A shearlet-based fast thresholded Landweber algorithm for deconvolution

Abstract

Image deconvolution is an important problem, which has seen plenty of progress in the last decades. Due to its ill-posedness, a common approach is to formulate the reconstruction as an optimization problem[Formula: see text] regularized by an additional sparsity-enforcing term. This term is often modeled as an [Formula: see text] norm measured in the domain of a suitable signal transform. The resulting optimization problem can be solved by an iterative approach via Landweber iterations with soft thresholding of the transform coefficients. Previous approaches focused on thresholding in the wavelet-domain. In particular, recent work [C. Vonesch and M. Unser, A fast thresholded Landweber algorithm for wavelet-regularized multidimensional deconvolution, IEEE Trans. Image Process. 17(4) (2008) 539–549.] has shown that the use of Shannon wavelets results in particularly efficient reconstruction algorithms. The present paper extends this approach to Shannon shearlets, which we also introduce in this work. We show that for anisotropic blurring filters, such as the motion blur, the novel shearlet-based approach allows for further a improvement in efficiency. In particular, we observe that for such kernels using shearlets instead of wavelets improves the quality of image restoration and SERG, when compared after the same number of iterations.