the scientist waveletslhf.impa.br/abel2017/velho-meyer.pdf · 2017. 8. 14. · orthogonal bases....
TRANSCRIPT
![Page 1: The Scientist Waveletslhf.impa.br/abel2017/velho-meyer.pdf · 2017. 8. 14. · orthogonal bases. Coifman, Daubechies and Stéphane Mallat went on to develop applications to many problems](https://reader033.vdocuments.net/reader033/viewer/2022052817/60aff4ed90577624d843170a/html5/thumbnails/1.jpg)
Yves Meyer
WaveletsLuiz Velho
IMPA
ContributionsYves Meyer was awarded the 2010 Gauss Prize for fundamental contributions to number theory, operator theory and harmonic analysis, and his pivotal role in the development of wavelets and
multiresolution analysis.
He also received the 2017 Abel Prize “for his pivotal role in the development of the
mathematical theory of wavelets.”
Stéphane Mallat calls him a “visionary” whose work cannot be labelled either
pure or applied mathematics, nor computer science either, but simply
“amazing”.
The ScientistFrom the mid-1980s, in what he called a
“second scientific life”, Meyer, together with Daubechies and Coifman, brought together
earlier work on wavelets into a unified picture.
In 1986 Meyer and Pierre Gilles Lemarié-Rieusset showed that wavelets may form mutually
independent sets of mathematical objects called orthogonal bases. Coifman, Daubechies and
Stéphane Mallat went on to develop applications to many problems in signal and image processing.
Wavelets
The Meyer Wavelet• First Non-Trivial Orthogonal Wavelet Basis
• Continuously Differentiable
• Non-Compact Support
• Defined in Frequency Domain
• Continuous Wavelet Transform
(1985)
Definition
![Page 2: The Scientist Waveletslhf.impa.br/abel2017/velho-meyer.pdf · 2017. 8. 14. · orthogonal bases. Coifman, Daubechies and Stéphane Mallat went on to develop applications to many problems](https://reader033.vdocuments.net/reader033/viewer/2022052817/60aff4ed90577624d843170a/html5/thumbnails/2.jpg)
The Function
• Space Domain
The Next Frontier
• “Big Questions”:
- How to Discretize the Wavelet Transform?
- How to Systematically Construct Wavelet Basis?
• The Answer
- Multiresolution Analysis
Multiresolution Analysis
• Yves Meyer and Stephane Mallat
• Foundations of Modern Signal Processing
• Many Applications in Science and Engineering
(1988)
Multiresolution Analysis and
Filter Banks
Outline• Scale Spaces
• Multiresolution Analysis
• Dilation Equations
• Fast Wavelet Transform • Two-Channel Filter Banks
• Matrix Implementation
• Examples
Scale
Natural Concept
• Physical
Measurements, Data Acquisition
• Perception
Focus, Features
• Applications
Units, Computation
![Page 3: The Scientist Waveletslhf.impa.br/abel2017/velho-meyer.pdf · 2017. 8. 14. · orthogonal bases. Coifman, Daubechies and Stéphane Mallat went on to develop applications to many problems](https://reader033.vdocuments.net/reader033/viewer/2022052817/60aff4ed90577624d843170a/html5/thumbnails/3.jpg)
Scale Spaces
Vs : Space of Scale s • Scaling Function
-1 1 2
)(||
1)(sx
sxs φφ =
1)( =∫ dxxsφ
{ })(| , kxsks −φφ
• Basis of Vs
Representation of a Function in VS
• Orthogonal Projection
• Representation Operator
• Reconstruction Operator
)( ,)(Proj ,, xff ksk
ksVs φφ∑=
φφ ∗== ffc kssk ,,
)()( , xcxfk
kssks ∑= φ
(continuous function)
(discrete sequence)
Resolution and Multiresolution
Scale ↔ Resolution
• Data Representation
Computation - Discrete Elements (Samples)
Resolution - Samples / Unit of Scale
• Need for Representation at Multiple Scales
Different Features - Different Scales
Efficient Computation - Coarse to Fine Scale
Sequence of Scale Spaces
1. Inclusion:
2. Scaling:
3. Density:
4. Maximality:
5. Scale Basis:
( Vj ) , j ∈ Z
Multiresolution Analysis
Vj ⊂ Vj-1
f(x) ∈ Vj ⇔ f(2 jx) ∈ V0
closure{∪j Vj } = L2(R)
∩j Vj = { 0 }
∃ φ (x) s.t. {φ (x-k) } basis of V0
Two Scale RelationAxioms 1. and 2. (Nested Spaces and Dyadic Scaling)
f(x) ∈ Vj ⇔ f(2x) ∈ Vj-1
f ∈ Vj ⇔ f ∈ Vj-1
Completeness
Axioms 3. and 4. (complete covering of function space)
closure{∪j Vj } = L2(R)
∩j Vj = { 0 }
![Page 4: The Scientist Waveletslhf.impa.br/abel2017/velho-meyer.pdf · 2017. 8. 14. · orthogonal bases. Coifman, Daubechies and Stéphane Mallat went on to develop applications to many problems](https://reader033.vdocuments.net/reader033/viewer/2022052817/60aff4ed90577624d843170a/html5/thumbnails/4.jpg)
Multiresolution Basis
φ j,k (x) = 2 - j/2 φ (2 - j x - k)
Vj+1
Vj
Vj-1
Dilations
Translations
Axioms 5. and 2. (basis family generated from a single φ)
From Multiresolution to Wavelets
Multiresolution Ladder • Coarse to Fine
• Fine to Coarse (information loss)
Wavelets: Complementary Subspaces Intuition: difference between two consecutive levels Wj : Details in Vj-1 that cannot be represented in Vj
f ∈ Vj ⇒ f ∈ Vj-1
f ∈ Vj-1 f ∉ Vj
Wavelet Spaces
Definition: Orthogonal Complement of Vj in V j-1
therefore: jjj WVV ⊕=−1
)(Proj)(Proj)(Proj wjVj1Vj- fff +=
Vj
f
( ) f
( ) f ProjVj
ProjWjWj
Vj-1
Wavelet BasisOrthogonality of Spaces • Decomposition of L2
• Basis of Wj
ψj,k (x) = 2 - j/2 ψ (2 - j x - k)
kj WW ⊥
jjWLZ
2 (R)∈⊕=
kj ≠
Multiresolution Decomposition• Scale and Wavelet Spaces • Function Decomposition
{0} … ⊂ V1 ⊂ V0 ⊂ V-1 ⊂ ... L2(R) | | | | | …
W1 W0 W-1 … …
f ∈ V0
f = ProjW0 ( f ) + ...+ ProjWn ( f ) + ProjVn ( f )
Two-Scale Operators
• Transform Representation Sequences
Discrete Operators
Move Between Consecutive Levels j and j+1
• Act on Multiresolution Hierarchy
Fine ↔ Coarse
Fine ↔ Detail
Key to Efficient Computation
![Page 5: The Scientist Waveletslhf.impa.br/abel2017/velho-meyer.pdf · 2017. 8. 14. · orthogonal bases. Coifman, Daubechies and Stéphane Mallat went on to develop applications to many problems](https://reader033.vdocuments.net/reader033/viewer/2022052817/60aff4ed90577624d843170a/html5/thumbnails/5.jpg)
Dilation Equations• Scaling Basis:
100 −⊂∈ VVφ
)(,1)(,1 ,100 , xkk
kxkk
k h−−− ∑=∑= φφφφφ
100 −⊂∈ VWψ
)(,1)(,1 ,100 , xkk
kxkk
k g−−− ∑=∑= φφφψψ
• Wavelet Basis:
Expression of φ and ψ
• Reflexive Definition
• ψ defined in terms of φ
)2(2)( kxk
kx h −= ∑ φφ
)2(2)( kxk
kx g −= ∑ φψ
Shape of φ and ψ
)(xφ
11 −−φh
)(xψ
00φh 11φh 11 −−φg 00φg 11φg
Fine to Coarse: Operator H• Recursive Computation of Inner Products with φ
kjn
knkj hff ,12, ,, −−∑= φφ
∑= −−n
kjkn fh ,12 ,φ
( ) 1 2 −↓= ∗ jn
jn cHc
• Discrete Convolution and Decimation
• Change of Representation
1−⎯⎯← jn
Hjn cc
1−⊂ jj VV
Fine to Detail: Operator G• Recursive Computation of Inner Products with ψ
kjn
knkj gff ,12, ,, −−∑= φψ
∑= −−n
kjkn fg ,12 ,φ
( ) 1 2 −↓= ∗ jn
jn cGd
• Discrete Convolution and Decimation
• Change of Representation
1−⎯⎯← jn
Gjn cd
1−⊂ jj VW
Wavelet Decomposition
• Assume:
!
!
21
21
++
+⎯→⎯+⎯→⎯
jn
jn
GG
jn
Hjn
Hjn
dd
ccc
kjjn fc ,,φ=• Start with:
• Decomposition Scheme
jj Vf ∈
![Page 6: The Scientist Waveletslhf.impa.br/abel2017/velho-meyer.pdf · 2017. 8. 14. · orthogonal bases. Coifman, Daubechies and Stéphane Mallat went on to develop applications to many problems](https://reader033.vdocuments.net/reader033/viewer/2022052817/60aff4ed90577624d843170a/html5/thumbnails/6.jpg)
Coarse and Detail to Fine• Recover c j-1 from c j and d j
njjj
n fc ,111 , −−− = φ
njkjjkkj
jk dc ,1,, ,
−∑+∑= φψφ
( ) ( ) jn
jn
jn dGcHc 221 ↑+↑=− ∗∗
• Reconstruction Operator ( Expansion and Convolution )
∑+∑= −− njkjjknjkj
jk dc ,1,,1, ,, φψφφ
∑+∑= −−jkkn
jkkn dgch 22
Wavelet Reconstruction
1
1
++
⎯→⎯+⎯→⎯+
jn
mjn
GG
jn
Hjn
Hmjn
dd
ccc
!
!
( )1,,, +++ jmjmj ddc …• Start with:
• Decomposition Scheme
)
Wavelets and Filters
A Look in Frequency Domain
• Operators H and G
Discrete Filters
• Fast Wavelet Transform
Sub-band Filtering
• Signal Processing Framework
Filter Design Techniques
Implementation Scheme
Low-Pass FilterDiscrete Operator H
• Impulse Response
• Transfer Function
( ) ( )…… ,,,, 101 hhhhk −=
∑=−
k
ikkehH ωω)(
Periodic-2 is πH
High-Pass Filter
Discrete Operator G • Impulse Response
• Transfer Function
( ) ( )…… ,,,, 101 ggggk −=
∑=−
k
ikkegG ωω)(
Periodic-2 is πG
Multiresolution in Frequency Domain
],[ˆjjjjj aafVf −∈∈ ⇒
],[ˆ111 +++
−∈ jjj aaf
],[],[ˆ 111 jjjjj aaaao+++
∪−−∈
ja− ja
1+− ja 1+ja
ja− 1+− ja 1+ja ja
![Page 7: The Scientist Waveletslhf.impa.br/abel2017/velho-meyer.pdf · 2017. 8. 14. · orthogonal bases. Coifman, Daubechies and Stéphane Mallat went on to develop applications to many problems](https://reader033.vdocuments.net/reader033/viewer/2022052817/60aff4ed90577624d843170a/html5/thumbnails/7.jpg)
Two Channel Filter Banks
• One-Level Transform: Sub-band Filtering Perfect Reconstruction
c cj j
j+1H H
G G↓2 ↑2
↑2↓2 c
j+1d
Analysis Bank
Synthesis Bank
Tree-Structured Filter Bank
Wavelet Transform
• Recursive Filtering • Same Filters (invariance: dyadic scale, integer shift)
c cj j
j+2
H
H
H
H
G
G
G
G
↓2
↓2
↑2
↑2↑2
↑2
↓2
↓2 c
j+1
j+2
d
d
Frequency Splitting• Recursive Filtering
Dyadic Bands Energy Preservation
0G1G
π
2H 2G
2π4π
Matrix Implementation of FWT
• Filtering with Subsampling
• Two Channel Filter Bank
• Tree-Structured Filter Banks
c j
j+2
H
H
G
G
↓2
↓2
↓2
↓2 c
j+1
j+2
d
d
Filtering with Subsampling
• Convolution and Decimation: y = (↓2)F x
• Filter Matrix : y = ↓F x
⎟⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜⎜
⎝
⎛
⎟⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜⎜
⎝
⎛
⎟⎟⎠
⎞⎜⎜⎝
⎛=⎟⎟⎠
⎞⎜⎜⎝
⎛
3
2
1
0
01
01
01
10
1
0
01000001
xxxx
ffff
ffff
yy
!!
!!
⎟⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜⎜
⎝
⎛
⎟⎟⎠
⎞⎜⎜⎝
⎛=⎟⎟⎠
⎞⎜⎜⎝
⎛
3
2
1
0
01
10
1
0
xxxx
ffff
yy
!!
Two-Channel Filter Bank
• Analysis Bank
• Analysis Matrix
⎟⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜⎜
⎝
⎛
⎟⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜⎜
⎝
⎛
=
⎟⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜⎜
⎝
⎛
03
02
01
00
01
10
01
10
11
10
11
10
cccc
gggg
hhhh
ddcc
!!
!!
⎟⎟⎠
⎞⎜⎜⎝
⎛↓↓
=⎟⎟⎠
⎞⎜⎜⎝
⎛↓↓
=GH
GHM)2()2(
• Synthesis BankT1 MMS == −
* Critical Sampling
![Page 8: The Scientist Waveletslhf.impa.br/abel2017/velho-meyer.pdf · 2017. 8. 14. · orthogonal bases. Coifman, Daubechies and Stéphane Mallat went on to develop applications to many problems](https://reader033.vdocuments.net/reader033/viewer/2022052817/60aff4ed90577624d843170a/html5/thumbnails/8.jpg)
Tree-Structured Filter Bank
• Decomposition and Reconstruction Matrices for Level j
• Example of Decomposition Sequence
( )⎟⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜⎜
⎝
⎛
⎟⎟⎠
⎞⎜⎜⎝
⎛⎟⎟⎠
⎞⎜⎜⎝
⎛=
⎟⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜⎜
⎝
⎛
07
06
05
04
03
02
01
00
13
12
11
10
21
20
30
30
84
4
6
2
c
c
c
c
c
c
c
c
d
d
d
d
d
d
d
c
MI
MI
M
⎟⎟⎠
⎞⎜⎜⎝
⎛=
−kn
kj I
MD
00
⎟⎟
⎠
⎞
⎜⎜
⎝
⎛==
−
−
kn
kjj I
MDR0
0T1
jNk 2with =
FWT Matrix Computation
• Direct Transform (Analysis Bank)
• Inverse Transform (Synthesis Bank)
• Data Vector( ))())(( mjjjj ddcv −= !
jj
j vDv =+1
110
0 += mmvRRRc !
1+= jj
j vRv
001
1 cDDDv mm !=+
)
Example
Haar: Simplest Case
• Scaling Function and Wavelet
• Transform Scheme
• Matrix Computation • Function Decomposition
• Change of Basis
Haar Wavelets
• Scaling Function and Wavelet
• Low / High Pass Filters
φ ψ
)1 ,1( 2
1=H )1- ,1( 2
1=G
Haar Transform
• Input:
c
c
c
c
c
c c0
0
0
1
1
2
0
1
0 00
0
0
11
2
1
1
2 3
d
d
d
• output: ),,,( 11
10
20
20 dddcf =
),,,( 03
02
01
00 ccccf =
Haar Decomposition
Projection
),,,( 03
02
01
00
0 ccccf =
),( 11
10
1 ddo =
),( 11
10
1 ccf =
)( 20
2 do =)( 20
2 cf =
Representation
![Page 9: The Scientist Waveletslhf.impa.br/abel2017/velho-meyer.pdf · 2017. 8. 14. · orthogonal bases. Coifman, Daubechies and Stéphane Mallat went on to develop applications to many problems](https://reader033.vdocuments.net/reader033/viewer/2022052817/60aff4ed90577624d843170a/html5/thumbnails/9.jpg)
Haar Matrix Computation
• Normalization:
⎟⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜⎜
⎝
⎛
=
⎟⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜⎜
⎝
⎛
⎟⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜⎜
⎝
⎛
−
−⎟⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜⎜
⎝
⎛
−
1423
0219
11
rrrr
rrrr
rrrr
21=r
First LevelSecond Level
• Wavelet Matrix: (H and G)
Change of Basis Interpretation
⎟⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜⎜
⎝
⎛
−
+
⎟⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜⎜
⎝
⎛
−+
⎟⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜⎜
⎝
⎛
−
−+
⎟⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜⎜
⎝
⎛
=
⎟⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜⎜
⎝
⎛
=
1100
1
0011
4
1111
2
1111
3
0219
f
⎟⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜⎜
⎝
⎛
+
⎟⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜⎜
⎝
⎛
+
⎟⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜⎜
⎝
⎛
+
⎟⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜⎜
⎝
⎛
=
⎟⎟⎟⎟⎟
⎠
⎞
⎜⎜⎜⎜⎜
⎝
⎛
=
1000
0
0100
2
0010
1
0001
9
0219
f
• Scale Representation
• Multiresolution Representation
Representation in the Two Bases
Scale Basis
0V2V
2W
1W
Multiresolution Basis
Bi-Orthogonal Wavelets
• Represent with one Basis, Reconstruct with the Other
Bi-Orthonormal Basis• Dual Basis
• Computationally Similar to Orthonormal Basis
• More Degrees of Freedom to Construct Basis
Meyer Bi-orthogonal Wavelets
![Page 10: The Scientist Waveletslhf.impa.br/abel2017/velho-meyer.pdf · 2017. 8. 14. · orthogonal bases. Coifman, Daubechies and Stéphane Mallat went on to develop applications to many problems](https://reader033.vdocuments.net/reader033/viewer/2022052817/60aff4ed90577624d843170a/html5/thumbnails/10.jpg)
Discrete Filter Pairs Conclusions
Computational Scheme for Wavelets
• Multiresolution Analysis • Non-Redundant Wavelets
• Two-Channel Filter Banks
• Function Decompostion
• Wavelet Transform as a Basis Change
• Bi-Orthogonal Wavelets