Image Reconstruction Tutorial Part 1: Sparse optimization and learning approaches zurica 1 and Bart Goossens 2 Aleksandra Piˇ 1 Group for Artificial Intelligence and Sparse Modelling (GAIM) 2 Image Processing and Interpretation (IPI) group TELIN, Ghent University - imec Yearly workshop FWO–WOG Turning images into value through statistical parameter estimation Ghent, Belgium, 20 September 2019
Outline Model-based iterative reconstruction algorithms 1 Sparse optimization Solution strategies: greedy methods vs. convex optimization Optimization methods in sparse image reconstruction Structured sparsity 2 Wavelet-tree sparsity Markov Random Field (MRF) priors Machine learning in image reconstruction 3 Main ideas and current trends A. Piˇ zurica and B. Goossens (UGent) Image Reconstruction Tutorial: Part 1 FWO-WOG TIVSPE 2019 2 / 45
A fairly general formulation Reconstruct a signal (image) x ∈ X from data y ∈ Y where y = T ( x ) + n X and Y are Hilbert spaces, T : X �→ Y is the forward operator and n is noise. A common model-driven approach is to minimize the negative log-likelihood L : min x ∈ X L ( T ( x ) , y ) Typically, ill-posed and leads to over-fitting. Variational regularization, also called model-based iterative reconstruction seeks to minimize a regularized objective function min x ∈ X L ( T ( x ) , y ) + τφ ( x ) φ : X �→ R ∪ {−∞ , ∞} is a regularization functional. τ ≥ 0 governs the influence of the a priori knowledge against the need to fit the data. A. Piˇ zurica and B. Goossens (UGent) Image Reconstruction Tutorial: Part 1 FWO-WOG TIVSPE 2019 3 / 45
Linear inverse problems Many image reconstruction problems can be formulated as a linear inverse problem. A noisy indirect observation y , of the original image x is then y = Ax + n Matrix A is the forward operator. x ∈ R n ; y , n ∈ R m (or x ∈ C n ; y , n ∈ C m ). Here, image pixels are stacked into vectors (raster scanning). In general, m � = n . Some examples CT: A is the system matrix modeling the X-ray transformation MRI: A is (partially sampled) Fourier operator OCT: A is the first Born approximation for the scattering Compressed sensing: A is a measurement matrix (dense or sparse) A. Piˇ zurica and B. Goossens (UGent) Image Reconstruction Tutorial: Part 1 FWO-WOG TIVSPE 2019 4 / 45
Linear inverse problems For the linear inverse problem y = Ax + n , model-based reconstruction seeks to solve: 1 2 � Ax − y � 2 min 2 + τφ ( x ) ( Tikhonov formulation) x Alternatively, � Ax − y � 2 min x φ ( x ) subject to 2 ≤ ǫ ( Morozov formulation) x � Ax − y � 2 min φ ( x ) ≤ δ ( Ivanov formulation) subject to 2 Under mild conditions, these are all equivalent [Figueiredo and Wright, 2013], and which one is more convenient is problem-dependent. A. Piˇ zurica and B. Goossens (UGent) Image Reconstruction Tutorial: Part 1 FWO-WOG TIVSPE 2019 5 / 45
Outline Model-based iterative reconstruction algorithms 1 Sparse optimization Solution strategies: greedy methods vs. convex optimization Optimization methods in sparse image reconstruction Structured sparsity 2 Wavelet-tree sparsity Markov Random Field (MRF) priors Machine learning in image reconstruction 3 Main ideas and current trends A. Piˇ zurica and B. Goossens (UGent) Image Reconstruction Tutorial: Part 1 FWO-WOG TIVSPE 2019 6 / 45
Sparse optimization A common assumption: x is sparse in a well-chosen transform domain. We refer to a wavelet representation meaning any wavelet-like multiscale representation, including curvelets and shearlets.. x = Ψ θ , θ ∈ R d , Ψ ∈ R n × d The columns of Ψ are the elements of a wavelet frame (an orthogonal basis or an overcomplete dictionary) The main results hold for learned dictionaries , trained on a set of representative examples to yield optimally sparse representation for a particular class of images. A. Piˇ zurica and B. Goossens (UGent) Image Reconstruction Tutorial: Part 1 FWO-WOG TIVSPE 2019 7 / 45
Compressed sensing with y , n ∈ R m , x ∈ R n , θ ∈ R d , Consider y = Ax + n , x = Ψ θ , m < n 1 ˆ 2 � AΨ θ − y � 2 x = Ψ ˆ θ = arg min 2 + τφ ( θ ) , ˆ θ θ θ � θ � 0 s.t. � AΨ θ − y � 2 1 2 � AΨ θ − y � 2 Commonly: min 2 ≤ ǫ or min 2 + τ � θ � 1 θ [Cand` es et al., 2006], [Donoho, 2006], [Lustig et al., 2007] A. Piˇ zurica and B. Goossens (UGent) Image Reconstruction Tutorial: Part 1 FWO-WOG TIVSPE 2019 8 / 45
Compressed sensing: recovery guarantees with y , n ∈ R m , x ∈ R n , m < n Consider y = Ax + n , x = Ψ θ , Matrix Φ = AΨ has K -restricted isometry property ( K-RIP ) with constant ǫ K < 1 if ∀ K -sparse (having only K non-zero entries) θ : ( 1 − ǫ K ) � θ � 2 2 ≤ � Φ θ � 2 2 ≤ ( 1 + ǫ K ) � θ � 2 2 Suppose matrix A ∈ R m × n is formed by subsampling a given sampling operator A ∈ R n × n . The mutual coherence between ¯ ¯ A and Ψ : µ (¯ i , j | a ⊤ A , Ψ ) = max i ψ j | If m > C µ 2 (¯ A , Ψ ) Kn log ( n ) , for some constant C > 0, then 1 2 � AΨ θ − y � 2 min 2 + τ � θ � 1 θ recovers x with high probability, given the K-RIP holds for Φ = AΨ . A. Piˇ zurica and B. Goossens (UGent) Image Reconstruction Tutorial: Part 1 FWO-WOG TIVSPE 2019 9 / 45
Analysis vs. synthesis formulation with y ∈ R m , x ∈ R n , θ ∈ R d Consider y = Ax + n , x = Ψ θ , Synthesis approach: 1 2 � AΨ θ − y � 2 min 2 + τφ ( θ ) θ or in the constrained form: � AΨ θ − y � 2 min θ φ ( θ ) 2 ≤ ǫ subject to Analysis approach: 1 2 � Ax − y � 2 min 2 + τφ ( x ) x or in the constrained form: � Ax − y � 2 min x φ ( x ) subject to 2 ≤ ǫ A. Piˇ zurica and B. Goossens (UGent) Image Reconstruction Tutorial: Part 1 FWO-WOG TIVSPE 2019 10 / 45
Analysis vs. synthesis formulation with y ∈ R m , x ∈ R n , θ ∈ R d Consider y = Ax + n , x = Ψ θ , Synthesis approach: 1 2 � AΨ θ − y � 2 min 2 + τφ ( θ ) θ or in a constrained form: � AΨ θ − y � 2 min θ φ ( θ ) 2 ≤ ǫ subject to Analysis approach: 1 2 � Ax − y � 2 min 2 + τφ ( x ) x or in a constrained form: � Ax − y � 2 min x φ ( x ) subject to 2 ≤ ǫ A. Piˇ zurica and B. Goossens (UGent) Image Reconstruction Tutorial: Part 1 FWO-WOG TIVSPE 2019 11 / 45
Analysis vs. synthesis formulation with y ∈ R m , x ∈ R n , θ ∈ R d Consider y = Ax + n , x = Ψ θ , Synthesis approach: 1 2 � AΨ θ − y � 2 min 2 + τφ ( θ ) θ or in a constrained form: � AΨ θ − y � 2 min θ φ ( θ ) subject to 2 ≤ ǫ Analysis approach that also applies to wavelet regularization 1 2 � Ax − y � 2 min 2 + τφ ( Px ) x or in a constrained form: � Ax − y � 2 min x φ ( Px ) 2 ≤ ǫ subject to A. Piˇ zurica and B. Goossens (UGent) Image Reconstruction Tutorial: Part 1 FWO-WOG TIVSPE 2019 12 / 45
Analysis vs. synthesis formulation with y ∈ R m , x ∈ R n , θ ∈ R d Consider y = Ax + n , x = Ψ θ , Synthesis approach: 1 2 � AΨ θ − y � 2 min 2 + τφ ( θ ) θ or in a constrained form: � AΨ θ − y � 2 min θ φ ( θ ) 2 ≤ ǫ subject to Analysis approach that also applies to wavelet regularization 1 2 � Ax − y � 2 min 2 + τφ ( Px ) x or in a constrained form: � Ax − y � 2 min x φ ( Px ) 2 ≤ ǫ subject to P : a wavelet transform operator or P = I (standard analysis) A. Piˇ zurica and B. Goossens (UGent) Image Reconstruction Tutorial: Part 1 FWO-WOG TIVSPE 2019 13 / 45
Outline Model-based iterative reconstruction algorithms 1 Sparse optimization Solution strategies: greedy methods vs. convex optimization Optimization methods in sparse image reconstruction Structured sparsity 2 Wavelet-tree sparsity Markov Random Field (MRF) priors Machine learning in image reconstruction 3 Main ideas and current trends A. Piˇ zurica and B. Goossens (UGent) Image Reconstruction Tutorial: Part 1 FWO-WOG TIVSPE 2019 14 / 45
Solution strategies: greedy methods vs. convex optimization Solution strategy is problem-dependent. For non-convex problems like � Ax − y � 2 x � x � 0 2 ≤ ǫ min subject to Greedy algorithms , e.g., Matching Pursuit ( MP ) [Mallat and Zhang, 1993] OMP [Tropp, 2004], CoSaMP [Needell and Tropp, 2009] IHT [Blumensath and Davies, 2009] or convex relaxation can be applied leading to: � Ax − y � 2 min x � x � 1 2 ≤ ǫ subject to 1 2 � Ax − y � 2 min 2 + τφ ( x ) x known as LASSO [Tibshirani, 1996] or BPDN [Chen et al., 2001] problem. A. Piˇ zurica and B. Goossens (UGent) Image Reconstruction Tutorial: Part 1 FWO-WOG TIVSPE 2019 15 / 45
Greedy methods: OMP OMP algorithm for solving min x � x � 0 subject to Ax = y Require: k = 1 , r ( 1 ) = y , Λ ( 0 ) = ∅ 1: repeat λ ( k ) = arg max j | A j · r ( k ) | 2: Λ ( k ) = Λ ( k − 1 ) ∪ { λ ( k ) } 3: x ( k ) = arg min x � A Λ k x − y � 2 4: y ( k ) = A Λ k x ( k ) ˆ 5: r ( k + 1 ) = r ( k ) − ˆ y ( k ) 6: k = k + 1 7: 8: until stopping criterion satisfied A j is the j -th column of A , and A Λ a sub-matrix of A with columns indicated in Λ . A. Piˇ zurica and B. Goossens (UGent) Image Reconstruction Tutorial: Part 1 FWO-WOG TIVSPE 2019 16 / 45
Recommend
More recommend