk(2m+c1):39

PAGE 45

2(Lc 2(Lc 2)hkhk2((Lc 2)=LXk=2m+c1hkhk(2m+c1)+LXk=L+12mhkhk(L+12m): 4ezeroorthogonalityconditionswhicharenecessaryinordertosatisfyorthogonalityofthelowpassportionofWN. Proof.Outofthec1rowsconsideredinLemma4.3.1(i.e.,c1 2non-wrappingrowsandc1 2wrappingrows),onlyc1 2uniqueorthogonalityconditionsresultfromtheirdotproductswithR1. Now,outoftheLc Weknowthatfortherstcase,whenLc 2+dLc 2+Lc 4e=dN2 4enecessaryzeroorthogonalityconditions. 40

PAGE 46

2ofthestandardzeroorthogonalityconditionstosatisfyorthogonalityinthelowpassportionofthematrix.Corollary4.3.5 2zeroorthogonalityconditions).Note4.3.6 Proof.RecallthatthereareL1 2standardzeroorthogonalityconditionsandonlydN2 4enecessaryzeroorthogonalityconditions.Toobtainthedegreesoffreedominoursystem,wesubtractthenumberofnecessaryconditionsfromthenumberofstandardconditions,asfollows:L1 2dN2 4e=2L2 4dL+c2 4e:(4.3.9) NotethatL+c2isalwayseven.Therefore,whenwedivideby4wegeteitheranintegeroranintegerandahalf.Thus,(4:3:9)=8<:2L2 4(L+c2 4);ifL+c2 42Z2L2 4(L+c2 4+1 2);ifL+c2 462Z 42ZLc 2;ifL+c2 462Z: 2isodd,thenLc 2fromtheoddnumberL+c2 2togeta(necessarilyodd)resultofLc 4isanintegeroranintegerandahalf). Thus,(4:3:9)=bLc 41

PAGE 47

2.42

PAGE 48

LetC=c1 2andL=Ll+1 2thenLPk=l+2(Ln)hkhk2(Ln)+LPk=l+2(n+C)hkhk2(n+C)=0forn=c+1 2;:::;Ll1 2. ForthestandardCoietlter,thevaluesofandareboth2K1.Therefore,forourmodiedlter,werstincreasethevalueofto2K,then,ifpossible,weincreasethevalueofto2K.Afterthatwewouldcontinueonebyoneincreasingthevalueofandthenincreasingthevalueof.Fromourempiricalobservations,weformthefollowingconjecture:Conjecture4.4.2 InFigure4.5,weprovideagraphicalcomparisonbetweenthemodulusgraphsofthestandardDaubechieslter,thestandardCoietlter,andthemodiedlter(K=2,c=3);allofthesearelength12lters.InFigure4.6,wecomparethemodulusoftheDaubechies,Coiet,andmodied(K=3,c=5)ltersoflength18.Inthesegraphicalrepresentations,itisinterestingtoseethatthegraphofjH(w)jcorrespondingtothemodiedlterconstruction\liesbetween"thegraphsofjH(w)jfortheDaubechiesandCoietltersofthesamelength. 43

PAGE 51

Figure4.6:ModuluscomparisonofDaubechieslength18,Coietlength18,andmodiedlength18(K=3,c=5)lters46

PAGE 52

Inthisthesis,wehavediscussedonlyorthogonallters.Sincetheconstructionoftheseltersrequiresalargenumberoforthogonalityconditions,therearenotmanydegreesoffreedomremaininginoursystemtobettersuittheapplicationweareworkingtowards.Therearehowevermanyothertypesofltersthathavebeenconstructed,includingthosecalledbiorthogonal.BiorthogonalltersareconstructedsothattheircorrespondingDWTisinvertible,butnotorthogonal[6].Theselterscanbeextremelyusefulinimageprocess-ingapplications.Actually,oneofthemostpowerfulandfrequentlyusedimagecompressiontoolsiscalledJPEG2000,anditusesbiorthogonalltersinitswavelettransforms.Forfurtherreadingontheconstructionofbiorthogonallters,see[6],andforfurtherreadingonJPEG2000anditsapplications,see[5].47

PAGE 53

Daub[L_]:=Module[{H,w,k,orthogsquare,orthog,m,i,remainingorthog,lowpasszero,lowpasspi,derivspi,derivpi,Solutions,R,n,P,z,zroot,mytable,mytable2},(*Definesamoduleandallofitslocalvariables*)H[w_]:=Sum[Subscript[h,k]*E^(I*k*w),{k,0,L}];(*TheFourierseriesforaDaubechiesfilteroflengthL+1*)orthogsquare=Sum[Subscript[h,k]^2,{k,0,L}]==1;(*ThefirstorthonormalityconditionfromProposition2.2.1*)orthog[m_]=Sum[Subscript[h,k]Subscript[h,k-2m],{k,2m,L}]==0/.{m->i};remainingorthog=Table[orthog[m],{i,1,(L-1)/2}];(*Thenext(L-1)/2orthonormalityconditionsfromProposition2.2.1*)lowpasszero=H[0]==Sqrt[2];(*ThelowpassconditionH(0)=Sqrt[2]*)lowpasspi=H[Pi]==0;48

PAGE 55

Endofcode

PAGE 56

G.Beylkin,R.Coifman,andV.Rokhlin,Fastwavelettransformsandnumericalalgorithms,CommunicationsonPureandAppliedMathematics44(1991),141-183.[2] A.BoggessandF.J.Narcowich,AFirstCourseinWaveletswithFourierAnalysis,PrenticeHall,UpperSaddleRiver,N.J.,2001.[3] C.S.BurrusandR.A.Gopinath,IntroductiontoWaveletsandWaveletTransforms:APrimer,PrenticeHall,UpperSaddleRiver,N.J.,1998.[4] C.S.BurrusandJ.E.Odegard,Coietsystemsandzeromoments,IEEETransactionsonSignalProcessing46(1998),no.3,761-766.[5] C.Christopoulos,A.Skodras,andT.Ebrahimi,TheJPEG2000stillimagecodingsystem:Anoverview,IEEETransactionsonConsumerElectronics46(2000),no.4,1103-1127.[6] A.Cohen,I.Daubechies,andJ.-C.Feauveau,Biorthogonalbasesofcompactlysup-portedwavelets,CommunicationsonPureandAppliedMathematics45(1992),no.5,485-560.[7] J.M.Combes,A.Grossman,andP.Tchamitchian,Eds.,Wavelets,Time-FrequencyMethodsandPhaseSpace,Springer-Verlag,Berlin,1989.[8] I.Daubechies,Orthonormalbasesofcompactlysupportedwavelets,CommunicationsonPureandAppliedMathematics41(1988),909-996.[9] I.Daubechies,Orthonormalbasesofcompactlysupportedwavelets,SIAMJournalonMathematicalAnalysis24(1993),no.2,499-519.[10] M.W.Frazier,AnIntroductiontoWaveletsThroughLinearAlgebra,Springer-Verlag,NewYork,N.Y.,1999.51

PAGE 57

D.Gabor,Theoryofcommunication,JournaloftheInstitutionofElectricalEngineers,PartIII93(1946),no.26,429-457.[12] A.GrossmanandJ.Morlet,DecompositionofHardyfunctionsintosquareintegrablewaveletsofconstantshape,SIAMJournalonMathematicalAnalysis15(1984),723-736.[13] A.Haar,Zurtheoriederorthogonalenfunktionensysteme,MathematischeAnnalen69(1910),331-371.[14] P.C.J.Hill,DennisGabor-Contributionstocommunicationtheory&signalprocess-ing,EUROCON,2007.TheInternationalConferenceon\ComputerasaTool"(2007),2632-2637.[15] Y.Meyer,Ed.,WaveletsandApplications,Springer-Verlag,Berlin,1992.[16] J.Morlet,Samplingtheoryandwavepropagation,IssuesonAcousticSignal/ImageProcessingandRecognition,C.H.Chen,Ed.,NATOASI,Springer-Verlag,NewYork,N.Y.,1983.[17] O.RioulandM.Vetterli,Waveletandsignalprocessing,IEEESignalProcessingMagazine8(1991),no.4,14-38.[18] M.B.Ruskai,G.Beylkin,R.Coifman,I.Daubechies,S.Mallat,Y.Meyer,andL.Raphael,Eds.,WaveletsandtheirApplications,JonesandBartlett,Boston,M.A.,1992.[19] J.TianandR.O.Wells,Vanishingmomentsandwaveletapproximation,TechnicalReportComputationalMathematicsLab,RiceUniversity,January1995.[20] P.J.VanFleet,DiscreteWaveletTransformations:AnElementaryApproachwithApplications,JohnWiley&SonsInc.,Hoboken,N.J.,2008.[21] D.F.Walnut,AnIntroductiontoWaveletAnalysis,Birkhauser,Cambridge,M.A.,2002.[22] WolframResearch,Inc.,MathematicaEdition:Version6.0,Champaign,I.L.,2007.52


printinsert_linkshareget_appmore_horiz

Download Options

close
Choose Size
Choose file type
Cite this item close

APA

Cras ut cursus ante, a fringilla nunc. Mauris lorem nunc, cursus sit amet enim ac, vehicula vestibulum mi. Mauris viverra nisl vel enim faucibus porta. Praesent sit amet ornare diam, non finibus nulla.

MLA

Cras efficitur magna et sapien varius, luctus ullamcorper dolor convallis. Orci varius natoque penatibus et magnis dis parturient montes, nascetur ridiculus mus. Fusce sit amet justo ut erat laoreet congue sed a ante.

CHICAGO

Phasellus ornare in augue eu imperdiet. Donec malesuada sapien ante, at vehicula orci tempor molestie. Proin vitae urna elit. Pellentesque vitae nisi et diam euismod malesuada aliquet non erat.

WIKIPEDIA

Nunc fringilla dolor ut dictum placerat. Proin ac neque rutrum, consectetur ligula id, laoreet ligula. Nulla lorem massa, consectetur vitae consequat in, lobortis at dolor. Nunc sed leo odio.