Plug-and-Play image restoration with Stochastic deNOising REgularization

Authors: Marien Renaud, Jean Prost, Arthur Leclaire, Nicolas Papadakis

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimentally, we prove that SNORE is competitive with respect to state-of-the-art methods on deblurring and inpainting tasks, both quantitatively and qualitatively.
Researcher Affiliation Academia 1Univ. Bordeaux, CNRS, INRIA, Bordeaux INP, IMB, UMR 5251, F-33400 Talence, France 2Universit e Paris-Cit e, CNRS, MAP5, France 3LTCI, T el ecom Paris, IP Paris.
Pseudocode Yes Algorithm 1 RED (Romano et al., 2017) ... Algorithm 2 SNORE ... Algorithm 3 Annealed SNORE
Open Source Code Yes The code used in these experiments can be found in https://github.com/Marien-RENAUD/SNORE.
Open Datasets Yes We evaluate each method on 10 images from CBSD68 (Martin et al., 2001) and 10 blur kernels presented in Figure 1. The denoiser proposed by (Hurault et al., 2022a) based on the DRUNet (Zhang et al., 2021) trained on a dataset of natural images composed of Berkeley segmentation dataset (CBSD) (Martin et al., 2001), Waterloo Exploration dataset (Ma et al., 2017), DIV2K dataset (Agustsson & Timofte, 2017) and Flick2K (Lim et al., 2017).
Dataset Splits No We evaluate each method on 10 images from CBSD68 (Martin et al., 2001) and 10 blur kernels presented in Figure 1.
Hardware Specification Yes Computing all the necessary experiments to generate Table 1 requires 9 hours and 40 minutes on a GPU NVIDIA Quadro RTX 8000.
Software Dependencies No We use the implementation of the Python library Deep Inverse modified to add a time tstart < T such as proposed by the authors (Zhu et al., 2023, Section 4.4). We use the Python library brisque , with which we sometimes observe some incoherence with some outlier outputs (smaller than 0 or larger than 100). We use the Python library lpips to compute this metric.
Experiment Setup Yes We set 1500 iterations and m = 16 annealing levels. To ensure convergence, we run 300 iterations with the last parameters σm 1, αm 1. For all input noise levels σy, we set σ0 = 1.8σy, σm 1 = 0.5σy, α0 = 0.1σ2 0σ 2 y and αm 1 = σ2 m 1σ 2 y . We initialize with the observation x0 = y and use a fixed step-size, δk = δ = 0.1. Table 4. Parameters setting for image deblurring for the different implemented methods.