Proximal Denoiser for Convergent Plug-and-Play Optimization with Nonconvex Regularization

Authors: Samuel Hurault, Arthur Leclaire, Nicolas Papadakis

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental These convergence results are confirmed with numerical experiments on deblurring, super-resolution and inpainting.
Researcher Affiliation Academia 1Univ. Bordeaux, Bordeaux INP, CNRS, IMB, UMR 5251,F33400 Talence, France. Correspondence to: Samuel Hurault <samuel.hurault@math.u-bordeaux.fr>.
Pseudocode Yes The iterative algorithms Pn P-PGD (8), Pn P-ADMM (10), and Pn P-DRS (11, 12, 14) are presented with structured mathematical equations that define the steps of the procedure.
Open Source Code Yes 1Code is available at https://github.com/ samuro95/Prox-Pn P.
Open Datasets Yes Average denoising PSNR performance of our proxdenoiser and compared methods on 256x256 center-cropped images from the CBSD68 dataset (Martin et al., 2001), for various noise levels σ.
Dataset Splits No The paper mentions using the CBSD68 dataset and training/testing, but does not explicitly provide the percentages or specific counts for the training, validation, and test splits.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No The paper does not list specific software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions).
Experiment Setup Yes The algorithm terminates when the relative difference between consecutive values of the objective function is less than ϵ = 10^-8 or the number of iterations exceeds K = 1000. We will use for evaluation Gaussian noise with 3 noise levels ν ∈ {2.55, 7.65, 12.75}/255 i.e. ν ∈ {0.01, 0.03, 0.05}. For each noise level, we propose default values for the parameters σ and λ that we keep for both deblurring and super-resolution. These values are explicitly given in Appendix G.2. Table 7: Choice of parameters (λ, σ) for both debluring and super-resolution experiments of Section 5.2. These parameters are only scaled with respect to the input noise level ν.