Iranian Journal of Numerical Analysis and Optimization

Iranian Journal of Numerical Analysis and Optimization

GPU-accelerated tensor global GMRES for matrix-free video restoration and dynamic medical imaging

Document Type : Research Article

Author
ENSA Oujda, Equipe MSN, Lab. LM2N, Université Mohammed Premier, Oujda, Morocco.
10.22067/ijnao.2026.101077.1987
Abstract
Reconstructing high-quality video sequences from degraded measurements in dynamic imaging applications—such as spatiotemporal CT and dynamic MRI—presents a computationally demanding, large-scale ill-posed inverse problem. Conventional matrix-flattening and vectorization approaches destroy the intrinsic multidimensional geometry of video data, resulting in the loss of spatial-temporal correlations and high memory overhead. To mitigate these limitations, we formulate large-scale regularized video restoration within a unified tensor-based framework using the Einstein tensor product, thereby preserving structural integrity and facilitating efficient GPU parallelization. In this work, we propose and investigate two specialized matrix-free tensor Krylov subspace solvers: \textbf{Tensor Gl-Aya-GMRES} (extending Ayachour's formulation to tensors) and \textbf{Tensor Gl-Simpler-GMRES}, evaluating their performance alongside the standard \textbf{Tensor Gl-Givens-GMRES} baseline. Operating via the Tensor Arnoldi process, these methods update solutions directly over tensor Krylov subspaces and enable low-cost implicit residual norm monitoring without explicitly reconstructing full video tensors at each inner iteration. Comprehensive numerical evaluations on simulated dynamic medical imaging sequences (spatiotemporal CT and MRI) and video tensors, implemented on GPU architectures, demonstrate that the proposed global GMRES variants achieve significantly faster convergence, reduced computational execution times, and lower GPU memory footprints compared to traditional approaches.
Keywords
Subjects

Send comment about this article
Enter Name.
Enter a valid email address.
Enter a vaid affiliation.
Enter comments (At leaset 10 words)
CAPTCHA Image
Enter Security Code Correctly.

Articles in Press, Accepted Manuscript
Available Online from 21 November 2026