SlideShare uma empresa Scribd logo
1 de 64
Sparse & Redundant Representation Modeling of Images: Theory and Applications   Michael Elad  The Computer Science Department  The Technion – Israel Institute of technology  Haifa 32000, Israel Seventh International Conference on     Curves and Surfaces Avignon - FRANCE June 24-30, 2010 This research was supported by the European Community's FP7-FET program SMALL under grant agreement no. 225913
2    This Talk Gives and Overview On …  A decade of tremendous progress in the field of  Sparse and Redundant Representations Numerical Problems Theory Applications Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
3 Agenda Part II – Theoretical & Numerical Foundations Part I – Denoising by Sparse & Redundant Representations Part III – Dictionary Learning         & The K-SVD Algorithm  Part IV 	– Back to Denoising … and Beyond – handling stills and video denoising & inpainting, demosaicing, super-res., and compression Part V	–Summary & Conclusions ,[object Object]
When used in image processing, they lead to state-of-the-art results. Today we will show that  Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
4 Part IDenoising by                              Sparse & Redundant                 Representations Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
5 ? Remove Additive Noise    Noise Removal? Our story begins with image denoising … ,[object Object]
Many Considered Directions: Partial differential equations, Statistical estimators, Adaptive filters, Inverse problems & regularization,          Wavelets, Example-based techniques, Sparse representations, …Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
6  Relation to measurements Prior or regularization ThomasBayes                                    1702 - 1761    Denoising By Energy Minimization  Many of the proposed image denoising algorithms are related to the minimization of an energy function of the form y : Given measurements   x : Unknown to be recovered ,[object Object]
Clearly, the wisdom in such an approach is within the choice of the prior – modeling the images of interest. Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
7 Energy Smoothness Adapt+ Smooth Robust Statistics ,[object Object]
 Compression algorithms as priors,
 …Total-Variation Wavelet Sparsity Sparse & Redundant    The Evolution of G(x) During the past several decades we have made all sort of guesses about the prior G(x) for images:    Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
8 ,[object Object]
The vector is generated randomly with few (say L) non-zeros at random locations and with random values. N N A sparse & random vector K A fixed Dictionary    Sparse Modeling of Signals  M ,[object Object],Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
9 M Multiply by D Sparseland  Signals are Special Interesting Model: ,[object Object]
Rich:A general model: the obtained signals are a union     of many low-dimensional Gaussians.
Familiar: We have been  using this model in other context for a while now (wavelet, JPEG, …).Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
10 As p  0 we  get a count         of the non-zeros in the vector 1 -1 +1    Sparse & Redundant Rep. Modeling? Our signal  model is thus:  Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
11 D-y=            -    Back to Our MAP Energy Function  ,[object Object]
The vector  is the                                                            representation (sparse/redundant)                                                     of the desired                                                                                               signal x.
The core idea: while few (L out of K) atoms can be merged        to form the true signal, the noise cannot be fitted well. Thus, we obtain an effective projection of the noise onto a very         low-dimensional space, thus getting denoising effect. Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
12 Wait! There are Some Issues  ,[object Object],					      or                                                                            or                                    ? ,[object Object]
Practical Problems: What dictionary D should we use, such that all this leads to effective denoising? Will all this work in applications?Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
Image Denoising & Beyond Via Learned  Dictionaries and Sparse representations By: Michael Elad 13 To Summarize So Far … Image denoising (and many other problems in image processing) requires a model for the desired image We proposed a model for signals/images based on sparse and redundant representations What do    we do?   There are some issues:  Theoretical How to approximate? What about D? Great!      No?
14 Part IITheoretical &                   Numerical Foundations  Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
Sparse and Redundant                     Signal Representation,  and Its Role in  Image Processing 15    Lets Start with the Noiseless Problem Suppose we build a signal by the relation We aim to find the signal’s representation:  Known  Why should we necessarily get            ? It might happen that eventually                    . Uniqueness
16 * Definition:Given a matrix D, =Spark{D} is the smallestnumber of columns that are linearly dependent. Donoho & E. (‘02)  Example: Spark = 3 *	In tensor decomposition, Kruskal defined something similar already in 1989.    Matrix “Spark” Rank  = 4 Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
17 Suppose this problem has been solved somehow Uniqueness If we found a representation that satisfy  Then necessarily it is unique (the sparsest). Donoho & E. (‘02)  M This result implies that if       generates signals using “sparse enough” , the solution of the above will find it exactly.    Uniqueness Rule Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
18 This is a combinatorial problem, proven to be NP-Hard!  Solve the LS problem  for each support                                   There are (K) such supports L    Our Goal   Here is a recipe for solving this problem: Gather all the supports {Si}i        of cardinality L    LS error ≤ ε2? Set L=1  Yes No Set L=L+1  Assume: K=1000, L=10 (known!), 1 nano-sec per each LS         We shall need ~8e+6 years to solve this problem !!!!! Done Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
19    Lets Approximate    Greedy methods Build the solution one non-zero element at a time Relaxation methods Smooth the L0 and use continuous optimization techniques Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
20    Relaxation – The Basis Pursuit (BP) Instead of solving Solve Instead ,[object Object]
The newly defined problem is convex (quad. programming).
Very efficient solvers can be deployed:
Interior point methods [Chen, Donoho, & Saunders (‘95)] [Kim, Koh, Lustig, Boyd, & D. Gorinevsky (`07)].
Sequential shrinkage for union of ortho-bases [Bruce et.al. (‘98)].
Iterative shrinkage [Figuerido & Nowak (‘03)] [Daubechies, Defrise, & De-Mole (‘04)]                     [E. (‘05)] [E., Matalon, & Zibulevsky (‘06)] [Beck & Teboulle (`09)] … Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
21    Go Greedy: Matching Pursuit (MP) ,[object Object]
Step 1: find the one atom that  best matches the signal.
Next steps: given the previously found atoms, find the next one to best fit the rsidual.
The algorithm stops when the error            is below the destination threshold.
The Orthogonal MP (OMP) is an improved version that re-evaluates the coefficients by Least-Squares after each round.Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
22 Pursuit Algorithms ? Why should they work There are various algorithms designed for approximating the solution of this problem:  ,[object Object]
Relaxation Algorithms: Basis Pursuit (a.k.a. LASSO), Dnatzig Selector & numerical ways to handle them [1995-today].
Hybrid Algorithms: StOMP, CoSaMP, Subspace Pursuit, Iterative Hard-Thresholding [2007-today].
…Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
23 D = DT DTD ,[object Object],The Mutual Coherence ,[object Object],Assume normalized columns ,[object Object],Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
24    BP and MP Equivalence (No Noise) Equivalence Given a signal x with a representation            , assuming that                         , BP and MP    are guaranteed to find the sparsest solution.  Donoho & E. (‘02)  Gribonval & Nielsen (‘03) Tropp (‘03)  Temlyakov (‘03) ,[object Object]
The above result corresponds to the worst-case, and as such, it is too pessimistic.
Average performance results are available too, showing much better bounds [Donoho (`04)] [Candes et.al. (‘04)] [Tanner et.al. (‘05)]             [E. (‘06)] [Tropp et.al. (‘06)] … [Candes et. al. (‘09)]. Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
25    BP Stability for the Noisy Case  Given a signal                with a representation satisfying                    and a white Gaussian  noise                   , BP will show  stability, i.e.,  Stability * Ben-Haim, Eldar & E. (‘09) * With very high    probability (as     K goes to ∞) ,[object Object]
This result is the oracle’s error, multuiplied by C·logK.
Similar results exist for other pursuit algorithms (Dantzig Selector, Orthogonal Matching Pursuit, CoSaMP, Subspace Pursuit, …)Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
Image Denoising & Beyond Via Learned  Dictionaries and Sparse representations By: Michael Elad 26 To Summarize So Far … Image denoising (and many other problems in image processing) requires a model for the desired image We proposed a model for signals/images based on sparse and redundant representations Problems? What do    we do?   We have seen that there are approximation methods to find the sparsest solution, and there are theoretical results that guarantee their success. The Dictionary D should be found somehow !!! What           next?
27 Part IIIDictionary Learning:                         The K-SVD Algorithm Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
28 D should be chosen such that it sparsifies the representations The approach we will take for building D is training it,   based on Learning from          Image Examples One approach to choose D is from a known set of transforms (Steerable wavelet, Curvelet, Contourlets, Bandlets, Shearlets…)    What Should D Be?  Our Assumption: Good-behaved Images                                      have a sparse representation Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
29 D X  A Each example has a sparse representation with no more than L atoms Each example is                    a linear combination                   of atoms from D Measure of Quality for D [Field & Olshausen (‘96)] [Engan et. al. (‘99)] [Lewicki & Sejnowski (‘00)] [Cotter et. al. (‘03)] [Gribonval et. al. (‘04)] [Aharon, E. & Bruckstein (‘04)] [Aharon, E. & Bruckstein (‘05)] Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
30 D Initialize         D Sparse Coding Nearest Neighbor XT Dictionary Update Column-by-Column by  Mean computation over the relevant examples    K–Means For Clustering  Clustering: An extreme sparse representation   Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
31 D Initialize         D Sparse Coding Use Matching Pursuit XT Dictionary Update Column-by-Column by  SVD computation over the relevant examples    The K–SVD Algorithm – General  [Aharon, E. & Bruckstein (‘04,‘05)] Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
32 D D is known!  For the jth item           we solve  XT    K–SVD: Sparse Coding Stage Solved by                            A Pursuit Algorithm Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
33 Fixing all A and D apart from the kth column, and seek both dk and the kth column in A to better fit the residual! We should solve:    K–SVD: Dictionary Update Stage We refer only to the examples that use the column dk D SVD Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
Image Denoising & Beyond Via Learned  Dictionaries and Sparse representations By: Michael Elad 34 To Summarize So Far … Image denoising (and many other problems in image processing) requires a model for the desired image We proposed a model for signals/images based on sparse and redundant representations Problems? What do    we do?   We have seen approximation methods that find the sparsest solution, and theoretical results that guarantee their success. We also saw a way to learn D Will it all work in applications?  What           next?
35 Part IVBack to Denoising …                 and Beyond –                     Combining it All Sparse and Redundant Representation Modeling of Signals – Theory and Applications  By: Michael Elad
36 k D N Extracts a patch in the ij location Our prior    From Local to Global Treatment ,[object Object]
So, how should large images be handled?

Mais conteúdo relacionado

Mais procurados

Yurii Pashchenko: Zero-shot learning capabilities of CLIP model from OpenAI
Yurii Pashchenko: Zero-shot learning capabilities of CLIP model from OpenAIYurii Pashchenko: Zero-shot learning capabilities of CLIP model from OpenAI
Yurii Pashchenko: Zero-shot learning capabilities of CLIP model from OpenAILviv Startup Club
 
Estimating Human Pose from Occluded Images (ACCV 2009)
Estimating Human Pose from Occluded Images (ACCV 2009)Estimating Human Pose from Occluded Images (ACCV 2009)
Estimating Human Pose from Occluded Images (ACCV 2009)Jia-Bin Huang
 
Deep Learning - A Literature survey
Deep Learning - A Literature surveyDeep Learning - A Literature survey
Deep Learning - A Literature surveyAkshay Hegde
 
Moshe Guttmann's slides on eigenface
Moshe Guttmann's slides on eigenfaceMoshe Guttmann's slides on eigenface
Moshe Guttmann's slides on eigenfacewolf
 
Deep Learning Using TensorFlow | TensorFlow Tutorial | AI & Deep Learning Tra...
Deep Learning Using TensorFlow | TensorFlow Tutorial | AI & Deep Learning Tra...Deep Learning Using TensorFlow | TensorFlow Tutorial | AI & Deep Learning Tra...
Deep Learning Using TensorFlow | TensorFlow Tutorial | AI & Deep Learning Tra...Edureka!
 
Extreme learning machine:Theory and applications
Extreme learning machine:Theory and applicationsExtreme learning machine:Theory and applications
Extreme learning machine:Theory and applicationsJames Chou
 
Donner - Deep Learning - Overview and practical aspects
Donner - Deep Learning - Overview and practical aspectsDonner - Deep Learning - Overview and practical aspects
Donner - Deep Learning - Overview and practical aspectsVienna Data Science Group
 
Deep Learning Class #0 - You Can Do It
Deep Learning Class #0 - You Can Do ItDeep Learning Class #0 - You Can Do It
Deep Learning Class #0 - You Can Do ItHolberton School
 
Deep Neural Networks 
that talk (Back)… with style
Deep Neural Networks 
that talk (Back)… with styleDeep Neural Networks 
that talk (Back)… with style
Deep Neural Networks 
that talk (Back)… with styleRoelof Pieters
 
Deep Learning for Information Retrieval: Models, Progress, & Opportunities
Deep Learning for Information Retrieval: Models, Progress, & OpportunitiesDeep Learning for Information Retrieval: Models, Progress, & Opportunities
Deep Learning for Information Retrieval: Models, Progress, & OpportunitiesMatthew Lease
 
Practical and Robust Stenciled Shadow Volumes for Hardware-Accelerated Rendering
Practical and Robust Stenciled Shadow Volumes for Hardware-Accelerated RenderingPractical and Robust Stenciled Shadow Volumes for Hardware-Accelerated Rendering
Practical and Robust Stenciled Shadow Volumes for Hardware-Accelerated RenderingMark Kilgard
 
NNFL - Guru Nanak Dev Engineering College
NNFL - Guru Nanak Dev Engineering CollegeNNFL - Guru Nanak Dev Engineering College
NNFL - Guru Nanak Dev Engineering CollegeMR. VIKRAM SNEHI
 
Introduction to soft computing
Introduction to soft computingIntroduction to soft computing
Introduction to soft computingAnkush Kumar
 

Mais procurados (20)

Yurii Pashchenko: Zero-shot learning capabilities of CLIP model from OpenAI
Yurii Pashchenko: Zero-shot learning capabilities of CLIP model from OpenAIYurii Pashchenko: Zero-shot learning capabilities of CLIP model from OpenAI
Yurii Pashchenko: Zero-shot learning capabilities of CLIP model from OpenAI
 
Review_Cibe Sridharan
Review_Cibe SridharanReview_Cibe Sridharan
Review_Cibe Sridharan
 
PIES_Profile_INDIA
PIES_Profile_INDIAPIES_Profile_INDIA
PIES_Profile_INDIA
 
Estimating Human Pose from Occluded Images (ACCV 2009)
Estimating Human Pose from Occluded Images (ACCV 2009)Estimating Human Pose from Occluded Images (ACCV 2009)
Estimating Human Pose from Occluded Images (ACCV 2009)
 
Deep Learning - A Literature survey
Deep Learning - A Literature surveyDeep Learning - A Literature survey
Deep Learning - A Literature survey
 
Moshe Guttmann's slides on eigenface
Moshe Guttmann's slides on eigenfaceMoshe Guttmann's slides on eigenface
Moshe Guttmann's slides on eigenface
 
Deep Learning Using TensorFlow | TensorFlow Tutorial | AI & Deep Learning Tra...
Deep Learning Using TensorFlow | TensorFlow Tutorial | AI & Deep Learning Tra...Deep Learning Using TensorFlow | TensorFlow Tutorial | AI & Deep Learning Tra...
Deep Learning Using TensorFlow | TensorFlow Tutorial | AI & Deep Learning Tra...
 
Recognition
RecognitionRecognition
Recognition
 
Extreme learning machine:Theory and applications
Extreme learning machine:Theory and applicationsExtreme learning machine:Theory and applications
Extreme learning machine:Theory and applications
 
Donner - Deep Learning - Overview and practical aspects
Donner - Deep Learning - Overview and practical aspectsDonner - Deep Learning - Overview and practical aspects
Donner - Deep Learning - Overview and practical aspects
 
Lecture17
Lecture17Lecture17
Lecture17
 
Deep Learning Class #0 - You Can Do It
Deep Learning Class #0 - You Can Do ItDeep Learning Class #0 - You Can Do It
Deep Learning Class #0 - You Can Do It
 
Deep Neural Networks 
that talk (Back)… with style
Deep Neural Networks 
that talk (Back)… with styleDeep Neural Networks 
that talk (Back)… with style
Deep Neural Networks 
that talk (Back)… with style
 
Deep Learning for Information Retrieval: Models, Progress, & Opportunities
Deep Learning for Information Retrieval: Models, Progress, & OpportunitiesDeep Learning for Information Retrieval: Models, Progress, & Opportunities
Deep Learning for Information Retrieval: Models, Progress, & Opportunities
 
Practical and Robust Stenciled Shadow Volumes for Hardware-Accelerated Rendering
Practical and Robust Stenciled Shadow Volumes for Hardware-Accelerated RenderingPractical and Robust Stenciled Shadow Volumes for Hardware-Accelerated Rendering
Practical and Robust Stenciled Shadow Volumes for Hardware-Accelerated Rendering
 
CS 354 Shadows
CS 354 ShadowsCS 354 Shadows
CS 354 Shadows
 
Lecture1 - Machine Learning
Lecture1 - Machine LearningLecture1 - Machine Learning
Lecture1 - Machine Learning
 
Neural Architectures for Video Encoding
Neural Architectures for Video EncodingNeural Architectures for Video Encoding
Neural Architectures for Video Encoding
 
NNFL - Guru Nanak Dev Engineering College
NNFL - Guru Nanak Dev Engineering CollegeNNFL - Guru Nanak Dev Engineering College
NNFL - Guru Nanak Dev Engineering College
 
Introduction to soft computing
Introduction to soft computingIntroduction to soft computing
Introduction to soft computing
 

Destaque

stduy Edge-Based Image Coarsening
stduy Edge-Based Image Coarseningstduy Edge-Based Image Coarsening
stduy Edge-Based Image CoarseningChiamin Hsu
 
Tuto part2
Tuto part2Tuto part2
Tuto part2Bo Li
 
Learning Sparse Representation
Learning Sparse RepresentationLearning Sparse Representation
Learning Sparse RepresentationGabriel Peyré
 
Ppt compressed sensing a tutorial
Ppt compressed sensing a tutorialPpt compressed sensing a tutorial
Ppt compressed sensing a tutorialTerence Gao
 
Super Resolution in Digital Image processing
Super Resolution in Digital Image processingSuper Resolution in Digital Image processing
Super Resolution in Digital Image processingRamrao Desai
 

Destaque (7)

stduy Edge-Based Image Coarsening
stduy Edge-Based Image Coarseningstduy Edge-Based Image Coarsening
stduy Edge-Based Image Coarsening
 
Tuto part2
Tuto part2Tuto part2
Tuto part2
 
Learning Sparse Representation
Learning Sparse RepresentationLearning Sparse Representation
Learning Sparse Representation
 
Single-image Super-resolution
Single-image Super-resolutionSingle-image Super-resolution
Single-image Super-resolution
 
Ppt compressed sensing a tutorial
Ppt compressed sensing a tutorialPpt compressed sensing a tutorial
Ppt compressed sensing a tutorial
 
Super resolution from a single image
Super resolution from a single imageSuper resolution from a single image
Super resolution from a single image
 
Super Resolution in Digital Image processing
Super Resolution in Digital Image processingSuper Resolution in Digital Image processing
Super Resolution in Digital Image processing
 

Semelhante a C&s sparse june_2010

IT6005 digital image processing question bank
IT6005   digital image processing question bankIT6005   digital image processing question bank
IT6005 digital image processing question bankGayathri Krishnamoorthy
 
Automatic Differentiation and SciML in Reality: What can go wrong, and what t...
Automatic Differentiation and SciML in Reality: What can go wrong, and what t...Automatic Differentiation and SciML in Reality: What can go wrong, and what t...
Automatic Differentiation and SciML in Reality: What can go wrong, and what t...Chris Rackauckas
 
Manifold learning with application to object recognition
Manifold learning with application to object recognitionManifold learning with application to object recognition
Manifold learning with application to object recognitionzukun
 
Mathematical Modeling for Practical Problems
Mathematical Modeling for Practical ProblemsMathematical Modeling for Practical Problems
Mathematical Modeling for Practical ProblemsLiwei Ren任力偉
 
Performance of Matching Algorithmsfor Signal Approximation
Performance of Matching Algorithmsfor Signal ApproximationPerformance of Matching Algorithmsfor Signal Approximation
Performance of Matching Algorithmsfor Signal Approximationiosrjce
 
Point Cloud Processing: Estimating Normal Vectors and Curvature Indicators us...
Point Cloud Processing: Estimating Normal Vectors and Curvature Indicators us...Point Cloud Processing: Estimating Normal Vectors and Curvature Indicators us...
Point Cloud Processing: Estimating Normal Vectors and Curvature Indicators us...Pirouz Nourian
 
Paper id 24201464
Paper id 24201464Paper id 24201464
Paper id 24201464IJRAT
 
17- Kernels and Clustering.pptx
17- Kernels and Clustering.pptx17- Kernels and Clustering.pptx
17- Kernels and Clustering.pptxssuser2023c6
 
Linear models for data science
Linear models for data scienceLinear models for data science
Linear models for data scienceBrad Klingenberg
 
Abstract of project 2
Abstract of project 2Abstract of project 2
Abstract of project 2Vikram Mandal
 
Latent Dirichlet Allocation
Latent Dirichlet AllocationLatent Dirichlet Allocation
Latent Dirichlet AllocationMarco Righini
 
人工智慧與物聯網的創新與服務模式
人工智慧與物聯網的創新與服務模式人工智慧與物聯網的創新與服務模式
人工智慧與物聯網的創新與服務模式Craig Chao
 
A Compressed Sensing Approach to Image Reconstruction
A Compressed Sensing Approach to Image ReconstructionA Compressed Sensing Approach to Image Reconstruction
A Compressed Sensing Approach to Image Reconstructionijsrd.com
 
Generalizing Scientific Machine Learning and Differentiable Simulation Beyond...
Generalizing Scientific Machine Learning and Differentiable Simulation Beyond...Generalizing Scientific Machine Learning and Differentiable Simulation Beyond...
Generalizing Scientific Machine Learning and Differentiable Simulation Beyond...Chris Rackauckas
 

Semelhante a C&s sparse june_2010 (20)

IT6005 digital image processing question bank
IT6005   digital image processing question bankIT6005   digital image processing question bank
IT6005 digital image processing question bank
 
Automatic Differentiation and SciML in Reality: What can go wrong, and what t...
Automatic Differentiation and SciML in Reality: What can go wrong, and what t...Automatic Differentiation and SciML in Reality: What can go wrong, and what t...
Automatic Differentiation and SciML in Reality: What can go wrong, and what t...
 
Manifold learning with application to object recognition
Manifold learning with application to object recognitionManifold learning with application to object recognition
Manifold learning with application to object recognition
 
Mathematical Modeling for Practical Problems
Mathematical Modeling for Practical ProblemsMathematical Modeling for Practical Problems
Mathematical Modeling for Practical Problems
 
Performance of Matching Algorithmsfor Signal Approximation
Performance of Matching Algorithmsfor Signal ApproximationPerformance of Matching Algorithmsfor Signal Approximation
Performance of Matching Algorithmsfor Signal Approximation
 
Point Cloud Processing: Estimating Normal Vectors and Curvature Indicators us...
Point Cloud Processing: Estimating Normal Vectors and Curvature Indicators us...Point Cloud Processing: Estimating Normal Vectors and Curvature Indicators us...
Point Cloud Processing: Estimating Normal Vectors and Curvature Indicators us...
 
Presentation v3.2
Presentation v3.2Presentation v3.2
Presentation v3.2
 
Presentation v3.2
Presentation v3.2Presentation v3.2
Presentation v3.2
 
Paper id 24201464
Paper id 24201464Paper id 24201464
Paper id 24201464
 
17- Kernels and Clustering.pptx
17- Kernels and Clustering.pptx17- Kernels and Clustering.pptx
17- Kernels and Clustering.pptx
 
Compressed Sensing - Achuta Kadambi
Compressed Sensing - Achuta KadambiCompressed Sensing - Achuta Kadambi
Compressed Sensing - Achuta Kadambi
 
L010628894
L010628894L010628894
L010628894
 
Linear models for data science
Linear models for data scienceLinear models for data science
Linear models for data science
 
Java and Deep Learning
Java and Deep LearningJava and Deep Learning
Java and Deep Learning
 
Abstract of project 2
Abstract of project 2Abstract of project 2
Abstract of project 2
 
Latent Dirichlet Allocation
Latent Dirichlet AllocationLatent Dirichlet Allocation
Latent Dirichlet Allocation
 
人工智慧與物聯網的創新與服務模式
人工智慧與物聯網的創新與服務模式人工智慧與物聯網的創新與服務模式
人工智慧與物聯網的創新與服務模式
 
A Compressed Sensing Approach to Image Reconstruction
A Compressed Sensing Approach to Image ReconstructionA Compressed Sensing Approach to Image Reconstruction
A Compressed Sensing Approach to Image Reconstruction
 
Generalizing Scientific Machine Learning and Differentiable Simulation Beyond...
Generalizing Scientific Machine Learning and Differentiable Simulation Beyond...Generalizing Scientific Machine Learning and Differentiable Simulation Beyond...
Generalizing Scientific Machine Learning and Differentiable Simulation Beyond...
 
DIP-Questions.pdf
DIP-Questions.pdfDIP-Questions.pdf
DIP-Questions.pdf
 

Último

Sports & Fitness Value Added Course FY..
Sports & Fitness Value Added Course FY..Sports & Fitness Value Added Course FY..
Sports & Fitness Value Added Course FY..Disha Kariya
 
BAG TECHNIQUE Bag technique-a tool making use of public health bag through wh...
BAG TECHNIQUE Bag technique-a tool making use of public health bag through wh...BAG TECHNIQUE Bag technique-a tool making use of public health bag through wh...
BAG TECHNIQUE Bag technique-a tool making use of public health bag through wh...Sapna Thakur
 
Web & Social Media Analytics Previous Year Question Paper.pdf
Web & Social Media Analytics Previous Year Question Paper.pdfWeb & Social Media Analytics Previous Year Question Paper.pdf
Web & Social Media Analytics Previous Year Question Paper.pdfJayanti Pande
 
Measures of Central Tendency: Mean, Median and Mode
Measures of Central Tendency: Mean, Median and ModeMeasures of Central Tendency: Mean, Median and Mode
Measures of Central Tendency: Mean, Median and ModeThiyagu K
 
BASLIQ CURRENT LOOKBOOK LOOKBOOK(1) (1).pdf
BASLIQ CURRENT LOOKBOOK  LOOKBOOK(1) (1).pdfBASLIQ CURRENT LOOKBOOK  LOOKBOOK(1) (1).pdf
BASLIQ CURRENT LOOKBOOK LOOKBOOK(1) (1).pdfSoniaTolstoy
 
Arihant handbook biology for class 11 .pdf
Arihant handbook biology for class 11 .pdfArihant handbook biology for class 11 .pdf
Arihant handbook biology for class 11 .pdfchloefrazer622
 
SOCIAL AND HISTORICAL CONTEXT - LFTVD.pptx
SOCIAL AND HISTORICAL CONTEXT - LFTVD.pptxSOCIAL AND HISTORICAL CONTEXT - LFTVD.pptx
SOCIAL AND HISTORICAL CONTEXT - LFTVD.pptxiammrhaywood
 
Student login on Anyboli platform.helpin
Student login on Anyboli platform.helpinStudent login on Anyboli platform.helpin
Student login on Anyboli platform.helpinRaunakKeshri1
 
Accessible design: Minimum effort, maximum impact
Accessible design: Minimum effort, maximum impactAccessible design: Minimum effort, maximum impact
Accessible design: Minimum effort, maximum impactdawncurless
 
Unit-IV- Pharma. Marketing Channels.pptx
Unit-IV- Pharma. Marketing Channels.pptxUnit-IV- Pharma. Marketing Channels.pptx
Unit-IV- Pharma. Marketing Channels.pptxVishalSingh1417
 
1029-Danh muc Sach Giao Khoa khoi 6.pdf
1029-Danh muc Sach Giao Khoa khoi  6.pdf1029-Danh muc Sach Giao Khoa khoi  6.pdf
1029-Danh muc Sach Giao Khoa khoi 6.pdfQucHHunhnh
 
microwave assisted reaction. General introduction
microwave assisted reaction. General introductionmicrowave assisted reaction. General introduction
microwave assisted reaction. General introductionMaksud Ahmed
 
Class 11th Physics NEET formula sheet pdf
Class 11th Physics NEET formula sheet pdfClass 11th Physics NEET formula sheet pdf
Class 11th Physics NEET formula sheet pdfAyushMahapatra5
 
Software Engineering Methodologies (overview)
Software Engineering Methodologies (overview)Software Engineering Methodologies (overview)
Software Engineering Methodologies (overview)eniolaolutunde
 
Interactive Powerpoint_How to Master effective communication
Interactive Powerpoint_How to Master effective communicationInteractive Powerpoint_How to Master effective communication
Interactive Powerpoint_How to Master effective communicationnomboosow
 
Disha NEET Physics Guide for classes 11 and 12.pdf
Disha NEET Physics Guide for classes 11 and 12.pdfDisha NEET Physics Guide for classes 11 and 12.pdf
Disha NEET Physics Guide for classes 11 and 12.pdfchloefrazer622
 
social pharmacy d-pharm 1st year by Pragati K. Mahajan
social pharmacy d-pharm 1st year by Pragati K. Mahajansocial pharmacy d-pharm 1st year by Pragati K. Mahajan
social pharmacy d-pharm 1st year by Pragati K. Mahajanpragatimahajan3
 

Último (20)

Sports & Fitness Value Added Course FY..
Sports & Fitness Value Added Course FY..Sports & Fitness Value Added Course FY..
Sports & Fitness Value Added Course FY..
 
BAG TECHNIQUE Bag technique-a tool making use of public health bag through wh...
BAG TECHNIQUE Bag technique-a tool making use of public health bag through wh...BAG TECHNIQUE Bag technique-a tool making use of public health bag through wh...
BAG TECHNIQUE Bag technique-a tool making use of public health bag through wh...
 
Web & Social Media Analytics Previous Year Question Paper.pdf
Web & Social Media Analytics Previous Year Question Paper.pdfWeb & Social Media Analytics Previous Year Question Paper.pdf
Web & Social Media Analytics Previous Year Question Paper.pdf
 
Measures of Central Tendency: Mean, Median and Mode
Measures of Central Tendency: Mean, Median and ModeMeasures of Central Tendency: Mean, Median and Mode
Measures of Central Tendency: Mean, Median and Mode
 
BASLIQ CURRENT LOOKBOOK LOOKBOOK(1) (1).pdf
BASLIQ CURRENT LOOKBOOK  LOOKBOOK(1) (1).pdfBASLIQ CURRENT LOOKBOOK  LOOKBOOK(1) (1).pdf
BASLIQ CURRENT LOOKBOOK LOOKBOOK(1) (1).pdf
 
Arihant handbook biology for class 11 .pdf
Arihant handbook biology for class 11 .pdfArihant handbook biology for class 11 .pdf
Arihant handbook biology for class 11 .pdf
 
SOCIAL AND HISTORICAL CONTEXT - LFTVD.pptx
SOCIAL AND HISTORICAL CONTEXT - LFTVD.pptxSOCIAL AND HISTORICAL CONTEXT - LFTVD.pptx
SOCIAL AND HISTORICAL CONTEXT - LFTVD.pptx
 
Student login on Anyboli platform.helpin
Student login on Anyboli platform.helpinStudent login on Anyboli platform.helpin
Student login on Anyboli platform.helpin
 
Accessible design: Minimum effort, maximum impact
Accessible design: Minimum effort, maximum impactAccessible design: Minimum effort, maximum impact
Accessible design: Minimum effort, maximum impact
 
Unit-IV- Pharma. Marketing Channels.pptx
Unit-IV- Pharma. Marketing Channels.pptxUnit-IV- Pharma. Marketing Channels.pptx
Unit-IV- Pharma. Marketing Channels.pptx
 
1029-Danh muc Sach Giao Khoa khoi 6.pdf
1029-Danh muc Sach Giao Khoa khoi  6.pdf1029-Danh muc Sach Giao Khoa khoi  6.pdf
1029-Danh muc Sach Giao Khoa khoi 6.pdf
 
microwave assisted reaction. General introduction
microwave assisted reaction. General introductionmicrowave assisted reaction. General introduction
microwave assisted reaction. General introduction
 
Class 11th Physics NEET formula sheet pdf
Class 11th Physics NEET formula sheet pdfClass 11th Physics NEET formula sheet pdf
Class 11th Physics NEET formula sheet pdf
 
Advance Mobile Application Development class 07
Advance Mobile Application Development class 07Advance Mobile Application Development class 07
Advance Mobile Application Development class 07
 
Mattingly "AI & Prompt Design: The Basics of Prompt Design"
Mattingly "AI & Prompt Design: The Basics of Prompt Design"Mattingly "AI & Prompt Design: The Basics of Prompt Design"
Mattingly "AI & Prompt Design: The Basics of Prompt Design"
 
Software Engineering Methodologies (overview)
Software Engineering Methodologies (overview)Software Engineering Methodologies (overview)
Software Engineering Methodologies (overview)
 
Interactive Powerpoint_How to Master effective communication
Interactive Powerpoint_How to Master effective communicationInteractive Powerpoint_How to Master effective communication
Interactive Powerpoint_How to Master effective communication
 
Código Creativo y Arte de Software | Unidad 1
Código Creativo y Arte de Software | Unidad 1Código Creativo y Arte de Software | Unidad 1
Código Creativo y Arte de Software | Unidad 1
 
Disha NEET Physics Guide for classes 11 and 12.pdf
Disha NEET Physics Guide for classes 11 and 12.pdfDisha NEET Physics Guide for classes 11 and 12.pdf
Disha NEET Physics Guide for classes 11 and 12.pdf
 
social pharmacy d-pharm 1st year by Pragati K. Mahajan
social pharmacy d-pharm 1st year by Pragati K. Mahajansocial pharmacy d-pharm 1st year by Pragati K. Mahajan
social pharmacy d-pharm 1st year by Pragati K. Mahajan
 

C&s sparse june_2010

  • 1. Sparse & Redundant Representation Modeling of Images: Theory and Applications Michael Elad The Computer Science Department The Technion – Israel Institute of technology Haifa 32000, Israel Seventh International Conference on Curves and Surfaces Avignon - FRANCE June 24-30, 2010 This research was supported by the European Community's FP7-FET program SMALL under grant agreement no. 225913
  • 2. 2 This Talk Gives and Overview On … A decade of tremendous progress in the field of Sparse and Redundant Representations Numerical Problems Theory Applications Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 3.
  • 4. When used in image processing, they lead to state-of-the-art results. Today we will show that Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 5. 4 Part IDenoising by Sparse & Redundant Representations Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 6.
  • 7. Many Considered Directions: Partial differential equations, Statistical estimators, Adaptive filters, Inverse problems & regularization, Wavelets, Example-based techniques, Sparse representations, …Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 8.
  • 9. Clearly, the wisdom in such an approach is within the choice of the prior – modeling the images of interest. Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 10.
  • 12. …Total-Variation Wavelet Sparsity Sparse & Redundant The Evolution of G(x) During the past several decades we have made all sort of guesses about the prior G(x) for images: Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 13.
  • 14.
  • 15.
  • 16. Rich:A general model: the obtained signals are a union of many low-dimensional Gaussians.
  • 17. Familiar: We have been using this model in other context for a while now (wavelet, JPEG, …).Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 18. 10 As p  0 we get a count of the non-zeros in the vector 1 -1 +1 Sparse & Redundant Rep. Modeling? Our signal model is thus: Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 19.
  • 20. The vector  is the representation (sparse/redundant) of the desired signal x.
  • 21. The core idea: while few (L out of K) atoms can be merged to form the true signal, the noise cannot be fitted well. Thus, we obtain an effective projection of the noise onto a very low-dimensional space, thus getting denoising effect. Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 22.
  • 23. Practical Problems: What dictionary D should we use, such that all this leads to effective denoising? Will all this work in applications?Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 24. Image Denoising & Beyond Via Learned Dictionaries and Sparse representations By: Michael Elad 13 To Summarize So Far … Image denoising (and many other problems in image processing) requires a model for the desired image We proposed a model for signals/images based on sparse and redundant representations What do we do? There are some issues: Theoretical How to approximate? What about D? Great! No?
  • 25. 14 Part IITheoretical & Numerical Foundations Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 26. Sparse and Redundant Signal Representation, and Its Role in Image Processing 15 Lets Start with the Noiseless Problem Suppose we build a signal by the relation We aim to find the signal’s representation: Known Why should we necessarily get ? It might happen that eventually . Uniqueness
  • 27. 16 * Definition:Given a matrix D, =Spark{D} is the smallestnumber of columns that are linearly dependent. Donoho & E. (‘02) Example: Spark = 3 * In tensor decomposition, Kruskal defined something similar already in 1989. Matrix “Spark” Rank = 4 Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 28. 17 Suppose this problem has been solved somehow Uniqueness If we found a representation that satisfy Then necessarily it is unique (the sparsest). Donoho & E. (‘02) M This result implies that if generates signals using “sparse enough” , the solution of the above will find it exactly. Uniqueness Rule Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 29. 18 This is a combinatorial problem, proven to be NP-Hard! Solve the LS problem for each support There are (K) such supports L Our Goal Here is a recipe for solving this problem: Gather all the supports {Si}i of cardinality L LS error ≤ ε2? Set L=1 Yes No Set L=L+1 Assume: K=1000, L=10 (known!), 1 nano-sec per each LS We shall need ~8e+6 years to solve this problem !!!!! Done Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 30. 19 Lets Approximate Greedy methods Build the solution one non-zero element at a time Relaxation methods Smooth the L0 and use continuous optimization techniques Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 31.
  • 32. The newly defined problem is convex (quad. programming).
  • 33. Very efficient solvers can be deployed:
  • 34. Interior point methods [Chen, Donoho, & Saunders (‘95)] [Kim, Koh, Lustig, Boyd, & D. Gorinevsky (`07)].
  • 35. Sequential shrinkage for union of ortho-bases [Bruce et.al. (‘98)].
  • 36. Iterative shrinkage [Figuerido & Nowak (‘03)] [Daubechies, Defrise, & De-Mole (‘04)] [E. (‘05)] [E., Matalon, & Zibulevsky (‘06)] [Beck & Teboulle (`09)] … Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 37.
  • 38. Step 1: find the one atom that best matches the signal.
  • 39. Next steps: given the previously found atoms, find the next one to best fit the rsidual.
  • 40. The algorithm stops when the error is below the destination threshold.
  • 41. The Orthogonal MP (OMP) is an improved version that re-evaluates the coefficients by Least-Squares after each round.Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 42.
  • 43. Relaxation Algorithms: Basis Pursuit (a.k.a. LASSO), Dnatzig Selector & numerical ways to handle them [1995-today].
  • 44. Hybrid Algorithms: StOMP, CoSaMP, Subspace Pursuit, Iterative Hard-Thresholding [2007-today].
  • 45. …Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 46.
  • 47.
  • 48. The above result corresponds to the worst-case, and as such, it is too pessimistic.
  • 49. Average performance results are available too, showing much better bounds [Donoho (`04)] [Candes et.al. (‘04)] [Tanner et.al. (‘05)] [E. (‘06)] [Tropp et.al. (‘06)] … [Candes et. al. (‘09)]. Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 50.
  • 51. This result is the oracle’s error, multuiplied by C·logK.
  • 52. Similar results exist for other pursuit algorithms (Dantzig Selector, Orthogonal Matching Pursuit, CoSaMP, Subspace Pursuit, …)Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 53. Image Denoising & Beyond Via Learned Dictionaries and Sparse representations By: Michael Elad 26 To Summarize So Far … Image denoising (and many other problems in image processing) requires a model for the desired image We proposed a model for signals/images based on sparse and redundant representations Problems? What do we do? We have seen that there are approximation methods to find the sparsest solution, and there are theoretical results that guarantee their success. The Dictionary D should be found somehow !!! What next?
  • 54. 27 Part IIIDictionary Learning: The K-SVD Algorithm Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 55. 28 D should be chosen such that it sparsifies the representations The approach we will take for building D is training it, based on Learning from Image Examples One approach to choose D is from a known set of transforms (Steerable wavelet, Curvelet, Contourlets, Bandlets, Shearlets…) What Should D Be? Our Assumption: Good-behaved Images have a sparse representation Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 56. 29 D X  A Each example has a sparse representation with no more than L atoms Each example is a linear combination of atoms from D Measure of Quality for D [Field & Olshausen (‘96)] [Engan et. al. (‘99)] [Lewicki & Sejnowski (‘00)] [Cotter et. al. (‘03)] [Gribonval et. al. (‘04)] [Aharon, E. & Bruckstein (‘04)] [Aharon, E. & Bruckstein (‘05)] Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 57. 30 D Initialize D Sparse Coding Nearest Neighbor XT Dictionary Update Column-by-Column by Mean computation over the relevant examples K–Means For Clustering Clustering: An extreme sparse representation Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 58. 31 D Initialize D Sparse Coding Use Matching Pursuit XT Dictionary Update Column-by-Column by SVD computation over the relevant examples The K–SVD Algorithm – General [Aharon, E. & Bruckstein (‘04,‘05)] Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 59. 32 D D is known! For the jth item we solve XT K–SVD: Sparse Coding Stage Solved by A Pursuit Algorithm Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 60. 33 Fixing all A and D apart from the kth column, and seek both dk and the kth column in A to better fit the residual! We should solve: K–SVD: Dictionary Update Stage We refer only to the examples that use the column dk D SVD Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 61. Image Denoising & Beyond Via Learned Dictionaries and Sparse representations By: Michael Elad 34 To Summarize So Far … Image denoising (and many other problems in image processing) requires a model for the desired image We proposed a model for signals/images based on sparse and redundant representations Problems? What do we do? We have seen approximation methods that find the sparsest solution, and theoretical results that guarantee their success. We also saw a way to learn D Will it all work in applications? What next?
  • 62. 35 Part IVBack to Denoising … and Beyond – Combining it All Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 63.
  • 64. So, how should large images be handled?
  • 65. The solution: Force shift-invariant sparsity - on each patch of size N-by-N (N=8) in the image, including overlaps. Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 66.
  • 67.
  • 68. Simply sweep through all patches of size N-by-N (overlapping blocks),
  • 69. Image of size 10002 pixels ~106 examples to use – more than enough.
  • 70. This works much better!Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 71. 38 ComputeD to minimize using SVD, updating one column at a time Compute x by which is a simple averaging of shifted patches Computeij per patch using the matching pursuit K-SVD K-SVD Image Denoising D? x=y and D known x and ij known D and ij known Complexity of this algorithm: O(N2×K×L×Iterations) per pixel. For N=8, L=1, K=256, and 10 iterations, we need 160,000 (!!) operations per pixel. Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 72.
  • 73. In a recent work that extended this algorithm to use joint sparse representation on the patches, the best published denoising performance are obtained [Mairal, Bach, Ponce, Sapiro & Zisserman (‘09)].Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 74.
  • 75. The solution with the above algorithm is simple – consider 3D patches or 8-by-8 with the 3 color layers, and the dictionary will detect the proper relations. Original Noisy (20.43dB) Result (30.75dB) Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 76. 41 Original Noisy (12.77dB) Result (29.87dB) Denoising (Color) [Mairal, E. & Sapiro (‘08)] Our experiments lead to state-of-the-art denoising results, giving ~1dB better results compared to [Mcauley et. al. (‘06)] which implements a learned MRF model (Field-of-Experts) Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 77. 42 Video Denoising [Protter & E. (‘09)] When turning to handle video, one could improve over the previous scheme in three important ways: Propagate the dictionary from one frame to another, and thus reduce the number of iterations; Use 3D patches that handle the motion implicitly; and Motion estimation and compensation can and should be avoided [Buades, Col, and Morel (‘06)]. Our experiments lead to state-of-the-art video denoising results, giving ~0.5dB better results on average compared to [Boades, Coll & Morel (‘05)] and comparable to [Rusanovskyy, Dabov, & Egiazarian (‘06)] Original Noisy (σ=25) Denoised (PSNR=27.62) Original Noisy (σ=15) Denoised (PSNR=29.98) Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 78.
  • 79. In medicine, CT projections are obtained by X-ray, and it typically requires a high dosage of radiation in order to obtain a good quality reconstruction.
  • 80. A lower-dosage projection implies a stronger noise (Poisson distributed) in data to work with.
  • 81. Armed with sparse and redundant representation modeling, we can denoise the data and the final reconstruction … enabling CT with lower dosage.Denoising of the sinogram and post-processing (another denoising stage) of the reconstruction PSNR=26.06dB FBP result with low dosage (one fifth) PSNR=22.31dB
  • 82.
  • 83. Missing values in x imply missing rows in this linear system.
  • 84. By removing these rows, we get .
  • 86. If α0 was sparse enough, it will be the solution of the above problem! Thus, computing Dα0recovers x perfectly.= Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 87.
  • 88. Assume: the signal x has been created by x=Dα0with very sparse α0.
  • 89. Multiply this set of equations by the matrix Q which reduces the number of rows.
  • 90. The new, smaller, system of equations is
  • 91. If α0 was sparse enough, it will be the sparsest solution of the new system, thus, computing Dα0recovers x perfectly.
  • 92. Compressed sensing focuses on conditions for this to happen, guaranteeing such recovery.Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 93. 46 Inpainting[Mairal, E. & Sapiro (‘08)] Our experiments lead to state-of-the-art inpainting results. Original 80% missing Result Original 80% missing Result Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 94. 47 Inpainting[Mairal, E. & Sapiro (‘08)] The same can be done for video, very much like the denoising treatment: (i) 3D patches, (ii) no need to compute the dictionary from scratch for each frame, and (iii) no need for explicit motion estimation Original 80% missing Result Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 95.
  • 96. Generalizing the inpainting scheme to handle demosaicing is tricky because of the possibility to learn the mosaic pattern within the dictionary.
  • 97. In order to avoid “over-fitting”, we handle the demosaicing problem while forcing strong sparsity and applying only few iterations. Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 98.
  • 99. General purpose methods (JPEG, JPEG2000) do not take into account the specific family.
  • 100. By adapting to the image-content (PCA/K-SVD), better results could be obtained.
  • 101. For these techniques to operate well, traindictionaries locally (per patch) using a training set of images is required.
  • 102. In PCA, only the (quantized) coefficients are stored, whereas the K-SVD requires storage of the indices as well.
  • 103. Geometric alignment of the image is very helpful and should be done [Goldenberg, Kimmel, & E. (‘05)]. Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 104. 50 Divide the image into disjoint 15-by-15 patches. For each compute mean and dictionary On the training set Per each patch find the operating parameters (number of atoms L, quantization Q) Warp, remove the mean from each patch, sparse code using L atoms, apply Q, and dewarp On the test image Image Compression Detect main features and warp the images to a common reference (20 parameters) Training set (2500 images) Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 105. 51 11.99 10.49 8.81 5.56 10.83 8.92 7.89 4.82 10.93 8.71 8.61 5.58 Image Compression Results Original JPEG JPEG-2000 Local-PCA K-SVD Results for 820 Bytes per each file Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 106. 52 15.81 13.89 10.66 6.60 14.67 12.41 9.44 5.49 15.30 12.57 10.27 6.36 Image Compression Results Original JPEG JPEG-2000 Local-PCA K-SVD Results for 550 Bytes per each file Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 107. 53 ? 18.62 12.30 7.61 ? 16.12 11.38 6.31 ? 16.81 12.54 7.20 Image Compression Results Original JPEG JPEG-2000 Local-PCA K-SVD Results for 400 Bytes per each file Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 108. Image Denoising & Beyond Via Learned Dictionaries and Sparse representations By: Michael Elad 54 Deblocking the Results [Bryt and E. (`09)] 550 bytes K-SVD results with and without deblocking K-SVD (6.60) K-SVD (11.67) K-SVD (6.45) K-SVD (5.49) Deblock (6.24) Deblock (11.32) Deblock (6.03) Deblock (5.27)
  • 109.
  • 110. Image scale-up using bicubic interpolation is far from being satisfactory for this task.
  • 111. Recently, a sparse and redundant representation technique was proposed [Yang, Wright, Huang, and Ma (’08)] for solving this problem, by training a coupled-dictionaries for the low- and high res. images.
  • 112. We extended and improved their algorithms and results.SR Result PSNR=16.95dB The training image: 717×717 pixels, providing a set of 54,289 training patch-pairs. Bicubic interpolation PSNR=14.68dB Given Image Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad 55
  • 113. Super-Resolution – Results (2) Given image Scaled-Up (factor 2:1) using the proposed algorithm, PSNR=29.32dB (3.32dB improvement over bicubic) Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad 56
  • 114. Super-Resolution – Results (2) The Original Bicubic Interpolation SR result Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad 57
  • 115. Super-Resolution – Results (2) The Original Bicubic Interpolation SR result Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad 58
  • 116. Image Denoising & Beyond Via Learned Dictionaries and Sparse representations By: Michael Elad 59 To Summarize So Far … Image denoising (and many other problems in image processing) requires a model for the desired image We proposed a model for signals/images based on sparse and redundant representations Well, does this work? What do we do? Yes! We have seen a group of applications where this model is showing very good results: denoising of bw/color stills/video, CT improvement, inpainting, super-resolution, and compression Well … many more things … So, what next?
  • 117. 60 Part V Summary and Conclusion Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 118.
  • 120. Demonstrating on other applications (graphics?)
  • 121. … What next? Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 122. 62 Thank You All this Work is Made Possible Due to my teachers and mentors colleagues & friends collaborating with me and my students A.M. Bruckstein D.L. Donoho G. Sapiro J.L. Starck I. Yavneh M. Zibulevsky M. Aharon O. Bryt J. Mairal M. Protter R. Rubinstein J. Shtok R. Giryes Z. Ben-Haim J. Turek R. Zeyde Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 123. 63 And Also … Thank you so much to the organizers of this event for inviting me to give this talk. Seventh International Conference on Curves and Surfaces Avignon - FRANCE June 24-30, 2010 Sparse and Redundant Representation Modeling of Signals – Theory and Applications By: Michael Elad
  • 124. Image Denoising & Beyond Via Learned Dictionaries and Sparse representations By: Michael Elad 64 If you are Interested … More on this topic (including the slides, the papers, and Matlab toolboxes) can be found in my webpage: http://www.cs.technion.ac.il/~elad A new book on this topic will be published on ~August.