Gabor Frames and Directional Time-Frequency Analysis
(joint work with Chris Sansing)
Loukas Grafakos
University of Missouri
http://www.math.missouri.edu/∼loukas
Marriage
Gabor functions
ψm,τ(x) = e2πim·xψ(x − τ )
Ridgelets (Candès)
Marriage
Gabor functions
ψm,τ(x) = e2πim·xψ(x − τ )
Ridgelets (Candès)
ψa,t,u(x) = a−1/2ψ((x · u − t)/a)
Marriage
Gabor functions
ψm,τ(x) = e2πim·xψ(x − τ )
Ridgelets (Candès)
ψa,t,u(x) = a−1/2ψ((x · u − t)/a)
Marriage
Gabor functions
ψm,τ(x) = e2πim·xψ(x − τ )
Ridgelets (Candès)
ψa,t,u(x) = a−1/2ψ((x · u − t)/a)
General Idea
Radon transform
Rf (u, t)
General Idea
Radon transform
Rf (u, t)
Time-Frequency Analysis
The Radon Transform
An integral transform which is sensitive to direction.
Rf (u, s) = Z
u·x=s
The Radon Transform
An integral transform which is sensitive to direction.
Rf (u, s) = Z
u·x=s
The Radon Transform
An integral transform which is sensitive to direction.
Rf (u, s) = Z
u·x=s
f (x)dx
The Radon Transform
An integral transform which is sensitive to direction.
Rf (u, s) = Z
u·x=s
f (x)dx
• This transform was introduced in 1917 by Johann Radon.
The Radon Transform
An integral transform which is sensitive to direction.
Rf (u, s) = Z
u·x=s
f (x)dx
• This transform was introduced in 1917 by Johann Radon.
First attempt at a directional TF representation
Let g ∈ S(R) be some window function (i.e. Gaussian). Define
First attempt at a directional TF representation
Let g ∈ S(R) be some window function (i.e. Gaussian). Define
gm,t(s) = e2πim(s−t)g(s − t)
First attempt at a directional TF representation
Let g ∈ S(R) be some window function (i.e. Gaussian). Define
gm,t(s) = e2πim(s−t)g(s − t)
First attempt at a directional TF representation
Let g ∈ S(R) be some window function (i.e. Gaussian). Define
gm,t(s) = e2πim(s−t)g(s − t)
First attempt at a directional TF representation
Let g ∈ S(R) be some window function (i.e. Gaussian). Define
gm,t(s) = e2πim(s−t)g(s − t)
gm,t,u(x) = gm,t(u · x) = ridge functions
Z Sn−1 Z R Z R
First attempt at a directional TF representation
Let g ∈ S(R) be some window function (i.e. Gaussian). Define
gm,t(s) = e2πim(s−t)g(s − t)
The Radon transform presents itself naturally
When we pair our function f ∈ L1 ∩ L2(Rn) with g ∈ S(R) the Radon transform appears:
The Radon transform presents itself naturally
When we pair our function f ∈ L1 ∩ L2(Rn) with g ∈ S(R) the Radon transform appears:
hf, gm,t,ui = hRuf, gm,ti
hf, gm,t,ui = Z
Rn
f (x)e−2πim(u·x−t)g(u · x − t)dx =
Z
R
Z
u·x=s
Attempt at Continuous Representation
Proposition 1 (Reproduction of the back-projection)
Let f ∈ L1 ∩ L2(Rn). Given functions g, ψ ∈ S(R), we have
Z Sn−1 Z R Z R
hf, gm,t,uiψm,t,udmdtdu = hg, ψi B(f )
where
Bf (x) = Z
Sn−1
Attempt at Continuous Representation
Proposition 1 (Reproduction of the back-projection)
Let f ∈ L1 ∩ L2(Rn). Given functions g, ψ ∈ S(R), we have
Z Sn−1 Z R Z R
hf, gm,t,uiψm,t,udmdtdu = hg, ψi B(f )
where
Bf (x) = Z
Sn−1
Ruf (u · x)du = R∗Rf (x)
Here R∗ is the adjoint of the Radon transform given by
R∗h(x) = Z
Sn−1
Attempt at Continuous Representation
Proposition 1 (Reproduction of the back-projection)
Let f ∈ L1 ∩ L2(Rn). Given functions g, ψ ∈ S(R), we have
Z Sn−1 Z R Z R
hf, gm,t,uiψm,t,udmdtdu = hg, ψi B(f )
where
Bf (x) = Z
Sn−1
Ruf (u · x)du = R∗Rf (x)
Here R∗ is the adjoint of the Radon transform given by
R∗h(x) = Z
Sn−1
h(x · u, u) du
which satisfies for a function h on the cylinder R × Sn−1
What is
B(f )
?
What is
B(f )
?
What is
B(f )
?
B(f ) = R∗R(f ) = c f ∗ | · |−1(x)
• We are “off by a filter.”
• The blurring effect is caused by the weight which amplifies low (blurry) frequencies and attenuates the higher (sharp)
What is
B(f )
?
Weighted Gabor Ridge Functions
Weighted Gabor Ridge Functions
We need to introduce a filter (weight) to achieve perfect reconstruction. For this reason we define
Gm,t(s) = Dn−1 2 (g
m,t)(s),
Weighted Gabor Ridge Functions
We need to introduce a filter (weight) to achieve perfect reconstruction. For this reason we define
Gm,t(s) = Dn−1 2 (g
m,t)(s),
for some g ∈ S(R)
where m, t are real numbers and
Dn−1 2 (g
Weighted Gabor Ridge Functions
We need to introduce a filter (weight) to achieve perfect reconstruction. For this reason we define
Gm,t(s) = Dn−1 2 (g
m,t)(s),
for some g ∈ S(R)
where m, t are real numbers and
Dn−1 2 (g
m,t) = ( dgm,t(σ)|σ|n−1 2 )∨
Also define the weighted Gabor ridge functions
Weighted Gabor Ridge Functions
We need to introduce a filter (weight) to achieve perfect reconstruction. For this reason we define
Gm,t(s) = Dn−1 2 (g
m,t)(s),
for some g ∈ S(R)
where m, t are real numbers and
Dn−1 2 (g
m,t) = ( dgm,t(σ)|σ|n−1 2 )∨
Also define the weighted Gabor ridge functions
Exact Continuous Representation
Proposition 2 (Continuous Representation)
Given f ∈ L1 ∩ L2(Rn), and Gm,t,u, Ψm,t,u weighted Gabor ridge functions, we have 2f (x) = 1 hg, ψi Z Sn−1 Z R Z R
Exact Continuous Representation
Proposition 2 (Continuous Representation)
Given f ∈ L1 ∩ L2(Rn), and Gm,t,u, Ψm,t,u weighted Gabor ridge functions, we have 2f (x) = 1 hg, ψi Z Sn−1 Z R Z R
hf, Gm,t,uiΨm,t,udmdtdu
1 hg, ψi Z Sn−1 Z R Z R
hf, Gm,t,uiΨm,t,udmdtdu
= 1 hg, ψi Z Sn−1 Z R Z R
hRuf, Gm,tiΨm,t,udmdtdu
= 1 hg, ψi Z Sn−1 Z R Z R hDn−1 2 (Ruf ), g m,tiD n−1 2 (ψ m,t)(u · x)dmdtdu = Dn−1 2 D n−1 2 Z Sn−1
Filtered Backprojection formula
Z Sn−1 Dn−1(Ruf )(u · x)du = Z Sn−1 Z R d Ruf (σ)|σ|n−1e2πiσ(u·x)dσdu = Z Sn−1 Z R bFiltered Backprojection formula
Z Sn−1 Dn−1(Ruf )(u · x)du = Z Sn−1 Z R d Ruf (σ)|σ|n−1e2πiσ(u·x)dσdu = Z Sn−1 Z R bf (σu)|σ|n−1e2πiσ(u·x)dσdu = 2 Z Rn b f (ξ)e2πiξxdξ = 2f (x)
This continuous representation is valid for any g, ψ ∈ S(R) with
Filtered Backprojection formula
Z Sn−1 Dn−1(Ruf )(u · x)du = Z Sn−1 Z R d Ruf (σ)|σ|n−1e2πiσ(u·x)dσdu = Z Sn−1 Z R bf (σu)|σ|n−1e2πiσ(u·x)dσdu = 2 Z Rn b f (ξ)e2πiξxdξ = 2f (x)
This continuous representation is valid for any g, ψ ∈ S(R) with
Filtered Backprojection formula
Z Sn−1 Dn−1(Ruf )(u · x)du = Z Sn−1 Z R d Ruf (σ)|σ|n−1e2πiσ(u·x)dσdu = Z Sn−1 Z R bf (σu)|σ|n−1e2πiσ(u·x)dσdu = 2 Z Rn b f (ξ)e2πiξxdξ = 2f (x)
This continuous representation is valid for any g, ψ ∈ S(R) with
Filtered Backprojection formula
Z Sn−1 Dn−1(Ruf )(u · x)du = Z Sn−1 Z R d Ruf (σ)|σ|n−1e2πiσ(u·x)dσdu = Z Sn−1 Z R bf (σu)|σ|n−1e2πiσ(u·x)dσdu = 2 Z Rn b f (ξ)e2πiξxdξ = 2f (x)
This continuous representation is valid for any g, ψ ∈ S(R) with
Filtered Backprojection formula
Z Sn−1 Dn−1(Ruf )(u · x)du = Z Sn−1 Z R d Ruf (σ)|σ|n−1e2πiσ(u·x)dσdu = Z Sn−1 Z R bf (σu)|σ|n−1e2πiσ(u·x)dσdu = 2 Z Rn b f (ξ)e2πiξxdξ = 2f (x)
This continuous representation is valid for any g, ψ ∈ S(R) with
Semi-discrete Reproduction
Theorem 3 (Semi-discrete Reproduction)
There exist g, ψ ∈ S(R) and 0 < α, β < 1 such that for f ∈ L1 ∩ L2(Rn)
f = 1 2 Z Sn−1 X m∈Z X t∈Z
Semi-discrete Reproduction
Theorem 3 (Semi-discrete Reproduction)
There exist g, ψ ∈ S(R) and 0 < α, β < 1 such that for f ∈ L1 ∩ L2(Rn)
f = 1 2 Z Sn−1 X m∈Z X t∈Z
hf, Gβm,αt,uiΨβm,αt,udu
Semi-discrete Reproduction
Theorem 3 (Semi-discrete Reproduction)
There exist g, ψ ∈ S(R) and 0 < α, β < 1 such that for f ∈ L1 ∩ L2(Rn)
f = 1 2 Z Sn−1 X m∈Z X t∈Z
hf, Gβm,αt,uiΨβm,αt,udu
Semi-discrete Representation
Semi-discrete Representation
Integrating over Sn−1 we deduce Z Sn−1 Dn−1(Ruf )(u · x) du = Z Sn−1 X m∈Z X t∈Z hDn−1 2 (Ruf ), g αm,βtiD n−1 2 (ψ αm,βt)(u · x) du or 2f (x) = Z Sn−1 X m∈Z X t∈Z
Semi-discrete Frame Identity
We also have the identity
Semi-discrete Frame Identity
We also have the identity
Semi-discrete Frame Identity
We also have the identity
Examples of Reconstruction
Examples of Reconstruction
Examples of Reconstruction
Examples of Reconstruction
Examples of Reconstruction
Half-filtered Operators
Half-filtered Operators
We introduce the following half-filtered operators: • Half-filtered Radon transform
Rf = Dn−1
2 (Rf (u, ·))(t) = D n−1
Half-filtered Operators
We introduce the following half-filtered operators: • Half-filtered Radon transform
Half-filtered Operators
We introduce the following half-filtered operators: • Half-filtered Radon transform
Rf = Dn−1 2 (Rf (u, ·))(t) = D n−1 2 (Rf )(u, t) • Half-filtered Backprojection R∗g = Z Sn−1 Dn−1 2 (g)(u, u · x) du
Operator/Inverse Operator relationship
R∗
Representation of the associated frame operator
Theorem 4 (Representation formula analogous to Walnut’s)
For f ∈ L1 ∩ L2(Rn), g, ψ ∈ W (R) the frame operator
Sg,ψf = Z Sn−1 X m∈Z X t∈Z
hf, Gβm,αt,uiΨβm,αt,udu
Representation of the associated frame operator
Theorem 4 (Representation formula analogous to Walnut’s)
For f ∈ L1 ∩ L2(Rn), g, ψ ∈ W (R) the frame operator
Sg,ψf = Z Sn−1 X m∈Z X t∈Z
hf, Gβm,αt,uiΨβm,αt,udu
can be written as Sg,ψf = R∗ Qf where Qf = β−1 X r∈Z Gr(s)Rf (s − r β)
Functional Spaces measuring Gabor ridge coefficients
Let 1 ≤ p, q ≤ ∞ and g be a fixed window. Recall the modulation spaces with norm
Functional Spaces measuring Gabor ridge coefficients
Let 1 ≤ p, q ≤ ∞ and g be a fixed window. Recall the modulation spaces with norm
khkMp,q(R) = Z R Z R |Vgh(x, ξ)|p dx q/p dξ !1/q = kVghkLp,q(R2)
Functional Spaces measuring Gabor ridge coefficients
Let 1 ≤ p, q ≤ ∞ and g be a fixed window. Recall the modulation spaces with norm
khkMp,q(R) = Z R Z R |Vgh(x, ξ)|p dx q/p dξ !1/q = kVghkLp,q(R2)
One is tempted to define spaces the measure the TF coefficients of the Radon transform in the following way:
kf kΩp,q,r(Rn) =
Z
kRuf krMp,q(R) du
1/r
More appropriate definition: Weighted Spaces
Ω
p,q,rA function f is in the space Ωp,q,rs (Rn) if the following norm is finite:
More appropriate definition: Weighted Spaces
Ω
p,q,rA function f is in the space Ωp,q,rs (Rn) if the following norm is finite:
kf kΩp,q,r s (Rn) = Z Sn−1 k( dRuf ωcs)∨ krMp,q(R)du 1/r , where ωs = (|σ| n−1 2 (1 + |σ|2) s 2)∨.
Also, we can create a discrete (unweighted) version.
k{aβm,αt,u}kωp,q,r(Z×Z×Sn−1) =
Z
Sn−1
k{aβm,αt,u}krℓp,qdu
More appropriate definition: Weighted Spaces
Ω
p,q,rA function f is in the space Ωp,q,rs (Rn) if the following norm is finite:
kf kΩp,q,r s (Rn) = Z Sn−1 k( dRuf ωcs)∨ krMp,q(R)du 1/r , where ωs = (|σ| n−1 2 (1 + |σ|2) s 2)∨.
Also, we can create a discrete (unweighted) version.
k{aβm,αt,u}kωp,q,r(Z×Z×Sn−1) =
Z
Sn−1
k{aβm,αt,u}krℓp,qdu
More appropriate definition: Weighted Spaces
Ω
p,q,rA function f is in the space Ωp,q,rs (Rn) if the following norm is finite:
kf kΩp,q,r s (Rn) = Z Sn−1 k( dRuf ωcs)∨ krMp,q(R)du 1/r , where ωs = (|σ| n−1 2 (1 + |σ|2) s 2)∨.
Also, we can create a discrete (unweighted) version.
Analysis and Synthesis operators
Introduce an analysis operator Cg by
Analysis and Synthesis operators
Introduce an analysis operator Cg by
Cg(f ) = hf, Gβm,αt,ui m,t,u
and a synthesis operator Dψ by
Analysis and Synthesis operators
Introduce an analysis operator Cg by
Cg(f ) = hf, Gβm,αt,ui m,t,u
and a synthesis operator Dψ by
Dψ({cm,t,u}) = Z Sn−1 X m∈Z X t∈Z cβm,αt,uΨβm,αt,udu
Analysis and Synthesis operators
Introduce an analysis operator Cg by
Cg(f ) = hf, Gβm,αt,ui m,t,u
and a synthesis operator Dψ by
Dψ({cm,t,u}) = Z Sn−1 X m∈Z X t∈Z cβm,αt,uΨβm,αt,udu
Analysis and Synthesis operators
Introduce an analysis operator Cg by
Cg(f ) = hf, Gβm,αt,ui m,t,u
and a synthesis operator Dψ by
Dψ({cm,t,u}) = Z Sn−1 X m∈Z X t∈Z cβm,αt,uΨβm,αt,udu
• The associated frame operator is Sg,ψ = Dψ ◦ Cg. • The operators Cg and Dψ are adjoint to each other.
Analysis and Synthesis operators
Introduce an analysis operator Cg by
Cg(f ) = hf, Gβm,αt,ui m,t,u
and a synthesis operator Dψ by
Dψ({cm,t,u}) = Z Sn−1 X m∈Z X t∈Z cβm,αt,uΨβm,αt,udu
Boundedness of the Analysis and Synthesis Operators
Theorem 5 (Boundedness of Cg)
If g ∈ M1,1(R), then Cg : Ωp,q,r → ωp,q,r for 1 ≤ p, q, r ≤ ∞ and
Boundedness of the Analysis and Synthesis Operators
Theorem 5 (Boundedness of Cg)
If g ∈ M1,1(R), then Cg : Ωp,q,r → ωp,q,r for 1 ≤ p, q, r ≤ ∞ and
kCgkΩp,q,r→ωp,q,r ≤ C(α, β)kVggkW (L1) independently of p, q and r. Here
Boundedness of the Analysis and Synthesis Operators
Theorem 5 (Boundedness of Cg)
If g ∈ M1,1(R), then Cg : Ωp,q,r → ωp,q,r for 1 ≤ p, q, r ≤ ∞ and
kCgkΩp,q,r→ωp,q,r ≤ C(α, β)kVggkW (L1) independently of p, q and r. Here
kgkW (L1) = X k∈Z ess sup x∈[0,1] |g(x + k)| Theorem 6 (Boundedness of Dψ)
Boundedness of the Frame Operator on
Ω
p,q,rCorollary Boundedness of Sg,ψ = Dψ ◦ Cg
If g, ψ ∈ M1,1(R), then the semi-discrete frame operator Sg,ψ is bounded on
Ωp,q,r for all 1 ≤ p, q, r ≤ ∞ and α, β > 0 with the following norm estimate:
Frame Extension to
Ω
p,q,rTheorem 7 Extension of frame to Ωp,q,r Assume g, ψ ∈ M1,1(R) and
Sg,ψ = I on L2(Rn). Then f = 1 2 Z Sn−1 X m∈Z X t∈Z
Frame Extension to
Ω
p,q,rTheorem 7 Extension of frame to Ωp,q,r Assume g, ψ ∈ M1,1(R) and
Sg,ψ = I on L2(Rn). Then f = 1 2 Z Sn−1 X m∈Z X t∈Z
hf, Gβm,αt,uiΨβm,αt,udu
Also, there are constants A, B > 0 such that for all f ∈ Ωp,q,r
Applications
Applications
Applications
Applications
Image Enhancing