小波与信号处理.docx
《小波与信号处理.docx》由会员分享,可在线阅读,更多相关《小波与信号处理.docx(17页珍藏版)》请在冰豆网上搜索。
小波与信号处理
第七章小波与信号处理
WAVELET&SIGNALPROCESSINGWITHAPPLICATIONS
小波是应用数学的一个新领域
80年代法国地质石油工程师J.Morlet提出Wavelet的概念,通过物理直观概念和信号处理的实际需要建立了经验的反演公式。
但当时的数学家并未认可-1807年Fourier热力学工程师-FourierKingdom
1986年数学家Mayer偶然构造出一个真正的小波基函数,并与Mallat合作建立了构造小波基函数的统一方法-多尺度分析方法及Mallat金字塔算法,小波才得到真正的发展。
Daubechies-比利时女数学家-小波十讲-Daubechies小波
近代科学400年来-从复杂事物中寻找简单规律-几乎忘记了非线性用局部线性化来取代非线性,用片面的美掩盖了整体的美。
一叶障目不见泰山。
自然世界的本性,丰富多彩、神奇多样化都来源于非线性-分形、Fractals-Choas。
1963年气象学家洛伦兹用计算机仿真大气模型,仅忽略了一些看起来微不足道的参数,但结果却大相径庭。
-科学别忘记了非线性
小波经过近十年的研究,数学体系已经建立,理论基础扎实,应用广泛。
小波分析,时间频率局部变换,多尺度分析,解决了传统的傅立叶变换难以解决的许多难题-数学显微镜,信号分析,数据压缩、模式识别、特征提取。
AGeneralCommentonInformationRepresentation——APrelude
informationF(u)
Today'scomputingdevicesarelargelylimitedtoprocessf(u)typeofinformation.
Wherefrepresentstheinformation,anduisageneralindexatwhichtheinformationisdescribed.
examplesifindexu:
timet(speech,music,etc.)
positionx(stillimage,etc.)
(time,position)(t,x)(video,etc.)
otherphysicalparameters...
indexu:
continuousversusdiscrete
indexuversusinformationf:
whichoneismoreimportant
Acharacterofusefulinformation:
redundancy.
TheNoteswerepreparedbasedonthefollowingreferences:
1.J.S.walker,"Fourieranalysisandwaveletanalysis."NoticesofAMS,vol.44,NO.6,pp.558-670,June/July1997.
2.C.Mulcahy,"plottingandschemingwithwavelets."Math.Magazine,b9,Dec.1996.
3.G.strangandT.Nguyen,WaveletsandFilterBanks,Wellesley-CambridgePress,1996.
4.W.-S.Lu,ELEC639ACourseNotesonWavelets,DeptofElec.&Comp.Engn.,Univ.ofVictoria,Jan1997.
§7-1.什么是小波WhatAreWavelets?
Awaveletisafunction(t)whoseaverageisZero,i.e.,
Examples:
WaveletsthatareusefulinDSPandotherapplicationsoftenpossessadditionalfeatures:
Generate(viadilationandtranslation)orthonormalorbiorthonormal(basis)systemsinL2
Localbothintimeandfrequencydomain
FastDecomposition/ReconstructionAlgorithms
Vanishingmoments
§7-2.SignalRepresentation:
FourierSystems
7.2.1SystemFeatures
AnOrthonormalSyteminL2(R)generatedwithone(Mother)function:
ejt
OffersFrequencyAnalysisofSignalsandDSPsystem
FastAlgorithms(FFT,DFT,DCT,2D-DFT,……)
InfiniteResolutioninFrequencydomain
ZeroResolutioninTimedomain
不能表示突变信号,局部时间变化的信号
7.2.2Application:
SignalApproximation(LossyCommpression)
Considerdiscretesignal
where
TheDFTof{fk}isgivenby
Anapproximationof{fk}istheinverseDFTof
"CompressionRate"=M/N
ThefigurebelowshowsaplotoftheoriginalsignalF(t),andthreeapproximatesofF(t)withN=512,256,and128.NoticetheGibbsoscillationsforN=256and128.
7.2.3Application:
SignalDenoising
Considera"Modulated"Signal
representingthebitsequence1011011
Supposethesignalistransmittedthroughachannelthataddsacertainamountofnoisewhosefrequencyrangeisboundedby200HZ.Thenoise-contaminatedsignalcanthenberecoveredusingabandpassfilteringtechnique:
STFT短时傅氏变换(Gabor变换)
☆在频域和时域均有局部化功能;
☆但是、其时频窗口的大小是固定的,窗口没有自适应性,只适合分析所有特征尺度大致相同的的各种过程。
☆此外、其离散形式没有正交展开形式,难以实现高效算法。
短时傅氏变换,只是向时频分析走出了重要的一步。
Guass窗
参考书“子波变换与子波分析”赵松年
窗口函数
易见、当两个频率分量之差为2时,STFT可以分辨出来
而当两个频率分量之差为1时,STFT就不能分辨了。
§7-3SignalRepresentation:
TheHaarWavelet
7.3.1HaarScalingFunctionandHaarwavelet
TheHaarscalingfunctionisdefinedby
尺度函数
property:
TheHaarwaveletisdefinedby
property:
小波函数
7.3.2HaarScalingFunctionSeenFromSignalApproximation
Consideracontinuous-timesignalf(t)anditsapprox.
where(t)theHaarScalingfunction
Weseef0(t)asafinite-resolutionrepresentationoff(t)inthespaceV0spannedby{(t-k),kZ}
Note:
V0=thesetofallfunctionsthatarepiecewiseconstanton[k,k+1),kZ
and
Twoproblems:
(a)Whatiff0(t)asanapproximatelyisnotaccurateenough?
(b)Whatiff0(t)istooaccurate?
WecanmovefromV0uptospaceswithhigherresolutioninthecaseof(a).
Orinthecaseof(b)wemovetospaceswithlowerresolution.
Case(a)Anaturalwaytoobtainmoreaccuraterepresentationoff(t)istoincreasethenumberofsamples.Forexample,ifthesamplingrateisdoubled,thenweobtain
Where
HereweseeaspaceV1:
isorthonormal
Case(b)Wecanforinstanceusehalfofthesamplingratetorepresentthesignaliff0(t)is“tooaccurate”:
Where
NotethespaceV-1
Acomparisionof:
§7-3Haarwaveletseenfromsignalreconstruction
LetusconsiderfasignaloffinitelengthNinspaceV1:
Weapproximatef(t)withasignal
;lengthof
Astheaveragingisakindoflowpassfiltering,detailsoffarelostin
.Thelostinformationcanbedescribedas
Note:
lengthof
Thissuggestsasignaldecompositionscheme:
Lowpassfilteringplusdownsamplingby2
AndHighpassfilteringplusdownsamplingby2
Featuresofthedecomposition:
{informationamountoff}={infoamountofa0}+{infoamountofd0}
fcanbeperfectlyrecoveredbyd0anda0
d0isusuallySMALL
Usethisdecompositionasabuildingblock,weobtaina(asymmetric)treestructureforsignalanalysisandCOMMPRESSION:
Foradiscretesignalatlength2k,klevelsatdecompositionmaybeusedtoobtaind0,d-1,...d-k+1,a-k+1,suchthat
Theirtotallength=thelengthoff;
Manyofd0areSMALL
{d0,d-1,...d-k+1,a-k+1}canbeusedtoREconstructf.
SignalCompression/FunctionApproximation
★Mathematiciansseetheuseofthedecomposition/reconstructioninfunctionapproximationwithfewerterms:
★Engineersseetheuseofthedecomposition/reconstructioninsigalcoding(compression)
Theoriginalsignalis
andoneleveldecomposition
usingtheHaarwavelet;5-leveldecompositionusingtheHarrwavelet.
Noticethesmallernumberofwaveletcoefficientsinthe5-leveldecomposition.
Wenoelookattheabovedecompositionanalytically:
where
Thefigurebelowshowswhathappensintheinterval[n,n+2):
7.3.4Connectionofscalingfunctionsatdifferentresolutionlelvels
Basicequationfor
:
Use(DE)onecaneasilyconnect
to
7.3.5ConnectionatWavelettoScalingFunction
Bydefinitionwecanverify
ThisistheWAVELETEQUATIONfortheHaarwavelet.
Together,theDEandWEareofcriticalimportanceforthedevelopmentofapowerfulMULTIRESOLUTIONANALYSIS(MRA)ofsignals,andforgeneratingmanyimprovedwaveletsthatareofmoreuseinDSPandotherapplications.
73.6DEandWE:
AFilter-BankInterpretation
Thenormalizedscalingfunctions(inV0andV1)are
andthenormalizedwaveletinW0is
TheDE&WEintermsof
are
Thecoefficients{1/2,1/2}inDEinterpretedastheimpulseresponseoftelowpass(average)filter
Thecoefficients{1/2,1/2}inWEcanbeviewedastheimpulsereponseofthehighpass(difference)filter
ThefigurebelowshowstheamplituderesponsesofF0(z)(lowpass)andF1(z)(highpass).
NoticethatF0(ej)=0at=(0.5innormalizedfrequentcy).Thisisimportantaswecanshowthat
#ofzerosofF0(z)at=isequalto
#ofvanishingmomentsoftheassociatedwavelet.
相当于低通滤波器的零点。
Haar小波仅此一个零点,故不是很有用;
零点越多越有用,如信号压缩的倍数越高。
Daubechies在她的论文中给出了,在给定零点条件下,阶数最少的滤波器。
Haar小波的实际应用意义并不大;
但是,这里作为例子,用来帮助理解小波的原理是很有益的。