Nested Conjugate Gradient Algorithm With Nested Preconditioning for Non-Linear Image Restoration

We develop a novel optimization algorithm, which we call nested non-linear conjugate gradient (CG) algorithm (NNCG), for image restoration based on quadratic data fitting and smooth non-quadratic regularization. The algorithm is constructed as a nesting of two conjugate gradient iterations. The oute...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:IEEE transactions on image processing Ročník 26; číslo 9; s. 4471 - 4482
Hlavní autoři: Skariah, Deepak G., Arigovindan, Muthuvel
Médium: Journal Article
Jazyk:angličtina
Vydáno: United States IEEE 01.09.2017
Témata:
conjugate gradient method
Convergence
convex regularization
Cost function
Gradient methods
Image restoration
Minimization
Noise measurement
nonconvex regularization
optimization
preconditioning
ISSN:1057-7149, 1941-0042, 1941-0042
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:We develop a novel optimization algorithm, which we call nested non-linear conjugate gradient (CG) algorithm (NNCG), for image restoration based on quadratic data fitting and smooth non-quadratic regularization. The algorithm is constructed as a nesting of two conjugate gradient iterations. The outer iteration is constructed as a preconditioned non-linear CG algorithm; the preconditioning is performed by the inner CG iteration that is linear. The inner CG iteration, which performs preconditioning for outer CG iteration, itself is accelerated by an another FFT-based non-iterative preconditioner. We prove that the method converges to a stationary point for both convex and non-convex regularization functionals. We demonstrate experimentally that proposed method outperforms the well-known majorization-minimization method used for convex regularization, and a non-convex inertial-proximal method for non-convex regularization functional.
Bibliografie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 23
ISSN:1057-7149
1941-0042
1941-0042
DOI:10.1109/TIP.2017.2717182