fast nonstationary preconditioned iterative methods for ill-posed
TRANSCRIPT
Fast nonstationary preconditionediterative methods for ill-posed problems,with application to image deblurring
Marco Donatelli
Dipartimento di Scienza e Alta TecnologiaUniversita dell’Insubria - Como
Join work with Martin Hanke (U. Mainz, Germany)
SLA 2014Kalamata, 8-12 September 2014
Outline
Iterative solution of ill-posed equations
A nonstationary preconditioned iteration
Application to image deblurring
Outline
Iterative solution of ill-posed equations
A nonstationary preconditioned iteration
Application to image deblurring
The model problem
Consider the solution of ill-posed equations
Tx = y , (1)
where T : X → Y is a linear operator between Hilbert spaces.Assume that problem (1) has a solution x† of minimal norm.
GoalCompute an approximation of x† starting from approximate datay δ ∈ Y, instead of the exact data y ∈ Y, with
‖y δ − y‖ ≤ δ , (2)
where δ ≥ 0 is the corresponding noise level.
Regularization methods
The solution of Tx = y δ requires some sort of regularization
◮ Tikhonov regularization: usually expensive (one linear systemto solve per approximation) and difficult optimization of theregularization parameter.
◮ Iterative regularization: cheap (only matrix-vector products)and the regularization parameter is determined on the fly.
◮ Hybrid regularization: Tikhonov regularization on Krylovsubspaces of increasing dimension (O’Leary, Nagy, Reichel,Novati, . . . ).
Iterative regularization
◮ Landweber iteration: slow convergence.
◮ CGLS: (conjugate gradient for least squares) faster, but stillnot fast enough. . . .
Preconditioned regularization
Replace the original problem Tx = y δ with
P−1Tx = P−1y δ
such that
◮ inversion of P is cheap
◮ P ≈ T but not too much (T−1 unbounded while P−1 mustbe bounded!)
Preconditioners can be used to accelerate the convergence, but animprudent choice of preconditioner may spoil the achievable qualityof computed restorations.
Classical preconditioner
◮ Historically, the first attempt of this sort was by Hanke, Nagy,and Plemmons (1993): In that work
P = Cε,
where Cε is the optimal doubly circulant approximation of T ,with eigenvalues set to be one for frequencies above 1/ε.Very fast, but the choice of ε is delicate and not robust.
◮ Subsequently, other regularizing preconditioners have beensuggested: Kilmer and O’Leary (1999), Egger and Neubauer(2005), Brianzi, Di Benedetto, and Estatico (2008).
Outline
Iterative solution of ill-posed equations
A nonstationary preconditioned iteration
Application to image deblurring
Nonstationary iterated Tikhonov regularization
Given x0 compute for n = 0, 1, 2, . . .
zn = (T ∗T + αnI )−1T ∗rn , rn = y δ − Txn , (3a)
xn+1 = xn + zn . (3b)
This is some sort of regularized iterative refinement.
Choices of αn:
◮ αn = α > 0, ∀n, stationary.
◮ αn = αqn where α > 0 and 0 < q ≤ 1, geometric sequence(fastest convergence), Groetsch and Hanke (1998).
T ∗T + αI and TT ∗ + αI could be expensive to invert!
The starting idea
Approximate T by C and iterate
hn = (C ∗C + αnI )−1C ∗rn , rn = y δ − Txn , (4)
xn+1 = xn + hn . (5)
Preconditioner ⇒ P = (C ∗C + αnI )−1C ∗
Differences to previous preconditioners:
◮ gradual approximation of the optimal regularization parameter
◮ nonstationary scheme, not to be used in combinationwith cgls
◮ essentially as fast as nonstationary iterated Tikhonovregularization
Nonstationary preconditioning
The iterative refinement applied to the error equation Ten ≈ rnis correct up to noise, hence consider instead
Cen ≈ rn , (6)
possibly tolerating a slightly larger misfit.
Adaptive parameter choice strategy for αn
Choose αn s.t. the (6) is solved up to a certain relative amount:
‖rn − Chn‖ = qn‖rn‖ , (7)
where qn < 1, but not too small.
Choice of qn and approximation of T
Assumption: To derive the parameter qn in (7) we impose
‖(C − T )z‖ ≤ ρ ‖Tz‖ , z ∈ X , (8)
for some 0 < ρ < 1/2.
Consequence: Choose αn s.t.
‖rn − Chn‖ = qn‖rn‖ ,
with qn > ρ+ (1 + ρ)δ/‖rn‖ (discrepancy principle).
The Algorithm
Choose τ = (1 + 2ρ)/(1 − 2ρ) and fix q ∈ (2ρ, 1).While ‖rn‖ > τδ, let τn = ‖rn‖/δ, and compute αn s.t.
‖rn−Chn‖ = qn‖rn‖ , qn = max{
q, 2ρ+(1+ρ)/τn}
. (9a)
Then, updatehn = (C ∗C + αnI )
−1C ∗rn , (9b)
xn+1 = xn + hn , rn+1 = y δ − Txn+1 . (9c)
Details
◮ The parameter q prevents that rn decreases too rapidly.
◮ The unique αn can be computed by Newton iteration.
Theoretical results
◮ The norm of the iteration error en = x† − xn decreasesmonotonically as long as
‖rn‖ ≤ τδ ≤ ‖rn−1‖, τ > 1 fixed
⇒ discrepancy principle.
◮ For exact data (δ = 0) the iterates xn converges to a solutionx† of Tx = y .
◮ For noisy data (δ > 0), with the above stopping criterion, theresulting approximation xδ converges to x† as δ → 0.
Outline
Iterative solution of ill-posed equations
A nonstationary preconditioned iteration
Application to image deblurring
Image deblurring problems
yδ = T ∗ x + e
◮ T is large and severely ill-conditioned (discretizzation of aFredholm integral equations of the first kind)
◮ yδ known, measured data (blurred and noisy image)
◮ e (white Gaussian) noise, s.t. ‖e‖ = δ
Boundary Conditions (BCs)
zero Dirichlet Periodic
Reflective Antireflective
The matrix C
Space invariant point spread function (PSF)
⇓
T has a doubly Toeplitz-like structure that carries the “correct”boundary conditions.
◮ doubly circulant matrix C diagonalizzable by FFT, thatcorresponds to periodic BCs.
◮ The boundary conditions have a very local effect
T− C = E+ R , (10)
where E is of small norm and R of small rank.
Choice of parameters
◮ We do not know whether a rigorous estimate of
‖(T − C)z‖ ≤ ρ ‖Tz‖
will hold for all (relevant) vectors z and ρ < 1/2.According to our numerical tests, the estimate (??) is satisfiedwith ρ of a few percent, i.e., 10−3 or 10−2.
◮ q = 0.7, but all q ∈ [ 0.6, 0.8 ] gives comparable results.
Comparison with other methods:
◮ CGLS, P-CGLS, our algorithm with αn = 0.5 qn (‖T‖∞ ≈ 1).
◮ Stopping by discrepancy principle with τ = 1.01.
Numerical results
◮ We add to the blurred image a white Gaussian noise with therelative amount of noise
ν =δ
‖y‖
◮ To compare the quality of the restorations, we evaluate theirrelative restoration errors (RRE), i.e.,
RRE =‖x− x†‖
‖x†‖,
where x is the computed solution.
◮ CGLS and P-CGLS from the Matlab Toolbox RestoreToolsJ. Nagy et al. http://www.mathcs.emory.edu/~nagy/RestoreTools
Example 1
Size 256× 256, ν = 0.5%, ρ = 10−3, zero Dirichlet BCs.
True image PSF Observed image
RRE and αn
0 20 40 60 800.3
0.4
0.5
0.6
0.7
Algor. IGeometricCGLSPCGLS
0 5 10 15 20 2510
−4
10−3
10−2
10−1
100
Algor. IGeometric
RRE vs iterations αn vs n
◮ Algor. I is our nonstationary scheme with the automaticselection of αn by Newton.
◮ Geometric is our iteration but with αn = 0.5 qn.
Check of the assumption
Recall: αn chosen s.t. ‖rn − Chn‖ = qn‖rn‖ , whereqn = max
{
q, 2ρ+ (1 + ρ)/τn}
, under the assumption
‖(C − T )z‖ ≤ ρ ‖Tz‖ , ρ = 10−3
0 2 4 6 8 10
0.2
0.4
0.6
0.8
1
‖rn − Cen‖/‖rn‖qn
0 2 4 6 8
0.2
0.4
0.6
0.8
1
‖rn − Cen‖/‖rn‖qn
q = 0.7 q = 0.6
Example 2
◮ Size 452 × 452, ν = 1%, Antireflective BCs.
◮ The structure of T is more involved ⇒ ρ = 10−2
−5 0 5
5
0
−5
True image PSF Observed image
Avoid T∗
◮ Restorations in the range of T∗ come with unwantedboundary artefacts [D., Serra-Capizzano IP 2005].
◮ RestoreTools implements the rebluring strategy: replaces T∗
with T′ by PSF rotated by 180o, but T′T is not symmetric.
0 2 4 6 8 100.11
0.12
0.13
0.14
0.15
0.16
Algor. IGeometricCGLS
0 2 4 6 8 101
1.4
2
3
4
Algor. IGeometricCGLS
%
RRE vs iterations ‖rn‖/‖yδ‖ vs n
Restored images
CGLS stopped at the minimum RRE.
Our Algorithm Geometric Sequence cgls
RRE = 0.110, it. 6 RRE = 0.110, it. 9 RRE = 0.131, it. 5
Conclusions
◮ Under the assumption that an approximation C of T isavailable, our new scheme turns out to be fast and stable.
◮ The choice of ρ reflect how much we trust in the previousapproximation.
◮ A too small ρ can be detected by αn or ‖rn‖.
◮ The choice of q is not crucial (q = 0.7 balance between speedand stability).
◮ Our scheme does not require T ∗.
References
M. Donatelli and M. HankeFast nonstationary preconditioned iterative methods forill-posed problems, with application to image deblurring,Inverse Problems, 29 (2013) 095008.
The choice of αn is inspired by
M. HankeA regularizing Levenberg-Marquardt scheme, with applicationsto inverse groundwater filtration problems,Inverse Problems, 13 (1997), pp. 79–95.
Example 3
Size 237× 237, ν = 0.1%, ρ = 10−3, zero Dirichlet BCs.
True image PSF Observed image
Detect that ρ = 10−3 is too small for the assumption
‖rn‖ αn
0 5 10 15 20 25 30
0.1
1
10
Geometricρ = 0.01ρ = 0.001
%
0 5 10 15 20 2510
−6
10−4
10−2
100
Geometricρ = 0.01ρ = 0.001
◮ The algorithm often recovers (RRE = 0.254).
◮ Increasing ρ = 10−2 ⇒ the zigzagging disappears and theAlgorithm terminates with RRE = 0.261.
◮ Different ρ only affects the final stage of the iteration.