A memory gradient algorithm for l 2 -l 0 regularization with applications to image restoration E. Chouzenoux, A. Jezierska, J.-C. Pesquet and H. Talbot Université Paris-Est Lab. d Informatique Gaspard Monge UMR CNRS 8049 Journées GDR MOA-MSPC Optimisation et traitement d images Emilie Chouzenoux Journées GDR MOA-MSPC 1 / 32
Outline 1 General context 2 l 2 -l 0 regularization functions Existence of minimizers Epi-convergence property 3 Minimization of F δ Proposed algorithm Convergence results 4 Application to image restoration Image denoising Image deblurring Image segmentation Texture+Geometry decomposition 5 Conclusion Emilie Chouzenoux Journées GDR MOA-MSPC 2 / 32
Context Image restoration We observe data y R Q, related to the original image x R N through: y = Hx+u, H R Q N Objective: Restore the unknown original image x from H and y. y x Emilie Chouzenoux Journées GDR MOA-MSPC 3 / 32
Context Penalized optimization problem Find where ( ) min F(x) = Φ(Hx y)+λr(x), x R N Φ Fidelity to data term, related to noise R Regularization term, related to some a priori assumptions λ Regularization weight Assumption: x is sparse in a dictionary V of analysis vectors in R N F 0 (x) = Φ(Hx y)+λl 0 (V x) Emilie Chouzenoux Journées GDR MOA-MSPC 4 / 32
Context Penalized optimization problem Find where ( ) min F(x) = Φ(Hx y)+λr(x), x R N Φ Fidelity to data term, related to noise R Regularization term, related to some a priori assumptions λ Regularization weight Assumption: x is sparse in a dictionary V of analysis vectors in R N F δ (x) = Φ(Hx y)+λ C c=1 ψ δ (V c x) where ψ δ is a differentiable, non-convex approximation of the l 0 norm. Emilie Chouzenoux Journées GDR MOA-MSPC 4 / 32
Outline 1 General context 2 l 2 -l 0 regularization functions Existence of minimizers Epi-convergence property 3 Minimization of F δ Proposed algorithm Convergence results 4 Application to image restoration Image denoising Image deblurring Image segmentation Texture+Geometry decomposition 5 Conclusion Emilie Chouzenoux Journées GDR MOA-MSPC 5 / 32
Outline 1 General context 2 l 2 -l 0 regularization functions Existence of minimizers Epi-convergence property 3 Minimization of F δ Proposed algorithm Convergence results 4 Application to image restoration Image denoising Image deblurring Image segmentation Texture+Geometry decomposition 5 Conclusion Emilie Chouzenoux Journées GDR MOA-MSPC 6 / 32
l 2 -l 0 regularization functions We consider the following class of potential functions: 1 ( δ (0,+ )) ψ δ is differentiable. 2 ( δ (0,+ )) lim t ψ δ (t) = 1. 3 ( δ (0,+ )) ψ δ (t) = O(t 2 ) for small t. Examples: 1 0.8 ψ δ (t) = t2 2δ 2 +t 2 ψ δ (t) = 1 exp( t2 2δ 2 ) 0.6 0.4 0.2 0 5 2.5 0 2.5 5 t Emilie Chouzenoux Journées GDR MOA-MSPC 7 / 32
Existence of minimizers F δ (x) = Φ(Hx y)+λ C c=1 ψ δ (V c x) Difficulty: F δ is a non convex, non coercive function. Proposition Assume that and that lim Φ(x) = + x + KerH = {0} Then, for every δ > 0, F δ has a minimizer. Emilie Chouzenoux Journées GDR MOA-MSPC 8 / 32
Existence of minimizers F δ (x) = Φ(Hx y)+λ C c=1 ψ δ (V c x)+ Πx 2 Difficulty: F δ is a non convex, non coercive function. Proposition Assume that and that lim Φ(x) = + x + KerH KerΠ = {0} Then, for every δ > 0, F δ has a minimizer. Emilie Chouzenoux Journées GDR MOA-MSPC 8 / 32
Epi-convergence to the l 0 -penalized objective function Assumptions: 1 ( (δ 1,δ 2 ) (0,+ ) 2 ) δ 1 δ 2 ( t R) ψ δ1 (t) ψ δ2 (t) 0. { 0 if t = 0 2 ( t R) lim δ 0 ψ δ (t) = 1 otherwise. 3 Φ is coercive and KerH KerΠ = 0 Proposition Let (δ n ) n N be a decreasing sequence of positive real numbers converging to 0. Under the above assumptions, inff δn inff 0 as n + In addition, if for every n N, x n is a minimizer of F δn, then the sequence ( x n ) n N is bounded and all its cluster points are minimizers of F 0. Emilie Chouzenoux Journées GDR MOA-MSPC 9 / 32
Outline 1 General context 2 l 2 -l 0 regularization functions Existence of minimizers Epi-convergence property 3 Minimization of F δ Proposed algorithm Convergence results 4 Application to image restoration Image denoising Image deblurring Image segmentation Texture+Geometry decomposition 5 Conclusion Emilie Chouzenoux Journées GDR MOA-MSPC 10 / 32
Iterative minimization of F δ (x) Descent algorithm x k+1 = x k +α k d k, ( k {0,...,K}) d k : search direction satisfying g T k d k < 0 where g k F δ (x k ) Ex: Gradient, conjugate gradient, Newton, truncated Newton,... stepsize α k : approximate minimizer of f k,δ (α): α F δ (x k +αd k ) Emilie Chouzenoux Journées GDR MOA-MSPC 11 / 32
Iterative minimization of F δ (x) Descent algorithm x k+1 = x k +α k d k, ( k {0,...,K}) d k : search direction satisfying g T k d k < 0 where g k F δ (x k ) Ex: Gradient, conjugate gradient, Newton, truncated Newton,... stepsize α k : approximate minimizer of f k,δ (α): α F δ (x k +αd k ) Generalization : subspace algorithm [Zibulevsky10] r x k+1 = x k + s i,k d i k, ( k {0,...,K}) i=1 [d 1 k,...,dr k ] = D k: Set of search directions Ex: Super-memory gradient D k = [ g k,d k 1,..,d k m ] stepsize s k : approximate minimizer of f k,δ (s): s F δ (x k +D k s) Emilie Chouzenoux Journées GDR MOA-MSPC 11 / 32
Majorize-Minimize principle [Hunter04] Objective: Find ˆx Argmin x F δ (x) For all x, let Q(.,x ) a tangent majorant of F δ at x i.e., Q(x,x ) F δ (x), x, Q(x,x ) = F δ (x ) MM algorithm: j {0,...,J}, x j+1 Argmin x Q(x,x j ) Q(.,x j ) F δ x j x j+1 Emilie Chouzenoux Journées GDR MOA-MSPC 12 / 32
Quadratic tangent majorant function Assumption: For all x, there exists A(x ), definite positive, such that Q(x,x ) = F δ (x )+ F δ (x ) T (x x )+ 1 2 (x x ) T A(x )(x x ) is a quadratic tangent majorant of F δ at x. Construction of A(.) G(x) = c φ([v x w] c ) { φisc 1 onr φhasl LipschitzgradientonR A(x ) = av T V, a L Emilie Chouzenoux Journées GDR MOA-MSPC 13 / 32
Quadratic tangent majorant function Assumption: For all x, there exists A(x ), definite positive, such that Q(x,x ) = F δ (x )+ F δ (x ) T (x x )+ 1 2 (x x ) T A(x )(x x ) is a quadratic tangent majorant of F δ at x. Construction of A(.) G(x) = c φ([v x w] c ) { φisc 1 onr φhasl LipschitzgradientonR A(x ) = av T V, a L [Allain06] φisc 1 andevenonr φ(.)isconcaveonr + A(x ) = V T Diag{ω([V x w])}v ω(u) = φ(u)/u (0, ) Emilie Chouzenoux Journées GDR MOA-MSPC 13 / 32
Majorize-Minimize multivariate stepsize [Chouzenoux11] x k+1 = x k +D k s k ( k {0,...K}) D k : set of directions s k resulting from MM minimization of f k,δ (s): s F δ (x k +D k s) q k (s,s j k ) : Quadratic tangent majorant of f k,δ at s j k with Hessian: B s j = Dk TA(x k +D k s j k )D k k MM minimization in the subspace: s 0 k = 0, s j+1 k Argmin s q k (s,s j k ), ( j {0,...J 1}) s k = s J k. Emilie Chouzenoux Journées GDR MOA-MSPC 14 / 32
Proposed algorithm Majorize-Minimize Memory Gradient (MM-MG) algorithm For k = 1,...,K 1 Compute the set of directions, for example D k = [ g k,x k x k 1 ] 2 s 0 k = 0 3 j {0,...J 1}, B s j = Dk A(x k +D k s j k )D k k s j+1 k 4 s k = s J k = s j k B 1 f s j k,δ (s j k ) k 5 Update x k+1 = x k +D k s k Emilie Chouzenoux Journées GDR MOA-MSPC 15 / 32
Convergence results Assumptions ➀ Φ is coercive and KerH KerΠ = 0 ➁ The gradient of Φ is L-Lipschitzian ➂ ψ δ is even and ψ δ (.) is concave on R +. Moreover, there exists ω [0,+ ) such that ( t (0,+ )) 0 ψ δ (t) ωt. In addition, lim t 0 ψ δ (t)/t R. t 0 ➃ F δ satisfies the Lojasiewicz inequality [Attouch10a,Attouch10b]: For every x R N and every bounded neighborhood of E of x, there exist constants κ > 0, ζ > 0 and θ [0,1) such that F δ (x) κ F δ (x) F δ ( x) θ, for every x E such that F δ (x) F δ ( x) < ζ. Emilie Chouzenoux Journées GDR MOA-MSPC 16 / 32
Convergence results Theorem Under Assumptions ➀, ➁, and ➂, for all J 1, the MM-MG algorithm is such that lim k F δ(x k ) = 0. Furthermore, if Assumption ➃ is fulfilled, then The MM-MG algorithm generates a sequence converging to a critical point x of F δ. The sequence (x k ) k N has a finite length in the sense that + k=0 x k+1 x k < +. Emilie Chouzenoux Journées GDR MOA-MSPC 17 / 32
Outline 1 General context 2 l 2 -l 0 regularization functions Existence of minimizers Epi-convergence property 3 Minimization of F δ Proposed algorithm Convergence results 4 Application to image restoration Image denoising Image deblurring Image segmentation Texture+Geometry decomposition 5 Conclusion Emilie Chouzenoux Journées GDR MOA-MSPC 18 / 32
Image denoising Original image x Noisy image y 512 512 SNR = 15 db F δ (x) = 1 2 x y 2 +λ c ψ δ(v c x), ψ δ (t) = 1 exp( t 2 /2δ 2 ) MM-MG / Beck Teboulle / Half quadratic algorithms Comparison with discrete methods for ψ δ (t) = δ 2 min(t 2,δ 2 ) Emilie Chouzenoux Journées GDR MOA-MSPC 19 / 32
Results Denoised image SNR = 24.4 db Emilie Chouzenoux Journées GDR MOA-MSPC 20 / 32
Results Algorithm Time SNR (db) MM-MG 28 s 24.4 Beck-Teboulle [Beck09] 45 s 24.45 Half-Quadratic [Allain06] 97 s 24.42 Graph Cut Swap [Boykov02] 365 s 23.98 Tree-Reweighted [Felzenszwalb10] 181 s 24.28 Belief Propagation [Kolmogorov06] 1958 s 24.28 Emilie Chouzenoux Journées GDR MOA-MSPC 20 / 32
Image deblurring Original image x Noisy blurred image y 256 256 Motion blur (9 pixels) SNR = 10 db F δ (x) = 1 2 Hx y 2 +λ c ψ δ(v c x)+τ x 2, τ = 10 10 Non convex penalty: ψ δ (t) = 1 exp( t 2 /2δ 2 ) Convex penalty: ψ δ (t) = 1+t 2 /δ 2 1 Emilie Chouzenoux Journées GDR MOA-MSPC 21 / 32
Results: Non convex penalty Restored image SNR = 18.82 db Algorithm Time SNR (db) MM-MG 29 s 18.82 B-T 32 s 18.67 H-Q 141 s 18.52 Emilie Chouzenoux Journées GDR MOA-MSPC 22 / 32
Results: Convex penalty Denoised image SNR = 17.91 db Algorithm Time SNR (db) MM-MG 194 s 17.91 B-T 900 s 17.91 H-Q 1824 s 17.91 Emilie Chouzenoux Journées GDR MOA-MSPC 23 / 32
Image segmentation Original image x 256 256 F δ (x) = 1 2 x x 2 +λ c ψ δ(v c x) with large λ and small δ Non convex penalty: ψ δ (t) = 1 exp( t 2 /2δ 2 ) Convex penalty: ψ δ (t) = 1+t 2 /δ 2 1 Emilie Chouzenoux Journées GDR MOA-MSPC 24 / 32
Results Segmentation with non convex penalty Segmentation with convex penalty Emilie Chouzenoux Journées GDR MOA-MSPC 25 / 32
Results 50th line 200 180 160 140 120 100 0 50 100 150 200 250 Initial image Non convex Convex Emilie Chouzenoux Journées GDR MOA-MSPC 25 / 32
Results 50th line 200 180 160 140 120 100 0 50 100 150 200 250 Initial image Non convex Convex Algorithm Time Non convex Convex MM-MG 15 s 5 s Beck-Teboulle 21 s 50 s Half-Quadratic 39 s 31 s Emilie Chouzenoux Journées GDR MOA-MSPC 25 / 32
Texture+Geometry decomposition Original image x Noisy image y 256 256 SNR=15 db { ˆx geometry y = ˆx+ ˇx with ˇx texture + noise ( ˆx Argmin 1 x 2 1 (x y) 2 +λ c (ψ δ(vc T x)) ) [Osher03] Non convex penalty / convex penalty Emilie Chouzenoux Journées GDR MOA-MSPC 26 / 32
Results: Non convex penalty Geometry ˆx Texture + noise ˇx MM-MG algorithm: Convergence in 91 s Emilie Chouzenoux Journées GDR MOA-MSPC 27 / 32
Results: Convex penalty Geometry ˆx Texture + noise ˇx MM-MG algorithm: Convergence in 134 s Emilie Chouzenoux Journées GDR MOA-MSPC 28 / 32
Outline 1 General context 2 l 2 -l 0 regularization functions Existence of minimizers Epi-convergence property 3 Minimization of F δ Proposed algorithm Convergence results 4 Application to image restoration Image denoising Image deblurring Image segmentation Texture+Geometry decomposition 5 Conclusion Emilie Chouzenoux Journées GDR MOA-MSPC 29 / 32
Conclusion Majorize-Minimize memory gradient algorithm for l 2 -l 0 minimization Faster methods w.r.t. combinatorial optimization techniques Simplicity of implementation Future work Constrained case Non differentiable case How to ensure the global convergence? Emilie Chouzenoux Journées GDR MOA-MSPC 30 / 32
Bibliography M. Allain, J. Idier and Y. Goussard On global and local convergence of half-quadratic algorithms IEEE Transactions on Image Processing, 15(5):1130 1142, 2006 H. Attouch, J. Bolte and B. F. Svaiter Convergence of descent methods for semi-algebraic and tame problems: proximal algorithms, forward-backward splitting, and regularized Gauss-Seidel methods http://www.optimization-online.org, 2010 H. Attouch, J. Bolte, P. Redont and A. Soubeyran Proximal alternating minimization and projection methods for nonconvex problems. An approach based on the Kurdyka- Lojasiewicz inequality Mathematics of Operations Research, 35(2):438 457, 2010 A. Beck and M. Teboulle Fast gradient-based algorithms for constrained total variation image denoising and deblurring problems IEEE Transactions on Image Processing, 18(11):2419 2434, 2009 E. Chouzenoux, J. Idier and S. Moussaoui A Majorize-Minimize strategy for subspace optimization applied to image restoration IEEE Transactions on Image Processing, 20(6):1517 1528, 2011 Emilie Chouzenoux Journées GDR MOA-MSPC 31 / 32
Thanks for you attention! Emilie Chouzenoux Journées GDR MOA-MSPC 32 / 32