Convergent Bregman Plug-and-Play Image Restoration for Poisson Inverse Problems

Samuel Hurault, Ulugbek Kamilov, Arthur Leclaire, Nicolas Papadakis

Introduction

Plug-and-Play (PnP) frameworks have become a cornerstone of modern image restoration, allowing researchers to integrate the representative power of deep-learning-based denoisers into iterative optimization algorithms. However, the vast majority of PnP research is predicated on the assumption of additive white Gaussian noise. In many critical scientific fields—including medical imaging (PET scans) and astronomy—the noise is signal-dependent and follows a Poisson distribution.

Applying standard Euclidean PnP methods to Poisson inverse problems is mathematically problematic. The Poisson data-fidelity term, typically expressed as the Kullback-Leibler (KL) divergence, lacks a Lipschitz-continuous gradient. This violation of standard smoothness assumptions means that traditional PnP schemes often fail to converge or require unrealistic constraints on the neural network's architecture. This paper introduces a new framework that adapts the optimization geometry to the physics of Poisson noise using Bregman divergences.

Methodology

Bregman Geometry and Burg’s Entropy

To address the non-Lipschitz nature of the Poisson likelihood, the authors shift from standard Euclidean distances to the Bregman Proximal Gradient (BPG) method. They utilize Burg’s entropy as a potential function to define a specific Bregman divergence. This choice of geometry is mathematically compatible with the Poisson likelihood, satisfying the "NoLip" condition—a generalization of smoothness that ensures stable optimization in non-Euclidean spaces.

The Bregman Score Denoiser

A central innovation of this work is the Bregman Score Denoiser. Standard denoisers are typically trained to remove Gaussian noise, which is an ill-fit for the Poisson setting. The authors propose training a denoiser (based on the DRUNet architecture) specifically to remove "Bregman noise." For the geometry defined by Burg’s entropy, this corresponds to an Inverse Gamma distribution.

By deriving a generalized version of Tweedie’s formula for exponential families in Bregman space, the authors establish a direct link between the neural network's output and the gradient of a log-prior. This transformation turns a "black-box" deep learning model into a mathematically rigorous regularization term.

Proposed Algorithms: B-RED and B-PnP

The paper introduces two algorithmic variants:

  1. B-RED: A Bregman version of Regularization-by-Denoising.
  2. B-PnP: A Bregman version of the standard PnP proximal gradient.

By leveraging the Kurdyka-Lojasiewicz (KL) property and semi-algebraicity, the authors provide the first rigorous convergence guarantees for these algorithms in the Poisson setting. They prove that the iterates are guaranteed to converge to stationary points of an explicit global cost function, even when employing nonconvex deep-learning priors.

Experimental Evaluation

The framework was evaluated using the CBSD68 dataset, a standard benchmark of 68 natural images. The images were subjected to varying levels of synthetic Poisson noise (parameterized by $\alpha$, where lower values indicate higher noise) and four different blur kernels, including realistic camera shake and standard Gaussian blurs.

Key Results

  • Performance Gains: The proposed B-RED and B-PnP consistently outperformed standard PnP methods that rely on Gaussian priors. In deblurring tasks with moderate noise ($\alpha=60$), B-RED achieved a PSNR of 24.90 dB, significantly higher than the 24.45 dB produced by standard PnP-PGD.
  • Theoretical Validation: Experimental plots confirmed that the objective function decreases monotonically. This empirical evidence validates the theoretical proofs of convergence provided in the paper.
  • Denoiser Efficiency: The Bregman Score Denoiser maintained high restoration quality while satisfying the mathematical properties required for the convergence proof, demonstrating that rigorous theory does not have to come at the cost of performance.

Discussion and Limitations

While the Bregman framework marks a significant step forward, some challenges remain. The authors noted that at extreme noise levels ($\alpha=20$), the performance gap between Bregman and standard methods narrows. This suggests that the current deep prior may require additional "regularization capacity" to handle very low-photon counts.

Furthermore, the convergence proof relies on a local convexity condition for the deep potential. While this condition is empirically observed to hold near the image manifold, enforcing it globally across the entire image space remains an open area for future research.

Conclusion

This research provides a robust mathematical foundation for solving Poisson inverse problems using deep learning priors. By aligning the optimization geometry with the statistical nature of the noise, the authors have created a framework that is both high-performing and theoretically sound. This work paves the way for more reliable image restoration in medical and scientific imaging where Poisson statistics are dominant.

Summary updated

BibTeX
@inproceedings{hurault2023convergent,
  author = {Samuel Hurault and Ulugbek Kamilov and Arthur Leclaire and Nicolas Papadakis},
  title = {{Convergent Bregman Plug-and-Play Image Restoration for Poisson Inverse Problems}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2023},
  url = {https://openreview.net/forum?id=clCELP8zFb},
}
Figure