non local sparse models for image restoration
play

Non-local Sparse Models for Image Restoration Julien Mairal 1 Francis - PowerPoint PPT Presentation

Non-local Sparse Models for Image Restoration Julien Mairal 1 Francis Bach 1 Jean Ponce 2 Guillermo Sapiro 3 Andrew Zisserman 4 1 INRIA - WILLOW 2 Ecole Normale Sup erieure 3 University of Minnesota 4 Oxford University MSR-INRIA workshop,


  1. Non-local Sparse Models for Image Restoration Julien Mairal 1 Francis Bach 1 Jean Ponce 2 Guillermo Sapiro 3 Andrew Zisserman 4 1 INRIA - WILLOW 2 Ecole Normale Sup´ erieure 3 University of Minnesota 4 Oxford University MSR-INRIA workshop, January 25th 2010 Julien Mairal Non-local Sparse Models for Image Restoration 1/33

  2. What this talk is about Exploiting self-similarities in images and learned sparse representations. A fast online algorithm for learning dictionaries and factorizing matrices in general. Various formulations for image and video processing, leading to state-of-the-art results in image denoising and demosaicking. Julien Mairal Non-local Sparse Models for Image Restoration 2/33

  3. The Image Denoising Problem y x orig + w = ���� ���� ���� noise measurements original image Julien Mairal Non-local Sparse Models for Image Restoration 3/33

  4. Sparse representations for image restoration y = x orig + w ���� ���� ���� noise measurements original image Energy minimization problem || y − x || 2 E ( x ) = + ψ ( x ) 2 � �� � ���� data fitting term relation to image model Some classical priors Smoothness λ ||L x || 2 2 Total variation λ ||∇ x || 2 1 Wavelet sparsity λ || Wx || 1 . . . Julien Mairal Non-local Sparse Models for Image Restoration 4/33

  5. What is a Sparse Linear Model? Let x in R m be a signal. Let D = [ d 1 , . . . , d p ] ∈ R m × p be a set of normalized “basis vectors”. We call it dictionary . D is “adapted” to x if it can represent it with a few basis vectors—that is, there exists a sparse vector α in R p such that x ≈ D α . We call α the sparse code .   α [1]     α [2]    d 1    x ≈ d 2 · · · d p .    .  .   α [ p ] � �� � � �� � x ∈ R m D ∈ R m × p � �� � α ∈ R p , sparse Julien Mairal Non-local Sparse Models for Image Restoration 5/33

  6. The Sparse Decomposition Problem 1 2 || x − D α || 2 min + λψ ( α ) 2 α ∈ R p � �� � � �� � sparsity-inducing data fitting term regularization ψ induces sparsity in α . It can be △ the ℓ 0 “pseudo-norm”. || α || 0 = # { i s.t. α [ i ] � = 0 } (NP-hard) = � p △ the ℓ 1 norm. || α || 1 i =1 | α [ i ] | (convex) . . . This is a selection problem. Julien Mairal Non-local Sparse Models for Image Restoration 6/33

  7. Sparse representations for image restoration Designed dictionaries [Haar, 1910], [Zweig, Morlet, Grossman ∼ 70s], [Meyer, Mallat, Daubechies, Coifman, Donoho, Candes ∼ 80s-today]. . . (see [Mallat, 1999]) Wavelets, Curvelets, Wedgelets, Bandlets, . . . lets Learned dictionaries of patches [Olshausen and Field, 1997], [Engan et al., 1999], [Lewicki and Sejnowski, 2000], [Aharon et al., 2006] 1 � 2 || x i − D α i || 2 min + λψ ( α i ) 2 α i , D ∈C � �� � i � �� � sparsity reconstruction ψ ( α ) = || α || 0 (“ ℓ 0 pseudo-norm”) ψ ( α ) = || α || 1 ( ℓ 1 norm) Julien Mairal Non-local Sparse Models for Image Restoration 7/33

  8. Sparse representations for image restoration Solving the denoising problem [Elad and Aharon, 2006] Extract all overlapping 8 × 8 patches y i . Solve a matrix factorization problem: n 1 � 2 || y i − D α i || 2 min + λψ ( α i ) , 2 α i , D ∈C � �� � i =1 � �� � sparsity reconstruction with n > 100 , 000 Average the reconstruction of each patch. Julien Mairal Non-local Sparse Models for Image Restoration 8/33

  9. Sparse representations for image restoration K-SVD: [Elad and Aharon, 2006] Dictionary trained on a noisy version of the image boat. Julien Mairal Non-local Sparse Models for Image Restoration 9/33

  10. Sparse representations for image restoration Inpainting, [Mairal, Sapiro, and Elad, 2008b] Julien Mairal Non-local Sparse Models for Image Restoration 10/33

  11. Sparse representations for image restoration Inpainting, [Mairal, Elad, and Sapiro, 2008a] Julien Mairal Non-local Sparse Models for Image Restoration 11/33

  12. Sparse representations for image restoration Inpainting, [Mairal, Elad, and Sapiro, 2008a] Julien Mairal Non-local Sparse Models for Image Restoration 12/33

  13. Optimization for Dictionary Learning n 1 � 2 || x i − D α i || 2 min 2 + λ || α i || 1 α ∈ R p × n i =1 D ∈C = { D ∈ R m × p s.t. ∀ j = 1 , . . . , p , C △ || d j || 2 ≤ 1 } . Classical optimization alternates between D and α . Good results, but very slow! Julien Mairal Non-local Sparse Models for Image Restoration 13/33

  14. Optimization for Dictionary Learning n 1 � 2 || x i − D α i || 2 min 2 + λ || α i || 1 α ∈ R p × n i =1 D ∈C = { D ∈ R m × p s.t. ∀ j = 1 , . . . , p , C △ || d j || 2 ≤ 1 } . Classical optimization alternates between D and α . Good results, but very slow! [Mairal et al., 2009a]: Online learning can handle potentially infinite or dynamic datasets, be dramatically faster than batch algorithms. Julien Mairal Non-local Sparse Models for Image Restoration 13/33

  15. Optimization for Dictionary Learning n 1 � 2 || x i − D α i || 2 min 2 + λ || α i || 1 α ∈ R p × n i =1 D ∈C = { D ∈ R m × p s.t. ∀ j = 1 , . . . , p , C △ || d j || 2 ≤ 1 } . Classical optimization alternates between D and α . Good results, but very slow! [Mairal et al., 2009a]: Online learning can handle potentially infinite or dynamic datasets, be dramatically faster than batch algorithms. Try by yourself! http://www.di.ens.fr/willow/SPAMS/ Julien Mairal Non-local Sparse Models for Image Restoration 13/33

  16. Optimization for Dictionary Learning Inpainting a 12-Mpixel photograph Julien Mairal Non-local Sparse Models for Image Restoration 14/33

  17. Optimization for Dictionary Learning Inpainting a 12-Mpixel photograph Julien Mairal Non-local Sparse Models for Image Restoration 15/33

  18. Optimization for Dictionary Learning Inpainting a 12-Mpixel photograph Julien Mairal Non-local Sparse Models for Image Restoration 16/33

  19. Optimization for Dictionary Learning Inpainting a 12-Mpixel photograph Julien Mairal Non-local Sparse Models for Image Restoration 17/33

  20. Exploiting Image Self-Similarities Buades et al. [2006], Efros and Leung [1999], Dabov et al. [2007] Image pixels are well explained by a Nadaraya-Watson estimator: n K h ( y i − y j ) � ˆ x [ i ] = l =1 K h ( y i − y l ) y [ j ] , (1) � n j =1 Julien Mairal Non-local Sparse Models for Image Restoration 18/33

  21. Exploiting Image Self-Similarities Buades et al. [2006], Efros and Leung [1999], Dabov et al. [2007] Image pixels are well explained by a Nadaraya-Watson estimator: n K h ( y i − y j ) � ˆ x [ i ] = l =1 K h ( y i − y l ) y [ j ] , (1) � n j =1 Successful application to texture synthesis: Efros and Leung [1999] . . . to image denoising (Non-Local Means) : Buades et al. [2006] . . . to image demosaicking: Buades et al. [2009] Julien Mairal Non-local Sparse Models for Image Restoration 18/33

  22. Exploiting Image Self-Similarities Buades et al. [2006], Efros and Leung [1999], Dabov et al. [2007] Image pixels are well explained by a Nadaraya-Watson estimator: n K h ( y i − y j ) � ˆ x [ i ] = l =1 K h ( y i − y l ) y [ j ] , (1) � n j =1 Successful application to texture synthesis: Efros and Leung [1999] . . . to image denoising (Non-Local Means) : Buades et al. [2006] . . . to image demosaicking: Buades et al. [2009] Block-Matching with 3D filtering (BM3D) Dabov et al. [2007], Similar patches are jointly denoised with orthogonal wavelet thresholding + several (good) heuristics: = ⇒ state-of-the-art denoising results, less artefacts, higher PSNR. Julien Mairal Non-local Sparse Models for Image Restoration 18/33

  23. Non-local Sparse Image Models non-local means : stable estimator. Can fail when there are no self-similarities. sparse representations : “unique” patches also admit a sparse approximation on the learned dictionary. potentially unstable decompositions. Improving the stability of sparse decompositions is a current topic of research in statistics Bach [2008], Meinshausen and Buehlmann [2010]. Mairal et al. [2009b]: Similar patches should admit similar patterns: Sparsity vs. joint sparsity Julien Mairal Non-local Sparse Models for Image Restoration 19/33

  24. Non-local Sparse Image Models Sparsity vs. joint sparsity Joint sparsity is achieved through specific regularizerers such as k � △ || α i || 0 , || A || 0 , ∞ = (not convex, not a norm) i =1 (2) k � △ || α i || 2 . (convex norm) || A || 1 , 2 = i =1 Julien Mairal Non-local Sparse Models for Image Restoration 20/33

  25. Non-local Sparse Image Models Basic scheme for image denoising: 1 Cluster patches △ = { j = 1 , . . . , n s.t. || y i − y j || 2 S i 2 ≤ ξ } , (3) 2 Learn a dictionary with group-sparsity regularization n || A i || 1 , 2 � � || y j − D α ij || 2 min s.t. ∀ i 2 ≤ ε i (4) | S i | ( A i ) n i =1 , D ∈C i =1 j ∈ S i 3 Estimate the final image by averaging the representations Julien Mairal Non-local Sparse Models for Image Restoration 21/33

Recommend


More recommend