549. Resource-Constrained Plug-and-Play Imaging: a block proximal heavy ball approach
Invited abstract in session WB-4: Optimization and learning for estimation problems, stream Optimization for machine learning.
Wednesday, 10:30-12:30Room: B100/5013
Authors (first author is the speaker)
| 1. | Andrea Sebastiani
|
| Department of Physics, Informatics and Mathematics, University of Modena and Reggio Emilia | |
| 2. | Federica Porta
|
| Universita' di Modena e Reggio Emilia | |
| 3. | Simone Rebegoldi
|
| Dipartimento di Scienze Fisiche, Informatiche e Matematiche, Università di Modena e Reggio Emilia |
Abstract
In this talk, we propose a novel memory-efficient optimization framework for Plug-and-Play (PnP) approaches in computational imaging. We introduce a block proximal variant of an heavy ball method that enables effective image reconstruction in resource-constrained settings. In particular, the memory required to perform the denoising step on the single block is significantly reduced, compared to the full image counterpart. Theoretical convergence guarantees are established under mild assumptions, and extensive experimental results demonstrate the efficacy of our method across multiple imaging problems including deblurring and super-resolution. The proposed framework extends the applicability of PnP reconstruction techniques to limited-memory scenarios without sacrificing performance or convergence properties.
Keywords
- Optimization for learning and data analysis
- Data driven optimization
- Computational mathematical optimization
Status: accepted
Back to the list of papers