2. Invariant LEAst square Re-fitting Invariant Re-fitting [Deledalle, P. & Salmon 2015] 250 � Piece-wise affine mapping y �→ ˆ x ( y ) 200 150 ˆ x ( y ) = argmin F (Φ x, y ) + λG ( x ) 100 x 50 0 � Jacobian of the estimator: -50 0 20 40 60 80 100 120 140 160 180 200 x ( y ) ) ij = ∂ ˆ x ( y ) i ( J ˆ ∂y j � Model subspace: M ˆ x ( y ) = ˆ x ( y ) + Im[ J ˆ x ( y ) ] Tangent space of the mapping N. Papadakis CLEAR 9 / 41
2. Invariant LEAst square Re-fitting Invariant Re-fitting [Deledalle, P. & Salmon 2015] 250 � Piece-wise affine mapping y �→ ˆ x ( y ) 200 150 ˆ x ( y ) = argmin F (Φ x, y ) + λG ( x ) 100 x 50 0 � Jacobian of the estimator: -50 0 20 40 60 80 100 120 140 160 180 200 x ( y ) ) ij = ∂ ˆ x ( y ) i ( J ˆ ∂y j � Model subspace: M ˆ x ( y ) = ˆ x ( y ) + Im[ J ˆ x ( y ) ] � Invariant Least square Re-fitting: R inv � Φ x − y � 2 x ( y ) = argmin ˆ Tangent space of the mapping x ∈M ˆ x ( y ) N. Papadakis CLEAR 9 / 41
2. Invariant LEAst square Re-fitting Practical Re-fitting for ℓ 1 analysis minimization 1 2 || Φ x − y || 2 + λ || Γ x || 1 x ( y ) = argmin ˆ (1) x Remark: Γ = Id is the LASSO, Γ = [ ∇ x , ∇ y ] ⊤ is the Anisotropic TV. Numerical stability issue � Piecewise constant solution [Strong & Chan 2003, Caselles et al., 2009] x ( y ) requires the support: ˆ � Re-fitting ˆ I = { i ∈ [1 , m ] , s.t. | Γˆ x | i > 0 } � But in practice, ˆ x k x is only approximated through a converging sequence ˆ x k ≈ ˆ I k ≈ ˆ x � ˆ � Unfortunately, ˆ I � Illustration for Anisotropic TV denoising ( Φ = Id ): x k x k Blurry obs. y Biased ˆ x estimate ˆ Re-fitting ˆ x Re-fitting ˆ N. Papadakis CLEAR 10 / 41
2. Invariant LEAst square Re-fitting Practical Re-fitting for ℓ 1 analysis minimization 1 2 || Φ x − y || 2 + λ || Γ x || 1 x ( y ) = argmin ˆ (1) x Remark: Γ = Id is the LASSO, Γ = [ ∇ x , ∇ y ] ⊤ is the Anisotropic TV. Proposed approach � Provided Ker Φ ∩ Ker Γ= { 0 } , one has for (1) [Vaiter et al. 2016]: � J y = ∂ ˆ x ( y ) � R inv x ( y ) = J y [ y ] with � ˆ ∂y � y � Interpretation: R inv x ( y ) is the derivative of ˆ x ( y ) in the direction of y ˆ x k by chain rule as the derivative of ˆ � Algorithm: x k ( y ) Compute ˜ in the direction of y x k converge towards R inv � Question: Does ˜ x ( y ) ? ˆ N. Papadakis CLEAR 11 / 41
2. Invariant LEAst square Re-fitting Practical Re-fitting for ℓ 1 analysis minimization 1 2 || Φ x − y || 2 + λ || Γ x || 1 x ( y ) = argmin ˆ (1) x Remark: Γ = Id is the LASSO, Γ = [ ∇ x , ∇ y ] ⊤ is the Anisotropic TV. Proposed approach � Provided Ker Φ ∩ Ker Γ= { 0 } , one has for (1) [Vaiter et al. 2016]: � J y = ∂ ˆ x ( y ) � R inv x ( y ) = J y [ y ] with � ˆ ∂y � y � Interpretation: R inv x ( y ) is the derivative of ˆ x ( y ) in the direction of y ˆ x k by chain rule as the derivative of ˆ � Algorithm: x k ( y ) Compute ˜ in the direction of y x k converge towards R inv � Question: Does ˜ x ( y ) ? ˆ yes, at least for the ADMM or the Chambolle-Pock sequences N. Papadakis CLEAR 11 / 41
2. Invariant LEAst square Re-fitting Implementation for Anisotropic TV 1 2 || y − x || 2 + λ || Γ x || 1 ˆ x ( y ) = argmin x � Primal-dual implementation [Chambolle and Pock 2011]: � z k + σ Γ x k � z k +1 = Proj B λ x k + τ ( y − Γ ⊤ ( z k +1 )) x k +1 = 1+ τ � z i if | z i | � λ � Projection: Proj B λ ( z ) i = λ sign( z i ) otherwise N. Papadakis CLEAR 12 / 41
2. Invariant LEAst square Re-fitting Implementation for Anisotropic TV 1 2 || y − x || 2 + λ || Γ x || 1 ˆ x ( y ) = argmin x � Primal-dual implementation [Chambolle and Pock 2011]: � z k + σ Γ x k � z k +1 = Proj B λ � x k � z k + σ Γ˜ z k +1 ˜ = P z k + σ Γ x k ˜ x k + τ ( y − Γ ⊤ ( z k +1 )) x k +1 = 1+ τ x k + τ ( y − Γ ⊤ ( ˜ z k +1 )) ˜ x k +1 ˜ = 1+ τ � z i if | z i | � λ � Projection: Proj B λ ( z ) i = λ sign( z i ) otherwise � Id if | z i | � λ + β P z = 0 otherwise N. Papadakis CLEAR 12 / 41
2. Invariant LEAst square Re-fitting Implementation for Anisotropic TV 1 2 || y − x || 2 + λ || Γ x || 1 ˆ x ( y ) = argmin x � Primal-dual implementation [Chambolle and Pock 2011]: � z k + σ Γ x k � z k +1 = Proj B λ � x k � z k + σ Γ˜ z k +1 ˜ = P z k + σ Γ x k ˜ x k + τ ( y − Γ ⊤ ( z k +1 )) x k +1 = 1+ τ x k + τ ( y − Γ ⊤ ( ˜ z k +1 )) ˜ x k +1 ˜ = 1+ τ � z i if | z i | � λ � Projection: Proj B λ ( z ) i = λ sign( z i ) otherwise � Id if | z i | � λ + β P z = 0 otherwise x k converges to the re-fitting R inv Theorem: The sequence ˜ x ( y ) of ˆ x ( y ) , ˆ ∀ β > 0 s.t. β < σ | Γˆ x ( y ) | i , ∀ i ∈ I N. Papadakis CLEAR 12 / 41
2. Invariant LEAst square Re-fitting Implementation for Anisotropic TV 1 2 || y − x || 2 + λ || Γ x || 1 ˆ x ( y ) = argmin x � Primal-dual implementation [Chambolle and Pock 2011]: � z k + σ Γ x k � z k +1 = Proj B λ � x k � z k + σ Γ˜ z k +1 ˜ = P z k + σ Γ x k ˜ x k + τ ( y − Γ ⊤ ( z k +1 )) x k +1 = 1+ τ x k + τ ( y − Γ ⊤ ( ˜ z k +1 )) ˜ x k +1 ˜ = 1+ τ � z i if | z i | � λ � Projection: Proj B λ ( z ) i = λ sign( z i ) otherwise � Id if | z i | � λ + β P z = 0 otherwise x k converges to the re-fitting R inv Theorem: The sequence ˜ x ( y ) of ˆ x ( y ) , ˆ ∀ β > 0 s.t. β < σ | Γˆ x ( y ) | i , ∀ i ∈ I Complexity: twice that of the Chambolle-Pock algorithm. N. Papadakis CLEAR 12 / 41
2. Invariant LEAst square Re-fitting Anisotropic TV: illustration R inv y ˆ x ( y ) x ( y ) ˆ N. Papadakis CLEAR 13 / 41
2. Invariant LEAst square Re-fitting Anisotropic TV: Bias-variance trade-off PSNR: 22 . 45 , SSIM: 0 . 416 PSNR: 24 . 80 , SSIM: 0 . 545 PSNR: 27 . 00 , SSIM: 0 . 807 240 Original ˆ x ( y ) Re- fi tted R ˆ x ( y ) 220 Optimum original Optimum re- fi tted 200 Sub-optimum original Quadratic cost 180 Sub-optimum re- fi tted 160 140 120 100 80 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 PSNR: 28 . 89 , SSIM: 0 . 809 PSNR: 28 . 28 , SSIM: 0 . 823 Regularization parameter λ Anisotropic TV CLEAR N. Papadakis CLEAR 14 / 41
2. Invariant LEAst square Re-fitting Anisotropic TV: Bias-variance trade-off PSNR: 22 . 84 , SSIM: 0 . 312 PSNR: 35 . 99 , SSIM: 0 . 938 PSNR: 23 . 96 , SSIM: 0 . 694 PSNR: 38 . 22 , SSIM: 0 . 935 PSNR: 46 . 42 , SSIM: 0 . 986 Anisotropic TV CLEAR N. Papadakis CLEAR 15 / 41
2. Invariant LEAst square Re-fitting Anisotropic TV: Bias-variance trade-off PSNR: 22 . 84 , SSIM: 0 . 312 PSNR: 35 . 99 , SSIM: 0 . 938 PSNR: 23 . 96 , SSIM: 0 . 694 PSNR: 38 . 22 , SSIM: 0 . 935 PSNR: 46 . 42 , SSIM: 0 . 986 Anisotropic TV CLEAR N. Papadakis CLEAR 15 / 41
2. Invariant LEAst square Re-fitting Example: Anisotropic TGV x 0 y R inv ˆ x ( y ) x ( y ) ˆ N. Papadakis CLEAR 16 / 41
2. Invariant LEAst square Re-fitting Example: Isotropic TV � Restoration model for image y 1 2 || y − x || 2 + λ ||∇ x || 1 , 2 x ( y ) = argmin ˆ x � Model subspace of isotropic TV is the same than anisotropic TV: signals whose gradients share their support with ∇ ˆ x ( y ) N. Papadakis CLEAR 17 / 41
2. Invariant LEAst square Re-fitting Example: Isotropic TV � Restoration model for image y 1 2 || y − x || 2 + λ ||∇ x || 1 , 2 x ( y ) = argmin ˆ x � Model subspace of isotropic TV is the same than anisotropic TV: signals whose gradients share their support with ∇ ˆ x ( y ) R inv Noise free Noisy data y x ( y ) ˆ x ( y ) ˆ Non sparse support: noise is re-injected Illustration done with an ugly standard (i.e. non Condat and non Chambolle-Pock) discretization of isotropic TV N. Papadakis CLEAR 17 / 41
2. Invariant LEAst square Re-fitting Limitations Model subspace � Only captures linear invariances w.r.t. small perturbations of y Jacobian matrix � Captures desirable covariant relationships between the entries of y and the entries of ˆ x ( y ) that should be preserved [Deledalle, P., Salmon and Vaiter, 2017, 2019] N. Papadakis CLEAR 18 / 41
3. Covariant LEAst-square Re-fitting Introduction to Re-fitting Invariant LEAst square Re-fitting Covariant LEAst-square Re-fitting Practical considerations and experiments Conclusions N. Papadakis CLEAR
3. Covariant LEAst-square Re-fitting Least-square Re-fitting General problem ˆ x ( y ) = argmin F (Φ x, y ) + λG ( x ) x � Φ linear operator, F and G convex N. Papadakis CLEAR 19 / 41
3. Covariant LEAst-square Re-fitting Least-square Re-fitting General problem ˆ x ( y ) = argmin F (Φ x, y ) + λG ( x ) x � Φ linear operator, F and G convex Desirable properties of Re-fitting operator h ∈ H ˆ x iff 1 h ∈ M ˆ x ( y ) 2 Affine map: h ( y ) = Ay + b 3 Preservation of covariants: J h ( y ) = ρJ ˆ x ( y ) 4 Coherence: h (Φˆ x ( y )) = ˆ x ( y ) N. Papadakis CLEAR 19 / 41
3. Covariant LEAst-square Re-fitting Least-square Re-fitting General problem ˆ x ( y ) = argmin F (Φ x, y ) + λG ( x ) x � Φ linear operator, F and G convex Desirable properties of Re-fitting operator h ∈ H ˆ x iff 1 h ∈ M ˆ x ( y ) 2 Affine map: h ( y ) = Ay + b 3 Preservation of covariants: J h ( y ) = ρJ ˆ x ( y ) 4 Coherence: h (Φˆ x ( y )) = ˆ x ( y ) Covariant LEAst-square Re-fitting 1 R cov 2 || Φ x − y || 2 x ( y ) = argmin ˆ x ∈H ˆ x N. Papadakis CLEAR 19 / 41
3. Covariant LEAst-square Re-fitting Covariant LEAst-square Re-fitting Proposition The covariant Re-fitting has an explicit formulation 1 R cov 2 || Φ x − y || 2 x ( y ) = ˆ x ( y ) + ρJ ( y − Φˆ x ( y )) = argmin ˆ x ∈H ˆ x where for δ = y − Φˆ x ( y ) : � � Φ Jδ,δ � if Φ Jδ � = 0 || Φ Jδ || 2 ρ = 1 otherwise N. Papadakis CLEAR 20 / 41
3. Covariant LEAst-square Re-fitting Covariant LEAst-square Re-fitting Proposition The covariant Re-fitting has an explicit formulation 1 R cov 2 || Φ x − y || 2 x ( y ) = ˆ x ( y ) + ρJ ( y − Φˆ x ( y )) = argmin ˆ x ∈H ˆ x where for δ = y − Φˆ x ( y ) : � � Φ Jδ,δ � if Φ Jδ � = 0 || Φ Jδ || 2 ρ = 1 otherwise Properties � If Φ J is an orthogonal projector, ρ = 1 and Φ R cov x ( y ) = Φ R inv x ( y ) ˆ ˆ N. Papadakis CLEAR 20 / 41
3. Covariant LEAst-square Re-fitting Covariant LEAst-square Re-fitting Proposition The covariant Re-fitting has an explicit formulation 1 R cov 2 || Φ x − y || 2 x ( y ) = ˆ x ( y ) + ρJ ( y − Φˆ x ( y )) = argmin ˆ x ∈H ˆ x where for δ = y − Φˆ x ( y ) : � � Φ Jδ,δ � if Φ Jδ � = 0 || Φ Jδ || 2 ρ = 1 otherwise Properties � If Φ J is an orthogonal projector, ρ = 1 and Φ R cov x ( y ) = Φ R inv x ( y ) ˆ ˆ � If F convex, G convex and 1 -homogenous and x ( y ) = argmin ˆ F (Φ x − y ) + G ( x ) , x then J Φˆ x ( y ) = ˆ x ( y ) a.e. so that: R cov x ( y ) = (1 − ρ )ˆ x ( y ) + ρJy ˆ N. Papadakis CLEAR 20 / 41
3. Covariant LEAst-square Re-fitting Covariant LEAst-square Re-fitting Proposition The covariant Re-fitting has an explicit formulation 1 R cov 2 || Φ x − y || 2 x ( y ) = ˆ x ( y ) + ρJ ( y − Φˆ x ( y )) = argmin ˆ x ∈H ˆ x where for δ = y − Φˆ x ( y ) : � � Φ Jδ,δ � if Φ Jδ � = 0 || Φ Jδ || 2 ρ = 1 otherwise Properties � If Φ J is an orthogonal projector, ρ = 1 and Φ R cov x ( y ) = Φ R inv x ( y ) ˆ ˆ � If F convex, G convex and 1 -homogenous and x ( y ) = argmin ˆ F (Φ x − y ) + G ( x ) , x then J Φˆ x ( y ) = ˆ x ( y ) a.e. so that: R cov x ( y ) = (1 − ρ )ˆ x ( y ) + ρJy ˆ N. Papadakis CLEAR 20 / 41
3. Covariant LEAst-square Re-fitting Statistical interpretation Theorem (Bias reduction) If Φ J is an orthogonal projector or ρ satisfies technical conditions || Φ( E [ D ˆ x ( Y )] − x 0 ) || 2 � || Φ( E [ T ˆ x ( Y )] − x 0 ) || 2 N. Papadakis CLEAR 21 / 41
3. Covariant LEAst-square Re-fitting Example: Isotropic TV Noise free Noisy data y R inv R cov x ( y ) ˆ x ( y ) x ( y ) ˆ ˆ N. Papadakis CLEAR 22 / 41
3. Covariant LEAst-square Re-fitting Why not iterating as Boosting approaches? � Differentiable estimator w.r.t y : | 2 + λ | x 0 = ˆ ˜ x ( y ) = argmin | | x − y | |∇ x | | 1 , 2 x N. Papadakis CLEAR 23 / 41
3. Covariant LEAst-square Re-fitting Why not iterating as Boosting approaches? � Differentiable estimator w.r.t y : | 2 + λ | x 0 = ˆ ˜ x ( y ) = argmin | | x − y | |∇ x | | 1 , 2 x � Bregman iterations : | 2 + λD | x k +1 = argmin ˜ | | x − y | | 1 , 2 ( x, ˜ x k ) |∇ . | x N. Papadakis CLEAR 23 / 41
3. Covariant LEAst-square Re-fitting Why not iterating as Boosting approaches? � Differentiable estimator w.r.t y : | 2 + λ | x 0 = ˆ ˜ x ( y ) = argmin | | x − y | |∇ x | | 1 , 2 x � Bregman iterations : | 2 + λD | x k +1 = argmin ˜ | | x − y | | 1 , 2 ( x, ˜ x k ) |∇ . | x N. Papadakis CLEAR 23 / 41
3. Covariant LEAst-square Re-fitting Why not iterating as Boosting approaches? � Differentiable estimator w.r.t y : | 2 + λ | x 0 = ˆ ˜ x ( y ) = argmin | | x − y | |∇ x | | 1 , 2 x � Bregman iterations : | 2 + λD | x k +1 = argmin ˜ | | x − y | | 1 , 2 ( x, ˜ x k ) |∇ . | x N. Papadakis CLEAR 23 / 41
3. Covariant LEAst-square Re-fitting Why not iterating as Boosting approaches? � Differentiable estimator w.r.t y : | 2 + λ | x 0 = ˆ ˜ x ( y ) = argmin | | x − y | |∇ x | | 1 , 2 x � Bregman iterations : | 2 + λD | x k +1 = argmin ˜ | | x − y | | 1 , 2 ( x, ˜ x k ) |∇ . | x � Convergence: x k → y ˜ N. Papadakis CLEAR 23 / 41
3. Covariant LEAst-square Re-fitting Why not iterating as Boosting approaches? � Differentiable estimator w.r.t y : | 2 + λ | x 0 = ˆ ˜ x ( y ) = argmin | | x − y | |∇ x | | 1 , 2 x � Covariant iterations : x k +1 ( y ) = ˜ x k ( y ) + ρJ ( y − Φ˜ x k ( y )) ˜ N. Papadakis CLEAR 23 / 41
3. Covariant LEAst-square Re-fitting Why not iterating as Boosting approaches? � Differentiable estimator w.r.t y : | 2 + λ | x 0 = ˆ ˜ x ( y ) = argmin | | x − y | |∇ x | | 1 , 2 x � Covariant iterations : x k +1 ( y ) = ˜ x k ( y ) + ρJ ( y − Φ˜ x k ( y )) ˜ N. Papadakis CLEAR 23 / 41
3. Covariant LEAst-square Re-fitting Why not iterating as Boosting approaches? � Differentiable estimator w.r.t y : | 2 + λ | x 0 = ˆ ˜ x ( y ) = argmin | | x − y | |∇ x | | 1 , 2 x � Covariant iterations : x k +1 ( y ) = ˜ x k ( y ) + ρJ ( y − Φ˜ x k ( y )) ˜ � Convergence: x k ( y ) → R inv ˜ x ( y ) ˆ N. Papadakis CLEAR 23 / 41
4. Practical considerations and experiments Introduction to Re-fitting Invariant LEAst square Re-fitting Covariant LEAst-square Re-fitting Practical considerations and experiments Conclusions N. Papadakis CLEAR
4. Practical considerations and experiments How computing the covariant Re-fitting? � Explicit expression: R cov x ( y ) = ˆ x ( y ) + ρJδ ˆ with J = ∂ ˆ x ( y ) ∂y , δ = y − Φˆ x ( y ) and � � Φ Jδ,δ � if Φ Jδ � = 0 ρ = || Φ Jδ || 2 1 otherwise Main issue � Being able to compute Jδ N. Papadakis CLEAR 24 / 41
4. Practical considerations and experiments Application of the Jacobian matrix to a vector Algorithmic differentiation � Iterative algorithm to obtain ˆ x ( y ) : x k +1 = ψ ( x k , y ) � Differentiation in the direction δ : � x k +1 = ψ ( x k , y ) x k + ∂ 2 ψ ( x k , y ) δ x k +1 = ∂ 1 ψ ( x k , y )˜ ˜ � J x k ( y ) δ = ˜ x k � Joint estimation of x k and J x k ( y ) δ � Double the computational cost N. Papadakis CLEAR 25 / 41
4. Practical considerations and experiments Application of the Jacobian matrix to a vector Finite difference based differentiation � ˆ x ( y ) can be any black box algorithm � Directional derivative w.r.t to direction δ : x ( y ) δ ≈ ˆ x ( y + εδ ) − ˆ x ( y ) J ˆ ε � Need to apply twice the algorithm N. Papadakis CLEAR 26 / 41
4. Practical considerations and experiments Computation of the Re-fitting Covariant LEAst-square Re-fitting x ( y ) and ρ = � Jδ, δ � R cov x ( y ) = ˆ x ( y ) + ρJδ, with δ = y − Φˆ ˆ || Jδ || 2 2 Two-steps with any denoising algorithm 1 Apply algorithm to y to get ˆ x ( y ) and set δ = y − Φˆ x ( y ) 2 Compute Jδ (with algorithmic or finite difference based differentiation) N. Papadakis CLEAR 27 / 41
4. Practical considerations and experiments Computation of the Re-fitting Covariant LEAst-square Re-fitting x ( y ) and ρ = � Jδ, δ � R cov x ( y ) = ˆ x ( y ) + ρJδ, with δ = y − Φˆ ˆ || Jδ || 2 2 Two-steps with any denoising algorithm 1 Apply algorithm to y to get ˆ x ( y ) and set δ = y − Φˆ x ( y ) 2 Compute Jδ (with algorithmic or finite difference based differentiation) One-step for absolutely 1 -homogeneous regularizer Re-fitting simplifies to R cov x ( y ) = (1 − ρ )ˆ x ( y ) + ρJy ˆ 1 Estimate jointly ˆ x ( y ) and Jy with the differentiated algorithm N. Papadakis CLEAR 27 / 41
4. Practical considerations and experiments 1 step Implementation: Anisotropic TV � Model with 1 -homogeneous regularizer: 1 2 || y − x || 2 + λ ||∇ x || 1 x ( y ) = argmin ˆ x � Primal-dual implementation [Chambolle and Pock 2011]: � z k + σ ∇ x k � z k +1 = Proj B λ � z k + σ ∇ ˜ x k � z k +1 ˜ = P z k + σ ∇ x k ˜ x k + τ ( y +div ( z k +1 )) x k +1 = 1+ τ x k + τ ( y +div ( ˜ z k +1 )) ˜ x k +1 ˜ = 1+ τ � Projection: � z i if | z i | � λ Proj B λ ( z ) i = λ sign( z i ) otherwise � Id if | z i | � λ + β P z = 0 otherwise � x k → ˆ x k = J x k y → J ˆ x ( y ) and ˜ x y � J is an orthogonal projector: R cov x ( y ) = R inv x ( y ) = J ˆ x y ˆ ˆ N. Papadakis CLEAR 28 / 41
4. Practical considerations and experiments 1 step Implementation: Isotropic TV � Model with 1 -homogeneous regularizer: 1 2 || y − x || 2 + λ ||∇ x || 1 , 2 x ( y ) = argmin ˆ x � Primal-dual implementation [Chambolle and Pock 2011]: � z k + σ ∇ x k � z k +1 = Proj B λ x k + τ ( y +div ( z k +1 )) x k +1 = 1+ τ � z i � Projection: if || z i || 2 � λ Proj B λ ( z ) i = z i λ otherwise || z i || 2 N. Papadakis CLEAR 29 / 41
4. Practical considerations and experiments 1 step Implementation: Isotropic TV � Model with 1 -homogeneous regularizer: 1 2 || y − x || 2 + λ ||∇ x || 1 , 2 x ( y ) = argmin ˆ x � Primal-dual implementation [Chambolle and Pock 2011]: � z k + σ ∇ x k � z k +1 = Proj B λ � z k + σ ∇ ˜ x k � z k +1 ˜ = P z k + σ ∇ x k ˜ x k + τ ( y +div ( z k +1 )) x k +1 = 1+ τ x k + τ ( y +div ( ˜ z k +1 )) ˜ x k +1 ˜ = 1+ τ � z i � Projection: if || z i || 2 � λ Proj B λ ( z ) i = z i λ otherwise � Id || z i || 2 if | | z i | | � λ + β � � P z = Id − z i z ⊤ λ otherwise i || z i || 2 || z i || 2 2 � x k → ˆ x k = J x k y → ˜ x ( y ) and ˜ x N. Papadakis CLEAR 29 / 41
4. Practical considerations and experiments 1 step Implementation: Isotropic TV � Model with 1 -homogeneous regularizer: 1 2 || y − x || 2 + λ ||∇ x || 1 , 2 x ( y ) = argmin ˆ x � Primal-dual implementation [Chambolle and Pock 2011]: � z k + σ ∇ x k � z k +1 = Proj B λ � z k + σ ∇ ˜ x k � z k +1 ˜ = P z k + σ ∇ x k ˜ x k + τ ( y +div ( z k +1 )) x k +1 = 1+ τ x k + τ ( y +div ( ˜ z k +1 )) ˜ x k +1 ˜ = 1+ τ � z i � Projection: if || z i || 2 � λ Proj B λ ( z ) i = z i λ otherwise � Id || z i || 2 if | | z i | | � λ + β � � P z = Id − z i z ⊤ λ otherwise i || z i || 2 || z i || 2 2 � x k → ˆ x k = J x k y → ˜ x ( y ) and ˜ x with ρ = � ˜ x − ˆ x,y − ˆ x � � Covariant re-fitting: R cov x ( y ) = (1 − ρ )ˆ x + ρ ˜ x, ˆ x || 2 || ˜ x − ˆ 2 N. Papadakis CLEAR 29 / 41
4. Practical considerations and experiments Inpainting with Isotropic TV R cov y ˆ x ( y ) x ( y ) ˆ Attenuated structures Residual lost structures || R cov x 0 || ˆ x ( y ) − x 0 || 2 x ( y ) − x 0 || 2 ˆ N. Papadakis CLEAR 30 / 41
4. Practical considerations and experiments Extension to chrominance [Pierre, Aujol, Deledalle, P., 2017] Noise free image Noisy image Denoised image N. Papadakis CLEAR 31 / 41
4. Practical considerations and experiments Extension to chrominance [Pierre, Aujol, Deledalle, P., 2017] Noise free image Noisy image Re-fitting N. Papadakis CLEAR 31 / 41
4. Practical considerations and experiments 2 steps Implementation: Non-Local Means � Model without 1 -homogeneous regularizer: � j w y ij y j � 2 /h 2 � w y | 2 x ( y ) i = ˆ , i,j = exp −| |P i y − P j y | � j w y ij � Differentiate NLM code � Algorithm: � Re-fitting with algorithmic differentiation: 1: Run NLM code ˆ x ( y ) and set δ = y − ˆ x ( y ) 2: Run differentiated NLM code in the direction δ to get Jδ 3: Set ρ = � Jδ,δ � || Jδ || 2 2 4: Covariant re-fitting: R cov x ( y ) = ˆ x ( y ) + ρJδ ˆ N. Papadakis CLEAR 32 / 41
4. Practical considerations and experiments Non-Local Means R cov y x ( y ) ˆ x ( y ) ˆ Attenuated structures Residual lost structures || R cov x 0 || ˆ x ( y ) − x 0 || 2 x ( y ) − x 0 || 2 ˆ N. Papadakis CLEAR 33 / 41
4. Practical considerations and experiments Non-Local Means: Bias-variance trade-off PSNR: 22 . 18 , SSIM: 0 . 397 PSNR: 25 . 07 , SSIM: 0 . 564 PSNR: 26 . 62 , SSIM: 0 . 724 Original ˆ x ( y ) Re- fi tted R ˆ x ( y ) Optimum original 200 Optimum re- fi tted Quadratic cost Sub-optimum original Sub-optimum re- fi tted 150 100 1 2 3 4 5 6 PSNR: 30 . 12 , SSIM: 0 . 815 PSNR: 29 . 20 , SSIM: 0 . 823 Filtering parameter h Non-Local Means CLEAR N. Papadakis CLEAR 34 / 41
4. Practical considerations and experiments Bias-variance trade-off: Non-Local Means PSNR: 22 . 18 , SSIM: 0 . 847 PSNR: 24 . 68 , SSIM: 0 . 914 PSNR: 24 . 64 , SSIM: 0 . 910 PSNR: 27 . 04 , SSIM: 0 . 946 PSNR: 27 . 29 , SSIM: 0 . 955 Non-Local Means CLEAR N. Papadakis CLEAR 35 / 41
4. Practical considerations and experiments Bias-variance trade-off: Non-Local Means PSNR: 22 . 18 , SSIM: 0 . 847 PSNR: 24 . 68 , SSIM: 0 . 914 PSNR: 24 . 64 , SSIM: 0 . 910 PSNR: 27 . 04 , SSIM: 0 . 946 PSNR: 27 . 29 , SSIM: 0 . 955 Non-Local Means CLEAR N. Papadakis CLEAR 35 / 41
4. Practical considerations and experiments 2 steps Implementation for Black Box algorithm � Denoising algorithm: y �→ ˆ x ( y ) � Re-fitting with finite difference: 1: δ = y − ˆ x ( y ) 2: Jδ = ˆ x ( y + εδ ) − ˆ x ( y )) ε 3: ρ = � Jδ,δ � || Jδ || 2 2 4: R cov x ( y ) = ˆ x ( y ) + ρJδ ˆ N. Papadakis CLEAR 36 / 41
4. Practical considerations and experiments BM3D [Dabov et al. 2007, Lebrun 2012] PSNR: 22 . 17 , SSIM: 0 . 528 PSNR: 25 . 00 , SSIM: 0 . 7523 PSNR: 26 . 92 , SSIM: 0 . 861 250 Original ˆ x ( y ) Re- fi tted R [ˆ x ] ( y ) Optimum original 200 Optimum re- fi tted Quadratic cost Sub-optimum original Sub-optimum re- fi tted 150 100 50 -0.5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 Filtering parameter log γ PSNR: 30 . 29 , SSIM: 0 . 920 PSNR: 29 . 41 , SSIM: 0 . 918 BM3D CLEAR N. Papadakis CLEAR 37 / 41
4. Practical considerations and experiments DDID [Knaus & Zwicker 2013] PSNR: 22 . 16 , SSIM: 0 . 452 PSNR: 26 . 33 , SSIM: 0 . 716 PSNR: 26 . 60 , SSIM: 0 . 721 200 180 160 Quadratic cost 140 120 100 80 60 40 -2 -1 0 1 2 3 4 5 6 PSNR: 31 . 02 , SSIM: 0 . 858 PSNR: 29 . 91 , SSIM: 0 . 845 Filtering parameter log γ DDID CLEAR N. Papadakis CLEAR 38 / 41
4. Practical considerations and experiments DnCNN [Zhang et al., 2017] Residual Network learning noise to remove Noise level 25 DnCNN N. Papadakis CLEAR 39 / 41
4. Practical considerations and experiments DnCNN [Zhang et al., 2017] Residual Network learning noise to remove Noise level 25 CLEAR N. Papadakis CLEAR 39 / 41
4. Practical considerations and experiments DnCNN [Zhang et al., 2017] Residual Network learning noise to remove Noise level 50 DnCNN N. Papadakis CLEAR 39 / 41
4. Practical considerations and experiments DnCNN [Zhang et al., 2017] Residual Network learning noise to remove Noise level 50 CLEAR N. Papadakis CLEAR 39 / 41
4. Practical considerations and experiments DnCNN [Zhang et al., 2017] Residual Network learning noise to remove Noise level 150 DnCNN N. Papadakis CLEAR 39 / 41
4. Practical considerations and experiments DnCNN [Zhang et al., 2017] Residual Network learning noise to remove Noise level 150 CLEAR N. Papadakis CLEAR 39 / 41
4. Practical considerations and experiments DnCNN [Zhang et al., 2017] Residual Network learning noise to remove Noise level 150 CLEAR � No interesting structural information to recover from noise model N. Papadakis CLEAR 39 / 41
4. Practical considerations and experiments FFDNet [Zhang et al., 2018] Network learning denoised image for Gaussian noise of variance [0; 75] Noise level 25 FFDNet N. Papadakis CLEAR 40 / 41
4. Practical considerations and experiments FFDNet [Zhang et al., 2018] Network learning denoised image for Gaussian noise of variance [0; 75] Noise level 25 CLEAR N. Papadakis CLEAR 40 / 41
4. Practical considerations and experiments FFDNet [Zhang et al., 2018] Network learning denoised image for Gaussian noise of variance [0; 75] Noise level 50 FFDNet N. Papadakis CLEAR 40 / 41
4. Practical considerations and experiments FFDNet [Zhang et al., 2018] Network learning denoised image for Gaussian noise of variance [0; 75] Noise level 50 CLEAR N. Papadakis CLEAR 40 / 41
4. Practical considerations and experiments FFDNet [Zhang et al., 2018] Network learning denoised image for Gaussian noise of variance [0; 75] Noise level 150 FFDNet N. Papadakis CLEAR 40 / 41
4. Practical considerations and experiments FFDNet [Zhang et al., 2018] Network learning denoised image for Gaussian noise of variance [0; 75] Noise level 150 CLEAR N. Papadakis CLEAR 40 / 41
5. Conclusions Introduction to Re-fitting Invariant LEAst square Re-fitting Covariant LEAst-square Re-fitting Practical considerations and experiments Conclusions N. Papadakis CLEAR
5. Conclusions Conclusions Covariant LEAst-square Re-fitting � Correct part of the bias of restoration models � No aditionnal parameter � Stability for a larger range of parameters � Double the computational cost N. Papadakis CLEAR 41 / 41
5. Conclusions Conclusions Covariant LEAst-square Re-fitting � Correct part of the bias of restoration models � No aditionnal parameter � Stability for a larger range of parameters � Double the computational cost When using re-fitting? � Differentiable estimators: no algorithm with quantization � Regularization prior adapted to data � Respect data range: oceanography, radiotherapy... N. Papadakis CLEAR 41 / 41
5. Conclusions Main related references [Brinkmann, Burger, Rasch and Sutour] Bias-reduction in variational regularization. JMIV , 2017. [Deledalle, P. and Salmon] On debiasing restoration algorithms: applications to total-variation and nonlocal-means. SSVM , 2015. [Deledalle, P., Salmon and Vaiter] CLEAR: Covariant LEAst-square Re-fitting. SIAM SIIMS , 2017. [Deledalle, P., Salmon and Vaiter] Refitting solutions with block penalties, SSVM , 2019. [Osher, Burger, Goldfarb, Xu, and Yin] An iterative regularization method for total variation-based image restoration. SIAM MMS , 2005 [Romano and Elad] Boosting of image denoising algorithms. SIAM SIIMS , 2015. [Talebi, Zhu and Milanfar] How to SAIF-ly boost denoising performance. IEEE TIP , 2013. [Vaiter, Deledalle, Peyré, Fadili and Dossal] The degrees of freedom of partly smooth regularizers. Annals of the Institute of Statistical Mathematics , 2016. N. Papadakis CLEAR
5. Conclusions Sign in TV models [Brinkmann et al. 2016] � Differentiable estimator w.r.t y : | 2 + λ | x ( y ) = argmin ˆ | | x − y | |∇ x | | 1 , 2 x N. Papadakis CLEAR
5. Conclusions Sign in TV models [Brinkmann et al. 2016] � Differentiable estimator w.r.t y : | 2 + λ | x ( y ) = argmin ˆ | | x − y | |∇ x | | 1 , 2 x � Orientation preservation: O ˆ x ( y ) = { x | ( ∇ x ) i = α i ( ∇ ˆ x ) i , ∀ i ∈ I} N. Papadakis CLEAR
5. Conclusions Sign in TV models [Brinkmann et al. 2016] � Differentiable estimator w.r.t y : | 2 + λ | x ( y ) = argmin ˆ | | x − y | |∇ x | | 1 , 2 x � Orientation preservation: O ˆ x ( y ) = { x | ( ∇ x ) i = α i ( ∇ ˆ x ) i , ∀ i ∈ I} � Infimal Convolution of Bregman divergences x ICB = argmin | 2 ˜ | | x − y | x ∈O ˆ x ( y ) N. Papadakis CLEAR
Recommend
More recommend