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

Plug-and-Play (PnP) methods are efficient iterative algorithms for solving ill-posed image inverse problems. PnP methods are obtained by using deep Gaussian denoisers instead of the proximal operator or the gradient-descent step within proximal algorithms. Current PnP schemes rely on data-fidelity terms that have either Lipschitz gradients or closed-form proximal operators, which is not applicable to Poisson inverse problems. Based on the observation that the Gaussian noise is not the adequate noise model in this setting, we propose to generalize PnP using theBregman Proximal Gradient (BPG) method. BPG replaces the Euclidean distance with a Bregman divergence that can better capture the smoothness properties of the problem. We introduce the Bregman Score Denoiser specifically parametrized and trained for the new Bregman geometry and prove that it corresponds to the proximal operator of a nonconvex potential. We propose two PnP algorithms based on the Bregman Score Denoiser for solving Poisson inverse problems. Extending the convergence results of BPG in the nonconvex settings, we show that the proposed methods converge, targeting stationary points of an explicit global functional. Experimental evaluations conducted on various Poisson inverse problems validate the convergence results and showcase effective restoration performance.

Paper

Similar papers

Peer review

Reviewer oqFb6/10 · confidence 3/52023-06-29

Summary

Inspired by the No-Lips literature on the optimization of convex objectives which are not globally L-smooth, this paper adds to the PnP literature the “Bregman score denoiser” which extends the BPG algorithm by a Bregman-based prox-map along with convergence conditions despite NN-parametrized non-convex potentials are involved.

Strengths

Originality: This is a plausible extension of recent related work on PnP networks based on state-of-the-art theory from the field of numerical optimization. Quality: It is apparant that the authors do not both fields very well. Significance: The paper adds a new concept to the PnP literature.

Weaknesses

Clarity: The presentation intersperses references, top-level arguments and technical details in a confusing manner. I had to read few times forth and back in order to get an idea what this paper is about. Authors criticize “unrealistic assumptions” (e.g. line 92) in related work but have to admit later on that their own assumptions are hard to check as well, even for a simple scenario (lines 275-277). The need for backtracking line search is not convenient either. Significance: Regarding the theory, I did not get if some `generic’ properties of the Bregman-prox-maps exist, that play the very same role like, say, the firmly-nonexpansiveness of Euclidean prox-maps, to achieve convergence in the considered generalized scenarios. The concrete scenario (Poisson noise) is classical. A comparison to a related approach shows no improvement (Fig.1, (c), (d)). In particular, artificially corrupting images looks like old-style image denoising papers.

Questions

Before eq. (14), the convexity assumption regarding F and R, adopted in [Bauschke et al 2017] to which you refer, is missing. line 92: why is nonexpansiveness considered as an “unrealistic requirement” to ensure convergence pf PnP, in view of your own assumptions which are not easy to check either? How is the set C in (18) determined?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

2 fair

Contribution

3 good

Limitations

This point has not been explicitly addressed in the paper, apparently. Please comment.

Reviewer 5jWQ6/10 · confidence 4/52023-07-02

Summary

This paper studies an extension of the Plug-and-Play (PnP) framework for solving inverse imaging problems by considering descent schemes in metrics different from L2: Motivated by the fact that some data fidelity terms such as the Kullback-Leibler divergence allow for an efficient minimization with the Bregman Proximal Gradient (BPG) method (an extension of proximal gradient descent to arbitrary Bregman distances instead of squared L2), the authors develop a parametrization of a learnable denoiser which can be interpreted as a proximal operator (or descent step) of a cost function w.r.t. the underlying Bregman distance that combines well with a particular data fidelity term. Under some additional assumptions, this allows proving the convergence of the resulting Bregman PnP framework. Numerical experiments illustrate that the resulting scheme can successfully solve deblurring problems with Poisson noise.

Strengths

The paper is technically sound and presents the technical construction of the Bregman PnP approach very well. It is an elegant solution that closes a gap missing in the PnP (and RED) framework with learnable priors. I found it very convincing in terms of its theory and it even derives some general (smaller) missing pieces for convergence beyond the seminal works by Bolte, Bauschke, Teboulle and co-authors. Despite this strengths section being short in comparison to the weaknesses, I think that the strong theoretical contribution along with an illustration of the implementation outweighs the weaknesses in terms of (benchmark) results, such that I am leaning towards accepting the paper.

Weaknesses

The numerical experiments/results are not very convincing from a practical point of view. - First, the denoising performance of the dedicated network is not better than that of a plain network (might not be very important). - Second, the main paper does not compare to the plain PnP approaches with L2 fidelity. The supplementary material reports tiny differences only (with the proposed approach being slightly worse for large noise and slightly better for small noise) - According to line 275, the desired condition of $\phi_\gamma \circ \nabla h^*$ being convex is not enforced but seems to hold empirically when training the network. Thus, any convergence guarantee is lost. - Figure 5 in the supplementary material shows a concerning risk of ending up with a bad result. The mitigation strategy of first running 100 iterations with a first set of parameters and then switching to a different set of parameters represents a significant amount of fine-tuning (possibly exceeding the number of hyperparameters and amount of fine-tuning used for the standard L2 PnP approach), such that even the small improvements in table 3 of the supplementary material need to be viewed with care. Minor aspects: - In 225 the author decide to do backtracking line search to avoid estimating the NoLip constant, but would backtracking on the (differentiable but not L-smooth) data fidelity term considered here not work in the L2 case? - The authors mention that there is no result on the convergence of the PDHG method for nonconvex regularizers (line 81). As I was curious, I briefly searched online and found "Precompact convergence of the nonconvex Primal–Dual Hybrid Gradient algorithm" by Sun et al., Journal of Computational and Applied Mathematics, 2018. Precompactness of the primal variable (the image) would be easy to ensure if one restricts every value to [0,1]. Is their result applicable? (Honestly, I have not read the paper yet). - The condition $\lambda L_f <1 $ in Theorem 2 seems to limit the amount of data fidelity one can use in order to still have a convergent algorithm - is this a limitation?

Questions

Considering a difficult (2-stage) optimization with different parameters to avoid bad minimizers as shown in Fig. 5 of the supplement, a lack of strict convergence guarantee as the convexity condition cannot be enforced, and negligible difference to a plain PnP or RED approach in terms of the PSNR, what is the advantage of the proposed method?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

I don't think there is a potential negative societal impact of this work. In terms of general limitations, I think the authors should be more open about the fact, that the use of the Bregman distance framework did not result in improved results over prior algorithmic schemes.

Reviewer 6Scw6/10 · confidence 4/52023-07-06

Summary

This paper develops a Bregman Plug and Play image restoration algorithm for solving under-determined inverse problems in the presence of Poisson measurement noise. The framework trains the image denoising algorithms used within PnP iterations (the "Bregman Score Denoiser") to remove noise with an exponential distribution that depends on the distribution of the measurement noise (as opposed to the Gaussian noise used for training previous PnP denoisers). The paper integrates the proposed denoiser into iterative algorithms to form B-PnP and B-RED. Both algorithms effectively deblur images in the presence of Poisson noise and are provably convergent to fixed points.

Strengths

-Generally well-written -Well motivated -I believe technically sound -Evaluated on several different blur kernels, including those based on camera shake. (Blur kernels were known)

Weaknesses

-Based on table 3 in the appendix, the proposed method does not meaningfully improve performance over existing methods (though it does come with convergence guarantees) -Comparisons with existing methods are not particularly comprehensive -In practice, PnP algorithms are quite sensitive to how hyperparameters are chosen and setting them correctly can be a challenge. This algorithm introduces another parameter, gamma.

Questions

Typos Line 90: "see also a review Kamilov et al." --> "see also a review by Kamilov et al." Line 197: Puting (i) on a separate line would read cleaner

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Not discussed.

Reviewer 5jWQ2023-08-16

Please make limitations clear in the revised version - otherwise: thanks, nice work!

Dear authors, thanks a lot for the detailed answers and the new section on the limitations! I think the restriction on $\lambda$ could be mentioned explicitly as well. Just to make sure: The constraint on $\lambda$ was respected in the experiments, right? Otherwise, this crucially needs to be pointed out.

Authorsrebuttal2023-08-17

Thanks for the advise. We will update the limitation paragraph to mention the restriction on $\lambda$ for the B-PnP algorithm. The constraint $\lambda L_f<1$, which is specific to the B-PnP algorithm (B-RED converges without such constraint) , may not be respected in our experiments. The best estimation of a global NoLip constant $L_f$ for the Poisson data-fidelity term we could get is $L_f=||y||_1$ (see Appendix E.3). However, for an image, the value $||y||_1$ is large and the restriction on $\lambda < \frac{1}{||y||_1}$ leads to extremely small $\lambda$ values. This approximation of a *global* NoLip constant $L_f$ can be *locally* very lose. In particular, the majoration (127) used for estimating this constant, is, for most images, way over-estimated. In order to still guarantee the convergence, as mentioned Appendix E.3, we adopt the following backtracking-like strategy to adjust the reguralization parameter $\lambda$: - Choose an initial value for $\lambda > 0$ - At each iteration $k$ of the B-PnP algorithm. Check sufficient decrease of the objective function $F_{\lambda,\gamma} = \lambda f + \phi_\gamma$ i.e. $F_{\lambda,\gamma}(x_{k}) - F_{\lambda,\gamma}(x_{k+1}) < \delta D_h(x_{k+1},x_k)$ If at some iteration, this condition is not satisified before convergence, we alert the user and restart the algorithm with $\lambda \longleftarrow \eta \lambda$. We also let the user know that, for optimal performance, it might be necessary to adjust the regularization parameter $\gamma$ of the denoiser, in order to compensate for this decrease of $\lambda$ With the proposed default value of $\lambda=0.025$, over the variety of blur kernels and noise levels experimented, the sufficient decrease property was always verified and this backtracking algorithm was never activated. This illustrates that $||y||_1$ is a bad approximation of the NoLip constant. In order to clarify this point we propose to include the previous discussion in a new Appendix and add the following paragraph in the main paper : In our experiments, the constraint $\lambda L_f<1$ of the B-PnP algorithm may not be respected. The *global* NoLip constant $L_f$ can indeed be *locally* very lose. As explained in the Appendix, we can adopt a backtracking-like strategy on the regularization parameter $\lambda$ to ensure convergence. Nevertheless, with the proposed default value $\lambda=0.025$, this backtracking algorithm was never activated over the variety of blur kernels and noise levels experimented."

Reviewer 6Scw2023-08-17

Thanks for the response. The proposed comparisons with existing methods will improve the paper.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC