contourlets: construction and...
TRANSCRIPT
Contourlets: Construction and Properties
Minh N. Do
Department of Electrical and Computer EngineeringUniversity of Illinois at Urbana-Champaign
www.ifp.uiuc.edu/∼[email protected]
Joint work with Martin Vetterli (EPFL and UC Berkeley),Arthur Cunha, Yue Lu, Jianping Zhou (UIUC)
Outline
1. Motivation
2. Discrete-domain construction using filter banks
3. Contourlets and directional multiresolution analysis
4. Contourlet approximation
5. Contourlet filter design with directional vanishing moments
1. Motivation 1
What Do Image Processors Do for Living?
Compression: At 158:1 compression ratio...
1. Motivation 2
What Do Image Processors Do for Living?
Denoising (restoration/filtering)
Noisy image Clean image
1. Motivation 3
What Do Image Processors Do for Living?
Feature extraction (e.g. for content-based image retrieval)
Featureextraction
Featureextraction
Images
Query
Similarity Measurement
Signatures
N best matched images
1. Motivation 4
Fundamental Question: Parsimonious Representation of
Visual Information
A randomly generated image A natural image
Natural images live in a very tiny bit of the huge “image space” (e.g.R
512×512×3)
1. Motivation 5
Mathematical Foundation: Sparse Representations
Fourier, Wavelets... = construction of bases for signal expansions:
f =∑
n
cnψn, where cn = 〈f, ψn〉.
Non-linear approximation:
fM =∑
n∈IM
cnψn, where IM : indexes of best M components.
Sparse representation: How fast ‖f − fM‖ → 0 as M →∞.
1. Motivation 6
Wavelets and Filter Banks
2
2
2
2x
analysis synthesis
H1
H0
G1
G0
y1
y0 x0
x1
+ xMallat
G12
G02
G12
G02
++ Daubechies
1. Motivation 7
The Success of Wavelets
• Wavelets provide a sparse representation for piecewise smooth signals.
• Multiresolution, tree structures, fast transforms and algorithms, etc.
• Unifying theory ⇒ fruitful interaction between different fields.
1. Motivation 8
Fourier vs. Wavelets
Non-linear approximation: N = 1024 data samples; keep M = 128coefficients
0 100 200 300 400 500 600 700 800 900 1000−20
0
20
40
Original
0 100 200 300 400 500 600 700 800 900 1000−20
0
20
40
Using wavelets (Daubechies−4): SNR = 37.67 dB
0 100 200 300 400 500 600 700 800 900 1000−20
0
20
40
Using Fourier (of size 1024): SNR = 22.03 dB
Approximation movie!
1. Motivation 9
Is This the End of the Story?
1. Motivation 10
Wavelets in 2-D
• In 1-D: Wavelets are well adapted to abrupt changes or singularities.
• In 2-D: Separable wavelets are well adapted to point-singularities (only).
But, there are (mostly) line- and curve-singularities...
1. Motivation 11
The Failure of Wavelets
Wavelets fail to capture the geometrical regularity in images andmultidimensional data.
1. Motivation 12
Edges vs. Contours
Wavelets (with nonlinear approximations) cannot “see” the differencebetween these two images.
• Edges: image points with discontinuity
• Contours: edges with localized and regular direction [Zucker et al.]
1. Motivation 13
Goal: Efficient Representation for Typical Images
with Smooth Contours
������������������������������������������������������������������������������������������������������������������������������������������������
������������������������������������������������������������������������������������������������������������������������������������������������regions
smooth
boundary
smooth
Goal: Exploring the intrinsic geometrical structure in natural images.
⇒ Action is at the edges!
1. Motivation 14
Wavelet vs. New Scheme
Wavelets New scheme
For images:
• Wavelet scheme... see edges but not smooth contours.
• New scheme... requires challenging non-separable constructions.
1. Motivation 15
And What The Nature Tells Us...
• Human visual system:
– Extremely efficient: 107 bits −→ 20-40 bits (per second).– Receptive fields are characterized as localized, multiscale and oriented.
• Sparse components of natural images (Olshausen and Field, 1996):
16 x 16 patches
from natural images
Search forSparse Code
1. Motivation 16
Recent Breakthrough from Harmonic Analysis:
Curvelets [Candes and Donoho, 1999]
• Optimal representation for functions in R2 with curved singularities.
• Key idea: parabolic scaling relation for C2 curves:
width ∝ length2
uv
l
w
u = u(v)
1. Motivation 17
“Wish List” for New Image Representations
• Multiresolution ... successive refinement
• Localization ... both space and frequency
• Critical sampling ... correct joint sampling
• Directionality ... more directions
• Anisotropy ... more shapes
Our emphasis is on discrete framework that leads toalgorithmic implementations.
1. Motivation 18
Outline
1. Motivation
2. Discrete-domain construction using filter banks
3. Contourlets and directional multiresolution analysis
4. Contourlet approximation
5. Contourlet filter design with directional vanishing moments
2. Discrete-domain construction using filter banks 19
Challenge: Being Digital!
Pixelization:
Digital directions:
2. Discrete-domain construction using filter banks 20
Proposed Computational Framework: Contourlets
In a nutshell: contourlet transform is an efficient directional multiresolutionexpansion that is digital friendly!
contourlets = multiscale, local and directional contour segments
• Starts with a discrete-domain construction that is amenable to efficientalgorithms, and then investigates its convergence to a continuous-domainexpansion.
• The expansion is defined on rectangular grids ⇒ seamless translationbetween the continuous and discrete worlds.
2. Discrete-domain construction using filter banks 21
Discrete-Domain Construction using Filter Banks
Idea: Multiscale and Directional Decomposition
• Multiscale step: capture point discontinuities, followed by...
• Directional step: link point discontinuities into linear structures.
2. Discrete-domain construction using filter banks 22
Analogy: Hough Transform in Computer Vision
Input image Edge image “Hough” image
=⇒ =⇒
Challenges:
• Perfect reconstruction.
• Fixed transform with low redundancy.
• Sparse representation for images with smooth contours.
2. Discrete-domain construction using filter banks 23
Multiscale Decomposition using Laplacian Pyramids
decomposition reconstruction
• Reason: avoid “frequency scrambling” due to (↓) of the HP channel.
• Laplacian pyramid as a frame operator → tight frame exists.
• New reconstruction: efficient filter bank for dual frame (pseudo-inverse).
2. Discrete-domain construction using filter banks 24
Directional Filter Banks (DFB)
• Feature: division of 2-D spectrum into fine slices using tree-structuredfilter banks.
37
6
5
4 7
6
5
4
0123
0 1 2
ω1
ω2 (π, π)
(−π,−π)
• Background: Bamberger and Smith (’92) cleverly used quincunx FB’s,modulation and shearing.
• We propose:
– a simplified DFB with fan FB’s and shearing– use DFB to construct directional bases
2. Discrete-domain construction using filter banks 25
Multichannel View of the Directional Filter Bank
+x
H0
H1
H2l−1
G0
G1
G2l−1
S0S0
S1S1
S2l−1S2l−1
y0
y1
y2l−1
x
Use two separable sampling matrices:
Sk =
[
2l−1 0
0 2
]
0 ≤ k < 2l−1 (“near horizontal” direction)
[
2 0
0 2l−1
]
2l−1 ≤ k < 2l (“near vertical” direction)
2. Discrete-domain construction using filter banks 26
General Bases from the DFB
An l-levels DFB creates a local directional basis of l2(Z2):
{
g(l)k [· − S
(l)k n]
}
0≤k<2l, n∈Z2
• G(l)k are directional filters:
• Sampling lattices (spatial tiling):
2
2
2l−1
2l−1
2. Discrete-domain construction using filter banks 27
Pyramidal Directional Filter Banks (PDFB)
Motivation: + add multiscale into the directional filter bank+ improve its non-linear approximation power.
image
subbandsdirectionalbandpass
directionalsubbands
bandpass
(2,2) ω1
ω2 (π, π)
(−π,−π)
Properties: + Flexible multiscale and directional representation for images(can have different number of directions at each scale!)
+ Tight frame with small redundancy (< 33%)
+ Computational complexity: O(N) for N pixels.
2. Discrete-domain construction using filter banks 28
Wavelets vs. Contourlets
50 100 150 200 250
50
100
150
200
250
50 100 150 200 250
50
100
150
200
250
2. Discrete-domain construction using filter banks 29
Examples of Discrete Contourlet Transform
2. Discrete-domain construction using filter banks 30
Outline
1. Motivation
2. Discrete-domain construction using filter banks
3. Contourlets and directional multiresolution analysis
4. Contourlet approximation
5. Contourlet filter design with directional vanishing moments
3. Contourlets and directional multiresolution analysis 31
Multiresolution Analysis: Laplacian Pyramid
w1
w2
Vj
Wj
Vj−1
Vj−1 = Vj ⊕Wj,
L2(R2) =⊕
j∈Z
Wj.
Vj has an orthogonal basis {φj,n}n∈Z2, where
φj,n(t) = 2−jφ(2−jt− n).
Wj has a tight frame {µj−1,n}n∈Z2 where
µj−1,2n+ki= ψ
(i)j,n, i = 0, . . . , 3.
3. Contourlets and directional multiresolution analysis 32
Directional Multiresolution Analysis: LP + DFB
��������������������������������������������������������
��������������������������������������������������������
���������������������������������������������������������������
���������������������������������������������������������������
replacements
Vj
Wj
Vj−1
W(lj)
j,k
ω1
ω2
2j
2j+lj−2
ρ(lj)
j,k,n
Wj =2
lj−1⊕
k=0
W(lj)
j,k
W(lj)
j,k has a tight frame{
ρ(l)j,k,n
}
n∈Z2where
ρ(l)j,k,n(t) =
∑
m∈Z2
g(l)k [m− S
(l)k n]
︸ ︷︷ ︸DFB basis
µj−1,m(t)︸ ︷︷ ︸
LP frame
= ρ(l)j,k(t− 2j−1
S(l)k n).
3. Contourlets and directional multiresolution analysis 33
Contourlet Frames
Theorem (Contourlet Frames) [DoV:03].{
ρ(lj)
j,k,n
}
j∈Z, 0≤k<2lj , n∈Z2
is a tight frame of L2(R2) for finite lj.
Theorem (Connection with Filter Banks) [DoV:04]Suppose x[n] = 〈f, φL,n〉, n ∈ Z
2, for some function f ∈ L2(R2).Furthermore, suppose
xPDFB−→ (aJ , d
(lj)
j,k )j=1,...,J; k=0,...,2
lj−1
where aJ is the lowpass subband, and d(lj)
j,k are bandpass directionalsubbands. Then
aJ [n] = 〈f, φL+J,n〉
d(lj)
j,k [n] = 〈f, ρ(lj)
L+j,k,n〉
for j = 1, . . . , J ; k = 0, . . . , 2lj − 1, n ∈ Z2.
3. Contourlets and directional multiresolution analysis 34
Sampling Grids of Contourlets
(a) (b)
(c) (d)
w
w
l
l
w/4
w/4
l/2
l/2
3. Contourlets and directional multiresolution analysis 35
Contourlet Features
• Defined via iterated filter banks ⇒ fast algorithms, tree structures, etc.
• Defined on rectangular grids ⇒ seamless translation between continuousand discrete worlds.
• Different contourlet kernel functions (ρj,k) for different directions.
• These functions are defined iteratively via filter banks.
• With FIR filters ⇒ compactly supported contourlet functions.
3. Contourlets and directional multiresolution analysis 36
Contourlet Packets
• Adaptive scheme to select the “best” tree for directional decomposition.
ω1
ω2 (π, π)
(−π,−π)
• Contourlet packets ⇒ directional multiresolution elements with differentshapes (aspect ratios).
– They do not necessary satisfy the parabolic relation.– They can include wavelets!
3. Contourlets and directional multiresolution analysis 37
Outline
1. Motivation
2. Discrete-domain construction using filter banks
3. Contourlets and directional multiresolution analysis
4. Contourlet approximation
5. Contourlet filter design with directional vanishing moments
4. Contourlet approximation 38
Contourlets with Parabolic Scaling
Support size of the contourlet function ρljj,k: width ≈ 2j and length ≈ 2lj+j
To satisfy the parabolic scaling (for C2 curved singularities):width ∝ length2, simply set:
the number of directions in the PDFB is doubled at every other finer
scale.
uv
l
w
u = u(v)
ω1
ω2 (π, π)
(−π,−π)
4. Contourlet approximation 39
Supports of Contourlet Functions
*
Contourlet
=*
DFBLP
=
Key point: Each generation doubles spatial resolutionas well as angular resolution.
4. Contourlet approximation 40
Contourlet Approximation
see a wavelet
ΠLΠL
Desire: Fast decay as contourlets turn away from the discontinuity direction
Key: Directional vanishing moments (DVMs)
4. Contourlet approximation 41
Geometrical Intuition
At scale 2j (j ≪ 0):
width ≈ 2j
length ≈ 2j/2
#directions ≈ 2−j/2
���������������
���������������
2j
2j/2 dj,k,n
d2j,k,n
θj,k,n
|〈f, ρj,k,n〉| ∼ 2−3j/4 · d3j,k,n
dj,k,n ∼ 2j/ sin θj,k,n ∼ 2j/2k−1 for k = 1, . . . , 2−j/2
=⇒ |〈f, ρj,k,n〉| ∼ 23j/4k−3
4. Contourlet approximation 42
How Many DVMs Are Sufficient?
d
d^2
������������������������
������������������������
alpha
�������������������������������������������������
�������������������������������������������������
�������������������������������������������������
�������������������������������������������������
ω1
ω2
Sufficient if the gap to a direction with DVM:
α . d ∼ 2j/2k−1 for k = 1, . . . , 2−j/2
This condition can be replaced with fast decay in frequency across directions.
It is still an open question if there is an FIR filter bank that satisfies thesufficient DVM condition
4. Contourlet approximation 43
Experiments with Decay Across Directions using
Near Ideal Frequency Filters
0 0.5 1 1.5 2 2.5 3 3.5 4−18
−16
−14
−12
−10
−8
−6Disc 256 coefficients decaying plot (average)
Log2
(|<
f, h>
|)
Log2(k)
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5−22
−20
−18
−16
−14
−12
−10
−8
−6Disc 256 coefficients decaying plot (average)
Log2
(|<
f, h>
|)
Log2(k)
4. Contourlet approximation 44
Nonlinear Approximation Rates
��������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������
��������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������C2 regions C2 boundary
Under the (ideal) sufficient DVM condition
|〈f, ρj,k,n〉| ∼ 23j/4k−3
with number of coefficients Nj,k ∼ 2−j/2k. Then
‖f − f(contourlet)M ‖2 ∼ (logM)3M−2
While ‖f − f(Fourier)M ‖2 ∼ O(M−1/2) and ‖f − f
(wavelet)M ‖2 ∼ O(M−1)
4. Contourlet approximation 45
Non-linear Approximation Experiments
Image size = 512× 512. Keep M = 4096 coefficients.
Original image Wavelets: Contourlets:PSNR = 24.34 dB PSNR = 25.70 dB
4. Contourlet approximation 46
Detailed Non-linear Approximations
Wavelets
M = 4 M = 16 M = 64 M = 256
Contourlets
M = 4 M = 16 M = 64 M = 256
4. Contourlet approximation 47
Denoising Experiments
original image noisy image (SNR = 9.55 dB)
wavelet denoising (SNR = 13.82 dB) contourlet denoising (SNR = 15.42 dB)
4. Contourlet approximation 48
Embedded Structure for Compression and Modeling
• So far, best-M term approximation:
fM =∑
λ∈IM
cλρλ, where IM is the set of indexes of the M -largest |cλ|.
• For compression, additional cost required to specify IM
– Naive approach: M · log2N bits– With embedded tree (wavelets): M bits
• Embedded trees for wavelets are crucial in state-of-the-art imagecompression (EZW,...), rate-distortion analysis (Cohen et al.), andmultiscale statistical modeling (Baraniuk et al.)
4. Contourlet approximation 49
Contourlet Embedded Tree Structure
Embedded tree data structure forcontourlet coefficients:
successively locate the positionand direction of image contours.
Since significant contourlet coefficientsare organized in trees, best M -treeapproximation (using M -node tree):
‖f − f(contourlet)M−tree ‖2 ≈ (logM)3M−2
⇒ D(R) ≈ (logR)3R−2
4. Contourlet approximation 50
Outline
1. Motivation
2. Discrete-domain construction using filter banks
3. Contourlets and directional multiresolution analysis
4. Contourlet approximation
5. Contourlet filter design with directional vanishing moments
5. Contourlet filter design with directional vanishing moments 51
Filter Bank Design Problem
ρ(l)j,k(t) =
∑
m∈Z2
c(l)k [m]φj−1,m(t)
ρ(l)j,k(t) has an L-order DVM along direction (u1, u2)
⇔ C(l)k (z1, z2) = (1− zu2
1 z−u12 )LR(z1, z2)
So far: Use good frequency selectivityto approximate DVMs.
Draw back: long filters...
��������������������������������������������������������
��������������������������������������������������������
���������������������������������������������������������������
���������������������������������������������������������������
ω1
ω2
Next: Design short filters that lead to many DVMs as possible.
5. Contourlet filter design with directional vanishing moments 52
Filters with DVMs = Directional Annihilating Filters
Input image and after being filtered by a directional annihilating filter
5. Contourlet filter design with directional vanishing moments 53
Perfect Reconstruction Two-Channel FBs with DVMs
Filter bank with order-L horizontal or vertical DVM:
T0
Q
Q
Q
Q
T1
Filter design amounts to solve
P (z) + P (−z) = 2,
where P (z) = H0(z)G0(z) = (1− z1)LR(z).
To get DVMs at other directions: Shearing or change of variables
≡
H0(ω) H0(RT0 ω)
R0 R1
5. Contourlet filter design with directional vanishing moments 54
Complete Characterization
Proposition [Cunha-D., 04]. Any FIR solution of
(1− z1)LR(z) + (1 + z1)
LR(−z) = 2 (1)
can be written as
R(z) = AL(z1) + (1 + z1)LB(z)
where AL(z1) is the minimum degree 1-D polynomial in z1 that solves (1)
AL(z1) =
L−1∑
i=0
(L+ i− 1
L− 1
)
2−(L+i−1)(1− z1)i,
and B(z) satisfy B(z) +B(−z) = 0.
Proof. Applying the Bezout theorem for each z2.
5. Contourlet filter design with directional vanishing moments 55
Design via Mapping (Cunha-D., 2004)
To avoid 2D factorization...
1. Start with a 1-D solution:
P (1D)(z) + P (1D)(−z) = 2 and P (1D)(z) = H(1D)0 (z)G
(1D)0 (z)
2. Apply a 1-D to 2-D mapping to each filter:
F (1D)(z) → F (2D)(z) = F (1D)(M(z1, z2))
• If M(z) = −M(−z) then PR is preserved
P (2D)(z) + P (2D)(−z) = 2
• Suppose H(1D)(z) = (1 + z)kR1(z) and M(z) + 1 = (1− z1)lm(z) then
H(2D)(M(z)) = (1− z1)klR2(z)
5. Contourlet filter design with directional vanishing moments 56
How To Design the Mapping
The mapping has to satisfy
M(z) = −M(−z), and
M(z) = (1− z1)lm(z)− 1
Thus,(1− z1)
lm(z) + (1 + z1)lm(−z) = 2
The last proposition tell us exactly how to solve this!
5. Contourlet filter design with directional vanishing moments 57
Design Examples
5. Contourlet filter design with directional vanishing moments 58
Directional Vanishing Moments Generated After
Iteration
Different expanding rules lead to different set of directions with DVMs.
5. Contourlet filter design with directional vanishing moments 59
Gain by using Filters with DVMs
0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
10
15
20
25
% coeff
SN
R [d
B]
dvmpkva
0 0.02 0.04 0.06 0.08 0.1 0.124
6
8
10
12
14
16
% coeff
SN
R [d
B]
dvmpkva
5. Contourlet filter design with directional vanishing moments 60
Summary
• Image processing relies on prior information about images.
– Geometrical structure is the key!
• Strong motivation for more powerful image representations:scale, space, and direction.
– New desideratum beyond wavelets: localized direction
• New two-dimensional discrete framework and algorithms:
– Flexible directional and multiresolution image representation.
– Effective for images with smooth contours ⇒ contourlets.
• Dream: Another fruitful interaction between harmonic analysis, computervision, and signal processing.
61
References
• M. N. Do,“Directional multiresolution image representations,” Ph.D.thesis, EPFL, December 2001.
• M. N. Do and M. Vetterli, “Contourlets,” Beyond Wavelets, G. V.Welland ed., Academic Press, 2003.
• M. N. Do and M. Vetterli, “The contourlet transform: an efficientdirectional multiresolution image representation,” IEEE Trans. on
Image Proc., submitted 2003.
• A. Cunha and M. N. Do, “Biorthogonal two-channel filter bankswith directional vanishing moments: characterization, design andapplications,” IEEE Trans. on Image Proc., submitted 2004.
• Software and downloadable papers: www.ifp.uiuc.edu/∼minhdo
62