Quasi-SVD: Learning a Lie-constrained matrix factorisation for real-time imaging

DOI10.48550/arXiv.2607.25967MaRDI QIDQ7284444arXivFDO

Authors Christopher Hahne

Publication date 28 July 2026

Published in arXiv

Tldr This paper introduces Quasi-SVD, a novel differentiable and fully parallelized matrix factorization framework for GPUs that achieves reconstruction fidelity of SSIM = 0.89-0.94 while accelerating computation by 3-20x compared to cuSOLVER and randomized SVD, enabling throughput above 25 FPS on two medical imaging tasks: ultrasound localisation microscopy and Mueller matrix polarimetry.

Full work available at URL https://doi.org/10.48550/arXiv.2607.25967

arXiv classification

cs.CVcs.LGmath.NA

Cites work

• A Hierarchically Blocked Jacobi SVD Algorithm for Single and Multiple Graphics Processing Units
• A Krylov-Schur approach to the truncated SVD
• Efficient Algorithms for Computing a Strong Rank-Revealing QR Factorization
• Fast low-rank modifications of the thin singular value decomposition
• Finding structure with randomness: probabilistic algorithms for constructing approximate matrix decompositions
• Jacobi’s Method is More Accurate than QR
• Matrix Computations
• Singular value decomposition and least squares solutions
• Tensor SVD: Statistical and Computational Limits
• The Parallel Evaluation of General Arithmetic Expressions
• The truncated SVD as a method for regularization
• Truncated Singular Value Decomposition Solutions to Discrete Ill-Posed Problems with Ill-Determined Numerical Rank

This page was built for publication: Quasi-SVD: Learning a Lie-constrained matrix factorisation for real-time imaging

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q7284444)