Wedgelets and Image Compression€¦ · symb. 12337 9253 7582 1671 3342 35856. J I First Results...
Transcript of Wedgelets and Image Compression€¦ · symb. 12337 9253 7582 1671 3342 35856. J I First Results...
![Page 1: Wedgelets and Image Compression€¦ · symb. 12337 9253 7582 1671 3342 35856. J I First Results (2) Circles, WC, 533 B, PSNR: 27.50 dB Peppers, WC, 16983 B, PSNR: 32.10 dB. J I Work](https://reader036.fdocuments.us/reader036/viewer/2022090605/605a22985d295f78072d24d7/html5/thumbnails/1.jpg)
J
I
Wedgelets and Image CompressionLaurent Demaret, Mattia Fedrigo, Hartmut Führ
Summer school: New Trends and Directions in Harmonic Analysis,Approximation Theory, and Image Analysis,
Inzell, Germany, 20 September 2007
![Page 2: Wedgelets and Image Compression€¦ · symb. 12337 9253 7582 1671 3342 35856. J I First Results (2) Circles, WC, 533 B, PSNR: 27.50 dB Peppers, WC, 16983 B, PSNR: 32.10 dB. J I Work](https://reader036.fdocuments.us/reader036/viewer/2022090605/605a22985d295f78072d24d7/html5/thumbnails/2.jpg)
J
I
Contents
1. Image Compression: Basis
2. Wedgelet Segmentations
3. Data structure and compression
4. first results
![Page 3: Wedgelets and Image Compression€¦ · symb. 12337 9253 7582 1671 3342 35856. J I First Results (2) Circles, WC, 533 B, PSNR: 27.50 dB Peppers, WC, 16983 B, PSNR: 32.10 dB. J I Work](https://reader036.fdocuments.us/reader036/viewer/2022090605/605a22985d295f78072d24d7/html5/thumbnails/3.jpg)
J
I
Classical Compression Methods
Original Image JPEG (6.8 KB) JPEG2000 (6.5 KB)DCT FWT + Contexts
![Page 4: Wedgelets and Image Compression€¦ · symb. 12337 9253 7582 1671 3342 35856. J I First Results (2) Circles, WC, 533 B, PSNR: 27.50 dB Peppers, WC, 16983 B, PSNR: 32.10 dB. J I Work](https://reader036.fdocuments.us/reader036/viewer/2022090605/605a22985d295f78072d24d7/html5/thumbnails/4.jpg)
J
I
Zoom
Original Image JPEG (6.8 KB) JPEG2000 (6.5 KB)DCT FWT + Contexts
![Page 5: Wedgelets and Image Compression€¦ · symb. 12337 9253 7582 1671 3342 35856. J I First Results (2) Circles, WC, 533 B, PSNR: 27.50 dB Peppers, WC, 16983 B, PSNR: 32.10 dB. J I Work](https://reader036.fdocuments.us/reader036/viewer/2022090605/605a22985d295f78072d24d7/html5/thumbnails/5.jpg)
J
I
Mathematical Framework
Justification Approximation Theory
Image: real-valued function, continuous domainAnsatz natural images have some regularity
f ∈ X (quasi?)-Banach space , X ⊂⊂ L2(Ω)
Approximation fn =∑
i αiϕi,ϕ ∈ A, n-approximationWe look for A, such that
‖f − fn‖22 = O
(1
nα
), for some α > 0, and f ∈ X
Critics
I Asymptotic results
I Continuous vs Discrete
![Page 6: Wedgelets and Image Compression€¦ · symb. 12337 9253 7582 1671 3342 35856. J I First Results (2) Circles, WC, 533 B, PSNR: 27.50 dB Peppers, WC, 16983 B, PSNR: 32.10 dB. J I Work](https://reader036.fdocuments.us/reader036/viewer/2022090605/605a22985d295f78072d24d7/html5/thumbnails/6.jpg)
J
I
Old and new Ansätze
Orthogonal Transforms
I FOURIER: non optimal (bad for local singularities)
I WAVELETS : optimal Non Linear Approximation ratesfor Besov spaces and Bounded Variation
+ in 2D Isotropic vs Anisotropic Methods=⇒ Structure of the set of singularities
Geometrical Methods
I TRIANGULATIONS: good theoretical Approximation rates(Mallat2004, Demaret-Iske 2006)
I CURVELETS: not adaptive but quasi-optimal(contains flexible geometrical features) (Candès2004)
I BANDELETS (Mallat-LePennec1999)
I WEDGELETS (Donoho1999), (Lissowska)
![Page 7: Wedgelets and Image Compression€¦ · symb. 12337 9253 7582 1671 3342 35856. J I First Results (2) Circles, WC, 533 B, PSNR: 27.50 dB Peppers, WC, 16983 B, PSNR: 32.10 dB. J I Work](https://reader036.fdocuments.us/reader036/viewer/2022090605/605a22985d295f78072d24d7/html5/thumbnails/7.jpg)
J
I
![Page 8: Wedgelets and Image Compression€¦ · symb. 12337 9253 7582 1671 3342 35856. J I First Results (2) Circles, WC, 533 B, PSNR: 27.50 dB Peppers, WC, 16983 B, PSNR: 32.10 dB. J I Work](https://reader036.fdocuments.us/reader036/viewer/2022090605/605a22985d295f78072d24d7/html5/thumbnails/8.jpg)
J
I
Wavelets and Contours
Wedge (left) and its Wavelet coefficients (right)
![Page 9: Wedgelets and Image Compression€¦ · symb. 12337 9253 7582 1671 3342 35856. J I First Results (2) Circles, WC, 533 B, PSNR: 27.50 dB Peppers, WC, 16983 B, PSNR: 32.10 dB. J I Work](https://reader036.fdocuments.us/reader036/viewer/2022090605/605a22985d295f78072d24d7/html5/thumbnails/9.jpg)
J
I
Geometrical Segmentations
S ⊂ Z2 set of pixelsf ∈ RS imageP family of partitions P ⊂ 2S of S
fP ∈ RS best constant approximation with fP |r constant, r ∈ PS segmentations (P , fP)
γ ≥ 0 penalisation parameter
I Goal: Efficient Minimisation of the penalised Functional
Hf,γ : S → R, (P , fP) 7→ γ · |P| + ‖f − fP‖22 (γ ≥ 0).
I Result(P , fP) ∈ argmin
(P ,fP)
Hz,γ
optimal tradeoff between penalisation and reconstruction quality
![Page 10: Wedgelets and Image Compression€¦ · symb. 12337 9253 7582 1671 3342 35856. J I First Results (2) Circles, WC, 533 B, PSNR: 27.50 dB Peppers, WC, 16983 B, PSNR: 32.10 dB. J I Work](https://reader036.fdocuments.us/reader036/viewer/2022090605/605a22985d295f78072d24d7/html5/thumbnails/10.jpg)
J
I
Wedgelet Segmentations
Hf,γ : S → R, (P , fP) 7→ γ · |P| + ‖f − fP‖22 (γ ≥ 0).
I Problem Size of the search space :(|P| > 2|S|!)
B MCMC: slow and not exact
I Restriction of the search space
B discrete wedges
B nested Quadtree structure
I fast moment computation: Green-like formula
![Page 11: Wedgelets and Image Compression€¦ · symb. 12337 9253 7582 1671 3342 35856. J I First Results (2) Circles, WC, 533 B, PSNR: 27.50 dB Peppers, WC, 16983 B, PSNR: 32.10 dB. J I Work](https://reader036.fdocuments.us/reader036/viewer/2022090605/605a22985d295f78072d24d7/html5/thumbnails/11.jpg)
J
I
Representation Elements
DCT basis (JPEG) (Haar) Wavelet basis Wedgelet partitions
![Page 12: Wedgelets and Image Compression€¦ · symb. 12337 9253 7582 1671 3342 35856. J I First Results (2) Circles, WC, 533 B, PSNR: 27.50 dB Peppers, WC, 16983 B, PSNR: 32.10 dB. J I Work](https://reader036.fdocuments.us/reader036/viewer/2022090605/605a22985d295f78072d24d7/html5/thumbnails/12.jpg)
J
I
Data Structure
Quadtree Partition dyadic Wedge Partition
![Page 13: Wedgelets and Image Compression€¦ · symb. 12337 9253 7582 1671 3342 35856. J I First Results (2) Circles, WC, 533 B, PSNR: 27.50 dB Peppers, WC, 16983 B, PSNR: 32.10 dB. J I Work](https://reader036.fdocuments.us/reader036/viewer/2022090605/605a22985d295f78072d24d7/html5/thumbnails/13.jpg)
J
I
Example
f (W , fW) fW
![Page 14: Wedgelets and Image Compression€¦ · symb. 12337 9253 7582 1671 3342 35856. J I First Results (2) Circles, WC, 533 B, PSNR: 27.50 dB Peppers, WC, 16983 B, PSNR: 32.10 dB. J I Work](https://reader036.fdocuments.us/reader036/viewer/2022090605/605a22985d295f78072d24d7/html5/thumbnails/14.jpg)
J
I
Compression: Algorithm
Idee Wedgelet representation contains too much redundancies=⇒ Correlation Model between neighbours
ALGORITHM
I Tree Coding
I Model Coding
I IF (Model = constant over square)
(quantised) mean value encoded
I IF (Modell = constant over each Wedge)
Angle Encoding and relative position
Coding of the (quantised) mean values
![Page 15: Wedgelets and Image Compression€¦ · symb. 12337 9253 7582 1671 3342 35856. J I First Results (2) Circles, WC, 533 B, PSNR: 27.50 dB Peppers, WC, 16983 B, PSNR: 32.10 dB. J I Work](https://reader036.fdocuments.us/reader036/viewer/2022090605/605a22985d295f78072d24d7/html5/thumbnails/15.jpg)
J
I
Compression: Features
I mixed Models (e.g. square constant, wedge constant,wedge linear ...)
I corresponding penalisation : estimation of the coding costs
Hγ : (f, (P , fP)) 7−→ γ(∑
i
|C(Wi)|+∑
j
|C(Qj)|)+‖f−fP‖22, γ ≥ 0,
Wi : wedge, Qj : square, C estimator for the coding costs
I Coding
B combinatorial encoding
B angle coding : resolution-adaptive
I Prediction Method
![Page 16: Wedgelets and Image Compression€¦ · symb. 12337 9253 7582 1671 3342 35856. J I First Results (2) Circles, WC, 533 B, PSNR: 27.50 dB Peppers, WC, 16983 B, PSNR: 32.10 dB. J I Work](https://reader036.fdocuments.us/reader036/viewer/2022090605/605a22985d295f78072d24d7/html5/thumbnails/16.jpg)
J
I
Prediction
I Observation Representation still strongly redundant
B "not natural", arbitrary quadtree structure
I Main idea
B Multiresolution differential codingonly "Brotherhood" correlations
B Extraction of spatial correlation between quadtree "cousins"
I Predictability of the current piece from the causal (i.e. alreadycoded) information
![Page 17: Wedgelets and Image Compression€¦ · symb. 12337 9253 7582 1671 3342 35856. J I First Results (2) Circles, WC, 533 B, PSNR: 27.50 dB Peppers, WC, 16983 B, PSNR: 32.10 dB. J I Work](https://reader036.fdocuments.us/reader036/viewer/2022090605/605a22985d295f78072d24d7/html5/thumbnails/17.jpg)
J
I
How to Code the Leaves ?
Tree Levels of the leaves
![Page 18: Wedgelets and Image Compression€¦ · symb. 12337 9253 7582 1671 3342 35856. J I First Results (2) Circles, WC, 533 B, PSNR: 27.50 dB Peppers, WC, 16983 B, PSNR: 32.10 dB. J I Work](https://reader036.fdocuments.us/reader036/viewer/2022090605/605a22985d295f78072d24d7/html5/thumbnails/18.jpg)
J
I
Predictive Coding: an Illustration
0 0 0 0 1 0 0 00 0 0 0 1 0 0 00 0 0 1 0 0 0 00 0 0 1 0 0 0 00 0 0 1 0 0 0 00 0 0 0 1 0 0 00 0 0 0 1 0 0 00 0 0 1 0 0 0 0
Context= 0 > 0
0 41 151 1 7
I Binary Tree : 45 bits
I Bottom to Top Non-Predictive: log2(64) + log2((64
8
)) = 39 bits
I Bottom to Top Predictive: log2(64) + log2((42
1
)) + log2(
(227
)) = 31 bits
![Page 19: Wedgelets and Image Compression€¦ · symb. 12337 9253 7582 1671 3342 35856. J I First Results (2) Circles, WC, 533 B, PSNR: 27.50 dB Peppers, WC, 16983 B, PSNR: 32.10 dB. J I Work](https://reader036.fdocuments.us/reader036/viewer/2022090605/605a22985d295f78072d24d7/html5/thumbnails/19.jpg)
J
I
First results (1)Comparison between "pure Wedge" and "Wedge+Constant" Models with higher penalisa-tion for Wedges versus Squares
C(Wi) = 3.5× C(Qi)
(a) Original Image (b) Squares: 87784 b, 30,54 dB(c) only Wedges: 84632 b, 30,42 dB (d)Wedges + Squares: 76184 b, 30,60 dB
Model Tree Models Const. Angles Line Wedge Totalvalues number values
pure squares bits 16960 70336 87296symb. 20801 15601 36402
pure wedges bits 5776 9808 20376 48296 84256symb. 7021 5266 5266 10532 28085
wedges + squares bits 10088 6440 34824 2928 6784 14904 75968symb. 12337 9253 7582 1671 1671 3342 35856
![Page 20: Wedgelets and Image Compression€¦ · symb. 12337 9253 7582 1671 3342 35856. J I First Results (2) Circles, WC, 533 B, PSNR: 27.50 dB Peppers, WC, 16983 B, PSNR: 32.10 dB. J I Work](https://reader036.fdocuments.us/reader036/viewer/2022090605/605a22985d295f78072d24d7/html5/thumbnails/20.jpg)
J
I
First Results (2)
Circles, WC, 533 B, PSNR: 27.50 dB Peppers, WC, 16983 B, PSNR: 32.10 dB
![Page 21: Wedgelets and Image Compression€¦ · symb. 12337 9253 7582 1671 3342 35856. J I First Results (2) Circles, WC, 533 B, PSNR: 27.50 dB Peppers, WC, 16983 B, PSNR: 32.10 dB. J I Work](https://reader036.fdocuments.us/reader036/viewer/2022090605/605a22985d295f78072d24d7/html5/thumbnails/21.jpg)
J
I
Work in Progress
I systematic investigation of the penalisation functional
B rate-distortionOptimisation
B Depends on the resolution
B Contexts change penalty
I Contextual Encoding
I Compression with richer regression models (e.g. linear)
B aim: avoid bloc artefacts
I Correct theoretical framework for discrete Data
B non asymptotical results