Non-overlapping domain decomposition methods for dual total variation based image denoising (Q2515535)

It is often said that a good algorithm is better than many hardware improvements. Sometimes, however, the continuous hardware improvements make it possible to treat many more amounts of data but in this case, algorithms that worked decently until then, cannot process such amount of big data sets. This is the case in the problem of image denoising. The existing state-of-the-art numerical algorithms for the recovery of an unknown image using the total variation as a regularization technique, perform well on small and medium-scale problems, but they are not capable of managing extremely large problems in realistic CPU-time. The natural alternative is to split the computational workload and to solve a sequence of smaller problems. After a short review of the existing splitting methods in the literature where the authors point out the pros and cons of each one, they introduce a sequential and parallel non-overlapping domain decomposition method closely related to the decomposition method considered by \textit{X.-C. Tai} [Numer. Math. 93, No. 4, 755--786 (2003; Zbl 1057.65040)] and by \textit{X.-C. Tai} and \textit{J. Xu} [Math. Comput. 71, No. 237, 105--124 (2002; Zbl 0985.65065)]. Nevertheless, since the proposed problem is convex but not strongly convex, the convergence theory used in the cited papers is not directly applicable. Thus, they had to show its convergence to a minimizer of the global problem using new techniques. Numerical experiments for image denoising show the effectiveness of the new method not only for 2D images but also for 3D image data sets coming from computerized tomography. In this 3D case, the parallel version of the algorithm seems to perform best.