infimal convolution type regularization for inverse
play

Infimal-convolution-type regularization for inverse problems in - PowerPoint PPT Presentation

Sep 23, 2022 •463 likes •2k views

INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Infimal-convolution-type regularization for inverse problems in imaging Kristian Bredies Institute for Mathematics and Scientific Computing University of Graz IHP semester The Mathematics

  1. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Infimal convolution for inverse problems Tikhonov regularization: Solve min u ∈ X S ( Ku , f ) + Φ( u ) S discrepancy for Ku = f , Φ regularization functional ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 10 / 51

  2. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Infimal convolution for inverse problems Tikhonov regularization: Solve min u ∈ X S ( Ku , f ) + Φ( u ) S discrepancy for Ku = f , Φ regularization functional Two viewpoints: 1 Monolithic regularization min u ∈ X S ( Ku , f ) + Φ( u ), Φ = Φ 1 � . . . � Φ m 2 Vector-valued regularization � � ( u 1 ,..., u m ) ∈ X m S min K ( u 1 + . . . + u m ) , f + Φ 1 ( u 1 ) + . . . Φ m ( u m ) ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 10 / 51

  3. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Infimal convolution for inverse problems Tikhonov regularization: Solve min u ∈ X S ( Ku , f ) + Φ( u ) S discrepancy for Ku = f , Φ regularization functional Two viewpoints: 1 Monolithic regularization min u ∈ X S ( Ku , f ) + Φ( u ), Φ = Φ 1 � . . . � Φ m 2 Vector-valued regularization � � ( u 1 ,..., u m ) ∈ X m S min K ( u 1 + . . . + u m ) , f + Φ 1 ( u 1 ) + . . . Φ m ( u m ) Monolithic regularization: Intrinsic decomposition & problem-adapted interpretation ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 10 / 51

  4. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING General existence result Theorem: X reflexive Banach space, Y normed space K : X → Y bounded linear S ( · , f ) : Y → [0 , ∞ ] proper, convex, l.s.c. Φ : X ∈ [0 , ∞ ] proper, convex, l.s.c. Φ 1-homogeneous, dim ker(Φ) < ∞ , and coercive, i.e., � u − Pu � X ≤ C Φ( u ) for P : X → ker(Φ) bounded linear projector Then: min u ∈ X S ( Ku , f ) + α Φ( u ) has a solution for each α > 0 ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 11 / 51

  5. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING General regularization properties min u ∈ X S ( Ku , f ) + α Φ( u ) Let: Y Hilbert space, S ( v , f ) = 1 2 � v − f � 2 Y Then: (Subsequential) stability for varying f Noise level δ → 0 and α → 0, δ 2 /α → 0 ⇒ (Subsequential) convergence Source condition K ∗ w ∈ ∂ Φ( u † ), α ∼ δ ⇒ O ( δ ) for the Bregman distance w.r.t. Φ ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 12 / 51

  6. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING General regularization properties min u ∈ X S ( Ku , f ) + α Φ( u ) Let: Y Hilbert space, S ( v , f ) = 1 2 � v − f � 2 Y Then: (Subsequential) stability for varying f Noise level δ → 0 and α → 0, δ 2 /α → 0 ⇒ (Subsequential) convergence Source condition K ∗ w ∈ ∂ Φ( u † ), α ∼ δ ⇒ O ( δ ) for the Bregman distance w.r.t. Φ Difficulties: dim ker(Φ) = ∞ ⇒ K has to be stably invertible on ker(Φ) Φ not coercive ⇒ K has to be stably invertible on X ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 12 / 51

  7. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Infimal convolution regularization Lemma: Φ 1 , Φ 2 : X → [0 , ∞ ] proper, convex, l.s.c. Φ i 1-homogeneous, dim ker(Φ) < ∞ for i = 1 , 2 Then: Φ 1 � Φ 2 : X → [0 , ∞ ] is exact, proper, convex, l.s.c., 1-homogeneous, ker Φ 1 � Φ 2 = ker Φ 1 + ker Φ 2 ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 13 / 51

  8. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Infimal convolution regularization Lemma: Φ 1 , Φ 2 : X → [0 , ∞ ] proper, convex, l.s.c. Φ i 1-homogeneous, dim ker(Φ) < ∞ for i = 1 , 2 Then: Φ 1 � Φ 2 : X → [0 , ∞ ] is exact, proper, convex, l.s.c., 1-homogeneous, ker Φ 1 � Φ 2 = ker Φ 1 + ker Φ 2 Lemma: Additionally: X ֒ − ֒ → Z , Z Banach space and � � � u � X ≤ C � u � Z + Φ i ( u ) for i = 1 , 2 and all u ∈ X Then: Φ 1 � Φ 2 is coercive, i.e., � u − Pu � X ≤ C Φ( u ) for all u ∈ X ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 13 / 51

  9. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Infimal convolution regularization Lemma: Φ 1 , Φ 2 : X → [0 , ∞ ] proper, convex, l.s.c. Φ i 1-homogeneous, dim ker(Φ) < ∞ for i = 1 , 2 Then: Φ 1 � Φ 2 : X → [0 , ∞ ] is exact, proper, convex, l.s.c., 1-homogeneous, ker Φ 1 � Φ 2 = ker Φ 1 + ker Φ 2 Lemma: Additionally: X ֒ − ֒ → Z , Z Banach space and � � � u � X ≤ C � u � Z + Φ i ( u ) for i = 1 , 2 and all u ∈ X Then: Φ 1 � Φ 2 is coercive, i.e., � u − Pu � X ≤ C Φ( u ) for all u ∈ X � sufficient conditions for a regularizer ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 13 / 51

  10. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Examples TV - TV 2 infimal convolution: TV and TV 2 have finite-dimensional kernels + embedding: � � � u � 1 + TV 2 ( u ) � u � BV ≤ � u � BV 2 ≤ C Consequently: TV � β TV 2 is a regularizer ( β > 0) ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 14 / 51

  11. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Examples TV - TV 2 infimal convolution: TV and TV 2 have finite-dimensional kernels + embedding: � � � u � 1 + TV 2 ( u ) � u � BV ≤ � u � BV 2 ≤ C Consequently: TV � β TV 2 is a regularizer ( β > 0) TV - L 1 ◦ ∆ infimal convolution: ker(TV � � · � M ◦ ∆) = { u ∈ BV | u harmonic } Infinite-dimensional kernel ⇒ not a regularizer ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 14 / 51

  12. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Examples TV - TV 2 infimal convolution: TV and TV 2 have finite-dimensional kernels + embedding: � � � u � 1 + TV 2 ( u ) � u � BV ≤ � u � BV 2 ≤ C Consequently: TV � β TV 2 is a regularizer ( β > 0) TV - L 1 ◦ ∆ infimal convolution: ker(TV � � · � M ◦ ∆) = { u ∈ BV | u harmonic } Infinite-dimensional kernel ⇒ not a regularizer TV - G -norm/ TV - H − 1 infimal convolution: � · � TV ∗ / � · � H − 1 not coercive in L p -spaces Infimal convolution not coercive ⇒ not a regularizer ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 14 / 51

  13. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Examples TV - TV 2 infimal convolution: TV and TV 2 have finite-dimensional kernels + embedding: � � � u � 1 + TV 2 ( u ) � u � BV ≤ � u � BV 2 ≤ C Consequently: TV � β TV 2 is a regularizer ( β > 0) TV - L 1 ◦ ∆ infimal convolution: ker(TV � � · � M ◦ ∆) = { u ∈ BV | u harmonic } Infinite-dimensional kernel ⇒ not a regularizer TV - G -norm/ TV - H − 1 infimal convolution: � · � TV ∗ / � · � H − 1 not coercive in L p -spaces Infimal convolution not coercive ⇒ not a regularizer � Develop regularizing cartoon/texture models + complex smoothness-based regularizers ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 14 / 51

  14. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Outline 1 Infimal convolution regularization 2 Total generalized variation Definition and properties Applications 3 Infimal convolution TGV Accelerated dynamic MRI 4 Infimal convolution of oscillation TGV Oscillation TGV and infimal convolution Numerical realization Applications 5 Summary ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 15 / 51

  15. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Total generalized variation Motivation: TV-based first-order regularization favors certain artifacts Total generalized variation: [B./Kunisch/Pock ’10] �� � u div k v d x TGV k � α ( u ) = sup noisy image � Ω c (Ω , Sym k (I R d )) , v ∈ C k � � div l v � ∞ ≤ α l , l = 0 , . . . , k − 1 α = ( α 0 , . . . , α k − 1 ) > 0 weights Second-order version: TGV 2 α ( u )= min α 1 �∇ u − w � M + α 0 �E w � M TV-regularization w ∈ BD(Ω) ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 16 / 51

  16. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Total generalized variation Motivation: TV-based first-order regularization favors certain artifacts Total generalized variation: [B./Kunisch/Pock ’10] �� � u div k v d x TGV k � α ( u ) = sup noisy image � Ω c (Ω , Sym k (I R d )) , v ∈ C k � � div l v � ∞ ≤ α l , l = 0 , . . . , k − 1 α = ( α 0 , . . . , α k − 1 ) > 0 weights Second-order version: TGV 2 α ( u )= min α 1 �∇ u − w � M + α 0 �E w � M TGV 2 α -regularization w ∈ BD(Ω) ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 16 / 51

  17. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Properties of TGV 2 α Basic properties: [B./Kunisch/Pock ’10] TGV 2 α is proper, convex, lower semi-continuous TGV 2 α is translation and rotation invariant TGV 2 α + � · � 1 gives the Banach space BGV 2 α (Ω) α ) = Π 1 affine functions ker(TGV 2 TGV 2 α measures piecewise affine only at the interfaces ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 17 / 51

  18. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Properties of TGV 2 α Basic properties: [B./Kunisch/Pock ’10] TGV 2 α is proper, convex, lower semi-continuous TGV 2 α is translation and rotation invariant TGV 2 α + � · � 1 gives the Banach space BGV 2 α (Ω) α ) = Π 1 affine functions ker(TGV 2 TGV 2 α measures piecewise affine only at the interfaces ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 17 / 51

  19. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Properties of TGV 2 α Basic properties: [B./Kunisch/Pock ’10] TGV 2 α is proper, convex, lower semi-continuous TGV 2 α is translation and rotation invariant TGV 2 α + � · � 1 gives the Banach space BGV 2 α (Ω) α ) = Π 1 affine functions ker(TGV 2 TGV 2 α measures piecewise affine only at the interfaces ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 17 / 51

  20. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Properties of TGV 2 α Basic properties: [B./Kunisch/Pock ’10] TGV 2 α is proper, convex, lower semi-continuous TGV 2 α is translation and rotation invariant TGV 2 α + � · � 1 gives the Banach space BGV 2 α (Ω) α ) = Π 1 affine functions ker(TGV 2 TGV 2 α measures piecewise affine only at the interfaces ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 17 / 51

  21. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Properties of TGV 2 α Basic properties: [B./Kunisch/Pock ’10] TGV 2 α is proper, convex, lower semi-continuous TGV 2 α is translation and rotation invariant TGV 2 α + � · � 1 gives the Banach space BGV 2 α (Ω) α ) = Π 1 affine functions ker(TGV 2 TGV 2 α measures piecewise affine only at the interfaces ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 17 / 51

  22. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Properties of TGV 2 α Basic properties: [B./Kunisch/Pock ’10] TGV 2 α is proper, convex, lower semi-continuous TGV 2 α is translation and rotation invariant TGV 2 α + � · � 1 gives the Banach space BGV 2 α (Ω) α ) = Π 1 affine functions ker(TGV 2 TGV 2 α measures piecewise affine only at the interfaces Advanced properties: [B./Valkonen ’11] BGV 2 α (Ω) = BV(Ω) in the sense of equivalent norms TGV 2 α is coercive in the sense � u − Pu � BV ≤ C TGV 2 α ( u ) for P : L 1 (Ω) → Π 1 continuous projection ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 17 / 51

  23. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Properties of TGV 2 α Basic properties: [B./Kunisch/Pock ’10] TGV 2 α is proper, convex, lower semi-continuous TGV 2 α is translation and rotation invariant TGV 2 α + � · � 1 gives the Banach space BGV 2 α (Ω) α ) = Π 1 affine functions ker(TGV 2 TGV 2 α measures piecewise affine only at the interfaces Advanced properties: [B./Valkonen ’11] BGV 2 α (Ω) = BV(Ω) in the sense of equivalent norms TGV 2 α is coercive in the sense � u − Pu � BV ≤ C TGV 2 α ( u ) for P : L 1 (Ω) → Π 1 continuous projection ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 17 / 51

  24. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Existence and stability Theorem: [B./Valkonen ’11]   1 < p ≤ d / ( d − 1) Optimization problem     K : L p (Ω) → H 1     2 � Ku − f � 2 min linear and continuous, ⇒ u ∈ L p (Ω) + TGV 2 α ( u )   H Hilbert space       K injective on Π 1 possesses a solution ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 18 / 51

  25. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Existence and stability Theorem: [B./Valkonen ’11]   1 < p ≤ d / ( d − 1) Optimization problem     K : L p (Ω) → H 1     2 � Ku − f � 2 min linear and continuous, ⇒ u ∈ L p (Ω) + TGV 2 α ( u )   H Hilbert space       K injective on Π 1 possesses a solution � u n ⇀ u in L p (Ω) (subseq.) Stability: f n → f in H ⇒ TGV 2 α ( u n ) → TGV 2 α ( u ) ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 18 / 51

  26. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Example: Denoising a “cartoon” image � u − f � 2 + TGV 2 Solve: min α ( u ) 2 u ∈ L 2 (Ω) noisy image ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 19 / 51

  27. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Example: Denoising a “cartoon” image � u − f � 2 + TGV 2 Solve: min α ( u ) 2 u ∈ L 2 (Ω) TV regularization ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 19 / 51

  28. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Example: Denoising a “cartoon” image � u − f � 2 + TGV 2 Solve: min α ( u ) 2 u ∈ L 2 (Ω) TGV 2 α regularization ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 19 / 51

  29. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Electron tomography Joint work with Georg Haberfehlner and Richard Huber Scanning Transmission Electron Microscopy (STEM): Focused electron beam High Angle Annular Dark Field (HAADF) � non Bragg-scattered electrons � proportional to mass-thickness � Reconstruction via Radon inversion ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 20 / 51

  30. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Regularized reconstruction Reconstruction of one slice: Sinogram Projections ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 21 / 51

  31. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Regularized reconstruction Reconstruction of one slice: Filtered back-projection Projections ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 21 / 51

  32. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Regularized reconstruction Reconstruction of one slice: TV-regularization Projections ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 21 / 51

  33. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Regularized reconstruction Reconstruction of one slice: TGV-regularization Projections ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 21 / 51

  34. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Outline 1 Infimal convolution regularization 2 Total generalized variation Definition and properties Applications 3 Infimal convolution TGV Accelerated dynamic MRI 4 Infimal convolution of oscillation TGV Oscillation TGV and infimal convolution Numerical realization Applications 5 Summary ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 22 / 51

  35. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Regularization for image sequences Motivation: Straightforward spatio-temporal regularization � � ǫ | ∂ x 1 u | 2 + ǫ | ∂ x 2 u | 2 + | ∂ t u | 2 d t d x TV ǫ x ( u ) = [0 , T ] × Ω ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 23 / 51

  36. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Regularization for image sequences Motivation: Straightforward spatio-temporal regularization � � ǫ | ∂ x 1 u | 2 + ǫ | ∂ x 2 u | 2 + | ∂ t u | 2 d t d x TV ǫ x ( u ) = [0 , T ] × Ω Compressed TV ǫ x ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 23 / 51

  37. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Regularization for image sequences Motivation: Straightforward spatio-temporal regularization � � | ∂ x 1 u | 2 + | ∂ x 2 u | 2 + ǫ | ∂ t u | 2 d t d x TV ǫ t ( u ) = [0 , T ] × Ω ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 23 / 51

  38. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Regularization for image sequences Motivation: Straightforward spatio-temporal regularization � � | ∂ x 1 u | 2 + | ∂ x 2 u | 2 + ǫ | ∂ t u | 2 d t d x TV ǫ t ( u ) = [0 , T ] × Ω Compressed TV ǫ t ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 23 / 51

  39. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Regularization for image sequences Motivation: Straightforward spatio-temporal regularization � � | ∂ x 1 u | 2 + | ∂ x 2 u | 2 + ǫ | ∂ t u | 2 d t d x TV ǫ t ( u ) = [0 , T ] × Ω TV ǫ t ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 23 / 51

  40. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Regularization for image sequences Motivation: Straightforward spatio-temporal regularization � � | ∂ x 1 u | 2 + | ∂ x 2 u | 2 + ǫ | ∂ t u | 2 d t d x TV ǫ t ( u ) = [0 , T ] × Ω ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 23 / 51

  41. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Regularization for image sequences Motivation: Straightforward spatio-temporal regularization � � | ∂ x 1 u | 2 + | ∂ x 2 u | 2 + ǫ | ∂ t u | 2 d t d x TV ǫ t ( u ) = [0 , T ] × Ω Approach: Optimal balancing with infimal convolution (IC) ICTV ǫ = TV ǫ x � TV ǫ t ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 23 / 51

  42. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Regularization for image sequences Motivation: Straightforward spatio-temporal regularization � � | ∂ x 1 u | 2 + | ∂ x 2 u | 2 + ǫ | ∂ t u | 2 d t d x TV ǫ t ( u ) = [0 , T ] × Ω Approach: Optimal balancing with infimal convolution (IC) ICTV ǫ = TV ǫ x � TV ǫ t Compressed ICTV ǫ ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 23 / 51

  43. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Regularization for image sequences Motivation: Straightforward spatio-temporal regularization � � | ∂ x 1 u | 2 + | ∂ x 2 u | 2 + ǫ | ∂ t u | 2 d t d x TV ǫ t ( u ) = [0 , T ] × Ω Approach: Optimal balancing with infimal convolution (IC) ICTV ǫ = TV ǫ x � TV ǫ t ICTV ǫ ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 23 / 51

  44. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Regularization for image sequences Motivation: Straightforward spatio-temporal regularization � � | ∂ x 1 u | 2 + | ∂ x 2 u | 2 + ǫ | ∂ t u | 2 d t d x TV ǫ t ( u ) = [0 , T ] × Ω Approach: Optimal balancing with infimal convolution (IC) ICTV ǫ = TV ǫ x � TV ǫ t u 1 u 2 ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 23 / 51

  45. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Regularization for image sequences Motivation: Straightforward spatio-temporal regularization � � | ∂ x 1 u | 2 + | ∂ x 2 u | 2 + ǫ | ∂ t u | 2 d t d x TV ǫ t ( u ) = [0 , T ] × Ω Approach: Optimal balancing with infimal convolution (IC) ICTV ǫ = TV ǫ x � TV ǫ t � Use infimal convolution of TGV u 1 u 2 ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 23 / 51

  46. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Infimal convolution TGV Definition: Anisotropic Total Generalized Variation �� � u div k v d x TGV k � � v ∈ C k c (Ω , Sym k (I R d )) , β ( u ) = sup � Ω � div l v � ∞ ,β ∗ l ≤ 1, l=0,. . . ,k-1 β = ( | · | β 0 , . . . , | · | β k − 1 ) tensor norms � v l � ∞ ,β ∗ l = sup x ∈ Ω | v l ( x ) | β ∗ l , | · | β ∗ l dual norm ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 24 / 51

  47. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Infimal convolution TGV Definition: Anisotropic Total Generalized Variation �� � u div k v d x TGV k � � v ∈ C k c (Ω , Sym k (I R d )) , β ( u ) = sup � Ω � div l v � ∞ ,β ∗ l ≤ 1, l=0,. . . ,k-1 β = ( | · | β 0 , . . . , | · | β k − 1 ) tensor norms � v l � ∞ ,β ∗ l = sup x ∈ Ω | v l ( x ) | β ∗ l , | · | β ∗ l dual norm Infimal convolution TGV: [Holler/Kunisch ’14] ICTGV k β = TGV k 1 β 1 � . . . � TGV k m β m ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 24 / 51

  48. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Infimal convolution TGV Definition: Anisotropic Total Generalized Variation �� � u div k v d x TGV k � v ∈ C k � c (Ω , Sym k (I R d )) , β ( u ) = sup � Ω � div l v � ∞ ,β ∗ l ≤ 1, l=0,. . . ,k-1 β = ( | · | β 0 , . . . , | · | β k − 1 ) tensor norms � v l � ∞ ,β ∗ l = sup x ∈ Ω | v l ( x ) | β ∗ l , | · | β ∗ l dual norm Infimal convolution TGV: [Holler/Kunisch ’14] ICTGV k β = TGV k 1 β 1 � . . . � TGV k m β m Regularization properties: Each TGV k i β i has finite-dimensional kernel + embedding � � � u � 1 + TGV k i � u � BV ≤ C β i ( u ) � ICTGV is a regularizer ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 24 / 51

  49. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Accelerated dynamic MRI Joint work with Martin Holler and Matthias Schl¨ ogl Dynamic MRI: Acquire image sequences of moving objects Due to limited acquisition speed, only low-resolution images are obtainable sum of squares ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 25 / 51

  50. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Accelerated dynamic MRI Joint work with Martin Holler and Matthias Schl¨ ogl Dynamic MRI: Acquire image sequences of moving objects Due to limited acquisition speed, only low-resolution images are obtainable sum of squares ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 25 / 51

  51. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Accelerated dynamic MRI Joint work with Martin Holler and Matthias Schl¨ ogl Dynamic MRI: Acquire image sequences of moving objects Due to limited acquisition speed, only low-resolution images are obtainable sum of squares Goal: Improve spatio-temporal resolution by undersampling reconstruction ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 25 / 51

  52. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Accelerated dynamic MRI Joint work with Martin Holler and Matthias Schl¨ ogl Dynamic MRI: Acquire image sequences of moving objects Due to limited acquisition speed, only low-resolution images are obtainable sum of squares Goal: Improve spatio-temporal resolution by undersampling reconstruction � Apply ICTGV regularized regularization ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 25 / 51

  53. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING The variational model Minimization problem: λ � 2 � K t , c ( u t ) − d t , c � 2 2 + ICTGV 2 min β ( u ) u t , c K t , c ( u t ) = M t F ( u t σ c ) masked Fourier transform ( σ c complex coil sensitivities) ICTGV 2 β = TGV 2 β 1 � TGV 2 β 2 ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 26 / 51

  54. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING The variational model Minimization problem: λ � 2 � K t , c ( u t ) − d t , c � 2 2 + ICTGV 2 min β ( u ) u t , c K t , c ( u t ) = M t F ( u t σ c ) masked Fourier transform ( σ c complex coil sensitivities) ICTGV 2 β = TGV 2 β 1 � TGV 2 β 2 Primal-dual algorithm: Guaranteed convergence, duality-based stopping criterion GPU-optimized version: ≈ 160 seconds including coil-sensitivity estimation (NVidia GeForce GTX770 with AGILE library) ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 26 / 51

  55. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Numerical test Acceleration factor 8: Reference data Unregularized reconstruction ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 27 / 51

  56. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Numerical test Acceleration factor 8: Reference data Unregularized reconstruction ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 27 / 51

  57. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Numerical test Acceleration factor 8: Low rank + sparse model Difference ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 27 / 51

  58. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Numerical test Acceleration factor 8: ICTGV model Difference ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 27 / 51

  59. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Numerical test Acceleration factor 16: ICTGV model Difference ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 27 / 51

  60. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Numerical test Acceleration factor 8: Slow component Fast component ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 27 / 51

  61. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Numerical test Acceleration factor 8: Slow component Fast component Favorable quantitative comparison 2nd place at the ISMRM 2013 reconstruction challenge ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 27 / 51

  62. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Outline 1 Infimal convolution regularization 2 Total generalized variation Definition and properties Applications 3 Infimal convolution TGV Accelerated dynamic MRI 4 Infimal convolution of oscillation TGV Oscillation TGV and infimal convolution Numerical realization Applications 5 Summary ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 28 / 51

  63. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Motivation: Denoising “barbara” noisy image ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 29 / 51

  64. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Motivation: Denoising “barbara” TGV 2 α regularization ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 29 / 51

  65. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Motivation: Denoising “barbara” TGV 2 α regularization � Capture oscillatory structures � TGV osci α,β, c ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 29 / 51

  66. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Oscillation TGV Joint work with Yiming Gao Idea: Choose differential equation with oscillatory functions as solutions ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 30 / 51

  67. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Oscillation TGV Joint work with Yiming Gao Idea: Choose differential equation with oscillatory functions as solutions R d , ω � = 0 ∇ 2 u + c u = 0 , c = ω ⊗ ω, ω ∈ I Solutions: u ( x ) = C 1 cos( ω · x ) + C 2 sin( ω · x ) , C 1 , C 2 ∈ I R ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 30 / 51

  68. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Oscillation TGV Joint work with Yiming Gao Idea: Choose differential equation with oscillatory functions as solutions R d , ω � = 0 ∇ 2 u + c u = 0 , c = ω ⊗ ω, ω ∈ I Solutions: u ( x ) = C 1 cos( ω · x ) + C 2 sin( ω · x ) , C 1 , C 2 ∈ I R Adapt second-order TGV to the corresponding operator ∇ 2 u + c u = 0 ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 30 / 51

  69. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Oscillation TGV Joint work with Yiming Gao Idea: Choose differential equation with oscillatory functions as solutions ∇ 2 u + c u = 0 , R d , ω � = 0 c = ω ⊗ ω, ω ∈ I Solutions: u ( x ) = C 1 cos( ω · x ) + C 2 sin( ω · x ) , C 1 , C 2 ∈ I R Adapt second-order TGV to the corresponding operator ∇ u − w = 0 , E w + c u = 0 ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 30 / 51

  70. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Oscillation TGV Joint work with Yiming Gao Idea: Choose differential equation with oscillatory functions as solutions ∇ 2 u + c u = 0 , R d , ω � = 0 c = ω ⊗ ω, ω ∈ I Solutions: u ( x ) = C 1 cos( ω · x ) + C 2 sin( ω · x ) , C 1 , C 2 ∈ I R Adapt second-order TGV to the corresponding operator ∇ u − w = 0 , E w + c u = 0 Total generalized variation (second order): � � TGV 2 α,β ( u ) = w ∈ BD(Ω) α min d |∇ u − w | + β d |E w | Ω Ω α > 0, β > 0 ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 30 / 51

  71. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Oscillation TGV Joint work with Yiming Gao Idea: Choose differential equation with oscillatory functions as solutions ∇ 2 u + c u = 0 , R d , ω � = 0 c = ω ⊗ ω, ω ∈ I Solutions: u ( x ) = C 1 cos( ω · x ) + C 2 sin( ω · x ) , C 1 , C 2 ∈ I R Adapt second-order TGV to the corresponding operator ∇ u − w = 0 , E w + c u = 0 Oscillation total generalized variation: � � TGV osci α,β, c ( u ) = w ∈ BD(Ω) α min d |∇ u − w | + β d |E w + c u | Ω Ω R d , ω � = 0 α > 0, β > 0, c = ω ⊗ ω, ω ∈ I ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 30 / 51

  72. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Properties of TGV osci α,β, c Basic properties: [Gao/B. ’17] TGV osci α,β, c is proper, convex, lower semi-continuous TGV osci α,β, c is translation and rotation invariant TGV osci α,β, c + � · � 1 gives the Banach space BGV osci α,β, c (Ω) ker(TGV osci α,β, c ) = span { x �→ cos( ω · x ) , x �→ sin( ω · x ) } TGV osci α,β, c measures piecewise oscillations only at interfaces ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 31 / 51

  73. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Properties of TGV osci α,β, c Basic properties: [Gao/B. ’17] TGV osci α,β, c is proper, convex, lower semi-continuous TGV osci α,β, c is translation and rotation invariant TGV osci α,β, c + � · � 1 gives the Banach space BGV osci α,β, c (Ω) ker(TGV osci α,β, c ) = span { x �→ cos( ω · x ) , x �→ sin( ω · x ) } TGV osci α,β, c measures piecewise oscillations only at interfaces ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 31 / 51

  74. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Properties of TGV osci α,β, c Basic properties: [Gao/B. ’17] TGV osci α,β, c is proper, convex, lower semi-continuous TGV osci α,β, c is translation and rotation invariant TGV osci α,β, c + � · � 1 gives the Banach space BGV osci α,β, c (Ω) ker(TGV osci α,β, c ) = span { x �→ cos( ω · x ) , x �→ sin( ω · x ) } TGV osci α,β, c measures piecewise oscillations only at interfaces ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 31 / 51

  75. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Properties of TGV osci α,β, c Basic properties: [Gao/B. ’17] TGV osci α,β, c is proper, convex, lower semi-continuous TGV osci α,β, c is translation and rotation invariant TGV osci α,β, c + � · � 1 gives the Banach space BGV osci α,β, c (Ω) ker(TGV osci α,β, c ) = span { x �→ cos( ω · x ) , x �→ sin( ω · x ) } TGV osci α,β, c measures piecewise oscillations only at interfaces ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 31 / 51

  76. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Properties of TGV osci α,β, c Basic properties: [Gao/B. ’17] TGV osci α,β, c is proper, convex, lower semi-continuous TGV osci α,β, c is translation and rotation invariant TGV osci α,β, c + � · � 1 gives the Banach space BGV osci α,β, c (Ω) ker(TGV osci α,β, c ) = span { x �→ cos( ω · x ) , x �→ sin( ω · x ) } TGV osci α,β, c measures piecewise oscillations only at interfaces ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 31 / 51

  77. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Properties of TGV osci α,β, c Basic properties: [Gao/B. ’17] TGV osci α,β, c is proper, convex, lower semi-continuous TGV osci α,β, c is translation and rotation invariant TGV osci α,β, c + � · � 1 gives the Banach space BGV osci α,β, c (Ω) ker(TGV osci α,β, c ) = span { x �→ cos( ω · x ) , x �→ sin( ω · x ) } TGV osci α,β, c measures piecewise oscillations only at interfaces Advanced properties: BGV osci α,β, c (Ω) = BV(Ω) in the sense of equivalent norms TGV osci α,β, c is coercive in the sense � u − Pu � BV ≤ C TGV osci α,β, c ( u ) for P : L 1 (Ω) → ker(TGV osci α,β, c ) continuous projection ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 31 / 51

  78. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Properties of TGV osci α,β, c Basic properties: [Gao/B. ’17] TGV osci α,β, c is proper, convex, lower semi-continuous TGV osci α,β, c is translation and rotation invariant TGV osci α,β, c + � · � 1 gives the Banach space BGV osci α,β, c (Ω) ker(TGV osci α,β, c ) = span { x �→ cos( ω · x ) , x �→ sin( ω · x ) } TGV osci α,β, c measures piecewise oscillations only at interfaces Advanced properties: BGV osci α,β, c (Ω) = BV(Ω) in the sense of equivalent norms TGV osci α,β, c is coercive in the sense � u − Pu � BV ≤ C TGV osci α,β, c ( u ) for P : L 1 (Ω) → ker(TGV osci α,β, c ) continuous projection ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 31 / 51

  79. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Infimal convolution of TGV osci Next steps: Separate cartoon components from oscillatory components Allow for multiple directions and frequencies � m -fold infimal convolution ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 32 / 51

  80. INSTITUTE FOR MATHEMATICS AND SCIENTIFIC COMPUTING Infimal convolution of TGV osci Next steps: Separate cartoon components from oscillatory components Allow for multiple directions and frequencies � m -fold infimal convolution Infimal convolution of TGV osci : ICTGV osci c ( u ) = (TGV osci α 1 ,β 1 , c 1 � . . . � TGV osci α m ,β m , c m )( u ) α,� � β,� R d α i > 0, β i > 0, c i = ω i ⊗ ω i , ω i ∈ I ICTGV osci Infimal convolution TGV ICTGV Summary K. Bredies 32 / 51

Recommend

Iterative regularization for general inverse problems Guillaume Garrigos with L. Rosasco and S.

Iterative regularization for general inverse problems Guillaume Garrigos with L. Rosasco and S.

Iterative regularization for general inverse problems Guillaume Garrigos with L. Rosasco and S. Villa CNRS, cole Normale Suprieure Sminaire CVN - Centrale Suplec - 23 Jan 2018 Regularization of inverse problems 1 Regularization by

719 views • 56 slides
A new approach for regularization of inverse problems in image processing I. Souopgui 1 , 2 , E.

A new approach for regularization of inverse problems in image processing I. Souopgui 1 , 2 , E.

Inverse problems : variational formulation Generalized Diffusion regularization Application Conclusion A new approach for regularization of inverse problems in image processing I. Souopgui 1 , 2 , E. Kamgnia 2 , F.-X. Le Dimet 1 , A. Vidard 1

380 views • 25 slides
Regularization of Inverse Problems Matthias J. Ehrhardt January 28, 2019 What is an Inverse

Regularization of Inverse Problems Matthias J. Ehrhardt January 28, 2019 What is an Inverse

Regularization of Inverse Problems Matthias J. Ehrhardt January 28, 2019 What is an Inverse Problem? A : U V mapping between Hilbert spaces U , V , A L ( U , V ) physical model A , cause u and effect A ( u ) = Au . Direct /

900 views • 31 slides
Inverse Problems and Regularization An Introduction Stefan Kindermann Industrial Mathematics

Inverse Problems and Regularization An Introduction Stefan Kindermann Industrial Mathematics

Inverse Problems and Regularization An Introduction Stefan Kindermann Industrial Mathematics Institute University of Linz, Austria Introduction to Regularization What are Inverse Problems ? One possible definition [Engl, Hanke, Neubauer

481 views • 47 slides
Inverse Problems Recovering x 0 R N from noisy observations y = x 0 + w R P Inverse

Inverse Problems Recovering x 0 R N from noisy observations y = x 0 + w R P Inverse

Low Complexity Regularization of Inverse Problems Joint work with: Samuel Vaiter Jalal Fadili Gabriel Peyr VISI N www.numerical-tours.com Inverse Problems Recovering x 0 R N from noisy observations y = x 0 + w R P Inverse

574 views • 55 slides
Page 1 Inverse Problems - Examples Inverse Problems - Theory inverse problem

Page 1 Inverse Problems - Examples Inverse Problems - Theory inverse problem

Outline Theory example 1D deconvolution Fourier method Algebraic method Inverse Problems discretization matrix properties regularization solution methods Ivo Ihrke Computed Tomography (CT) Radon

304 views • 12 slides
Non-quadratic Regularization of the Inverse Problem Associated to the Black-Scholes PDE UNDER

Non-quadratic Regularization of the Inverse Problem Associated to the Black-Scholes PDE UNDER

Non-quadratic Regularization of the Inverse Problem Associated to the Black-Scholes PDE UNDER CAPRICORN 1 Thanks to Antonio Leitao!!! V. Albani (IMPA) A. De Cezaro (FURG,Brazil) O. Scherzer (U.Vienna, Austria) Jorge P . Zubelli IMPA

1.61k views • 93 slides
On iterative regularization methods for nonlinear inverse problems Antonio Leito

On iterative regularization methods for nonlinear inverse problems Antonio Leito

On iterative regularization methods for nonlinear inverse problems Antonio Leito acgleitao@gmail.com Department of Math., Federal Univ. of St. Catarina, Brazil Workshop 2: Large Scale Inv. Probl. and Appl. in the Earth Sciences RICAM, Linz,

799 views • 32 slides
Spectral regularization methods for statistical inverse learning problems G. Blanchard

Spectral regularization methods for statistical inverse learning problems G. Blanchard

Spectral regularization methods for statistical inverse learning problems G. Blanchard Universtit at Potsdam van Dantzig seminar, 23/06/2016 Joint work with N. M ucke (U. Potsdam) Rates for statistical inverse learning van Dantzig

414 views • 38 slides
1 Convolution Convolution is an important operation in signal and image processing. Convolution

1 Convolution Convolution is an important operation in signal and image processing. Convolution

CS1114 Section 6: Convolution February 27th, 2013 1 Convolution Convolution is an important operation in signal and image processing. Convolution op- erates on two signals (in 1D) or two images (in 2D): you can think of one as the input

353 views • 6 slides
Convex regularization of discrete-valued inverse problems Christian Clason Faculty of

Convex regularization of discrete-valued inverse problems Christian Clason Faculty of

Convex regularization of discrete-valued inverse problems Christian Clason Faculty of Mathematics, Universitt Duisburg-Essen joint work with Thi Bich Tram Do, Florian Kruse, Karl Kunisch New Trends in Parameter Identification for Mathematical

925 views • 49 slides
Regularization of Inverse Problems Yoeri Boink *, Srirang Manohar , Leonie Zeune *, Leon

Regularization of Inverse Problems Yoeri Boink *, Srirang Manohar , Leonie Zeune *, Leon

Deep Inversion, Autoencoders for Learned Regularization of Inverse Problems Yoeri Boink *, Srirang Manohar , Leonie Zeune *, Leon Terstappen , Stephan van Gils*, Christoph Brune* Biomedical Photonic Imaging Medical Cell Biophysics

759 views • 53 slides
A Double Regularization Approach for Inverse Problems with Noisy Data and Inexact Operator Ismael

A Double Regularization Approach for Inverse Problems with Noisy Data and Inexact Operator Ismael

A Double Regularization Approach for Inverse Problems with Noisy Data and Inexact Operator Ismael Rodrigo Bleyer Prof. Dr. Ronny Ramlau Johannes Kepler Universit at - Linz Florian opolis - September, 2011. supported by Doctoral Program

599 views • 42 slides
Regularization via Spectral Filtering Lorenzo Rosasco MIT, 9.520 Class 7 L. Rosasco

Regularization via Spectral Filtering Lorenzo Rosasco MIT, 9.520 Class 7 L. Rosasco

Regularization via Spectral Filtering Lorenzo Rosasco MIT, 9.520 Class 7 L. Rosasco Regularization via Spectral Filtering About this class Goal To discuss how a class of regularization methods originally designed for solving ill-posed inverse

889 views • 48 slides
Regularization Overview Regularization Overview Problems &amp; Multicollinearity We will

Regularization Overview Regularization Overview Problems &amp; Multicollinearity We will

Regularization Overview Regularization Overview Problems &amp; Multicollinearity We will discuss three popular methods for obtaining better estimates of the linear model coefficients Regularization Techniques Principal

305 views • 12 slides
Regularization Theory Nicolas Rougon Institut Mines-Tlcom / Tlcom SudParis ARTEMIS

Regularization Theory Nicolas Rougon Institut Mines-Tlcom / Tlcom SudParis ARTEMIS

Direct &amp; Inverse problems Restoring well-posedness Regularization Regularization Theory Nicolas Rougon Institut Mines-Tlcom / Tlcom SudParis ARTEMIS Department; CNRS UMR 8145 nicolas.rougon@telecom-sudparis.eu May 10, 2020

1.03k views • 64 slides
Banach algebras of convolution type operators with PSO data on weighted Lebesgue spaces Yuri

Banach algebras of convolution type operators with PSO data on weighted Lebesgue spaces Yuri

Spaces PSO Banach algebra A p , w Central algebra Localization Local spectra Main results Banach algebras of convolution type operators with PSO data on weighted Lebesgue spaces Yuri Karlovich Universidad Autnoma del Estado de Morelos,

590 views • 21 slides
Regularization Regularization is a general approach to add a complexity parameter to a

Regularization Regularization is a general approach to add a complexity parameter to a

Regularization Regularization is a general approach to add a complexity parameter to a learning algorithm. Requires that the model parameters be continuous. (i.e., Regression OK, IAML: Regularization and Ridge Regression Decision trees

204 views • 3 slides
10. Regularization More on tradeoffs Regularization Effect of using different norms

10. Regularization More on tradeoffs Regularization Effect of using different norms

CS/ECE/ISyE 524 Introduction to Optimization Spring 201718 10. Regularization More on tradeoffs Regularization Effect of using different norms Example: hovercraft revisited Laurent Lessard (www.laurentlessard.com) Review of

407 views • 21 slides
Vision and Sound Computer Vision Fall 2018 Columbia University Single-modality video

Vision and Sound Computer Vision Fall 2018 Columbia University Single-modality video

Vision and Sound Computer Vision Fall 2018 Columbia University Single-modality video representations Vision Hearing 1D Convolution 3D Convolution 1D Convolution 3D Convolution 1D Convolution 3D Convolution Slide credit: Andrew Owens

1.07k views • 79 slides
Fighting ambiguity of inverse problems in seismic imaging Jrg Schleicher University of

Fighting ambiguity of inverse problems in seismic imaging Jrg Schleicher University of

Introduction Reflector-oriented regularization in slope tomography Decomposition of sensitivity kernels in full-waveform inversion Conclusions Fighting ambiguity of inverse problems in seismic imaging Jrg Schleicher University of Campinas

1.3k views • 95 slides
Space-variant Generalized Gaussian Regularization for Image Restoration Alessandro Lanza, Serena

Space-variant Generalized Gaussian Regularization for Image Restoration Alessandro Lanza, Serena

Space-variant Generalized Gaussian Regularization for Image Restoration Alessandro Lanza, Serena Morigi, Monica Pragliola, Fiorella Sgallari University of Bologna Computational Methods for Inverse Problems in Imaging Workshop July 16-18,

445 views • 26 slides
Manifold Regularization Lorenzo Rosasco MIT, 9.520 L. Rosasco Manifold Regularization About

Manifold Regularization Lorenzo Rosasco MIT, 9.520 L. Rosasco Manifold Regularization About

Manifold Regularization Lorenzo Rosasco MIT, 9.520 L. Rosasco Manifold Regularization About this class Goal To analyze the limits of learning from examples in high dimensional spaces. To introduce the semi-supervised setting and the use of

437 views • 31 slides
IllPosed Inverse Problems in Image Processing Introduction, Structured matrices, Spectral

IllPosed Inverse Problems in Image Processing Introduction, Structured matrices, Spectral

IllPosed Inverse Problems in Image Processing Introduction, Structured matrices, Spectral filtering, Regularization, Noise revealing a 1 , M. Ple singer 2 , Z. Strako s 3 I. Hn etynkov hnetynko@karlin.mff.cuni.cz ,

616 views • 58 slides

More recommend