Stochastic Unrolled Proximal Point Algorithm for linear image inverse problems - INRIA - Institut National de Recherche en Informatique et en Automatique Access content directly
Conference Papers Year : 2023

Stochastic Unrolled Proximal Point Algorithm for linear image inverse problems


Unrolled optimization methods have emerged as a way to combine classical iterative optimization techniques with learned priors to efficiently solve image restoration problems. However, learning the regularization prior along the unrolled iterations requires intensive memory usage due to the deep explicit backpropagation, hence making the number of unrolled iterations usually small in practice. Inspired by deep equilibrium models, unrolling models with implicit backpropagation have been considered for solving this issue. Nevertheless, while these methods yield good restoration quality with reduced memory usage, the theory of implicit backpropagation relies on the knowledge of the fixed point of the function to optimize, usually unknown and estimated by iterating until convergence. Therefore, these methods require intensive computation time to ensure a stable backpropagation. In this paper, we present an unrolled Proximal Point Algorithm method, where the end-toend optimization problem is redefined as a per unrolled iteration optimization problem. We prove that the proposed optimization strategy is memory-efficient and applicable for any number of computed unrolled iteration. We empirically show that our method achieves state-of-the-art image restoration quality.
Fichier principal
Vignette du fichier
Stochastic_Unrolled_Proximal_Point_Algorithm_for_linear_image_inverse_problems__EUSIPCO_edition_.pdf (395.2 Ko) Télécharger le fichier
Origin Files produced by the author(s)

Dates and versions

hal-04109191 , version 1 (29-05-2023)



  • HAL Id : hal-04109191 , version 1


Brandon Le Bon, Mikaël Le Pendu, Christine Guillemot. Stochastic Unrolled Proximal Point Algorithm for linear image inverse problems. EUSIPCO 2023 - 31st European Signal Processing Conference, Sep 2023, Helsinki, Finland. pp.1-5. ⟨hal-04109191⟩
130 View
132 Download


Gmail Mastodon Facebook X LinkedIn More