Alternating Krylov subspace image restoration methods (Q2428120)

Let \(f^\delta\) represent the available noise- and blur-contaminated image and \(\widehat u\) the associated image that is to recover. The model \[ f^\delta(x)= \int h(x,y)\widehat u(y)\,dy+ \eta^\delta(x),\quad x\in\Omega, \] with the noise \(\eta^\delta\) is assumed. In this paper, the unknown \(\widehat u\) is considered as a solution of the minimization problem \(\min J(u,w)\), where \[ J(u,w)= \int\int (h(x,y) u(y)\,dy- f^\delta(x))^2 dx+\alpha \int({\mathcal L}(u- w)(x)^2+ \beta|\nabla w(x)|)\,dx \] with a certain regularization operator \({\mathcal L}\). The authors solve a discrete version of this problem by alternating interative methods. Krylov subspace based two-way alternating methods are studied, that allow the application of regularization operators different from the identity in both deblurring and the denoising step. Deblurring applies a few steps of Golub-Kahan bidiagonalization, while denoising is performed by an adaptive total-variation diminishing method. Convergence of the methods is proved. Four numerical examples are discussed, these examples illustrate by impressive figures the computed restorations.