Research map / Image generation / Image restoration
Learning Proximal Operators: Using Denoising Networks for Regularizing Inverse Imaging Problems
Tim Meinhardt, Michael Moller, Caner Hazirbas, Daniel Cremers
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use a DnCNN-S denoising network as the proximal operator instead of BM3D in demosaicking | FlexISP-style PDHG · Bayer demosaicking, McMaster (18 images) | PSNR (dB): 36.12 -> 37.12 | Improved |
| replace BM3D with the DnCNN-S network as the proximal operator in deconvolution | FlexISP-style PDHG · Deconvolution, 11 grayscale images, average over experiments a-e | PSNR (dB): 24.997267 -> 25.11 | Improved |
| use one generic denoiser instead of MLP trained per experiment | Deconvolution · Deconvolution, 11 grayscale images, average over experiments a-e | PSNR (dB): 25.27 -> 25.11 | Hurt |
| use the generic denoiser instead of the specialized MLP on motion blur | Deconvolution · Deconvolution experiment e (motion blur) | PSNR (dB): 29.42 -> 30.17 | Improved |
| use the fixed denoiser versus a deep CNN trained for demosaicking | Demosaicking · Bayer demosaicking, McMaster (18 images) | PSNR (dB) | Hurt |
| train the denoising network on a smaller noise level (sigma 0.001 instead of 0.02) (ablation) | Ours · Bayer demosaicking, McMaster (18 images) | PSNR (dB): 37.123543 -> 36.050674 | Hurt |
| train the denoising network on sigma 0.05 instead of 0.02 (ablation) | Ours · Bayer demosaicking, McMaster (18 images) | PSNR (dB): 37.123543 -> 36.082635 | Hurt |
| add explicit TV and cross-channel priors besides the denoiser (ablation) | Ours · Bayer demosaicking, McMaster (18 images) | PSNR (dB): about 0.4 dB | Improved |
| train the denoiser on sigma 0.06 instead of 0.02 (ablation) | Ours · Deconvolution, 11 grayscale images, average over experiments a-e | PSNR (dB): 25.11 -> 25.3257 | Improved |
| train the denoiser on sigma 0.2 instead of 0.02 (ablation) | Ours · Deconvolution, 11 grayscale images, average over experiments a-e | PSNR (dB): 25.11 -> 24.2501 | Hurt |
| train the denoiser on sigma 0.01 instead of 0.02 (ablation) | Ours · Deconvolution, 11 grayscale images, average over experiments a-e | PSNR (dB): 25.11 -> 24.6253 | Hurt |
| use the neural denoiser in place of BM3D inside the optimization for runtime | Deconvolution · Deconvolution runtime | seconds: approx. 4s -> approx. 2.5s (37.5% improvement) | Improved |
Closest papers by meaning
- A Restoration Network as an Implicit Prior (ICLR 2024 · similarity 0.90)
- What's in a Prior? Learned Proximal Networks for Inverse Problems (ICLR 2024 · similarity 0.90)
- Unleashing the Denoising Capability of Diffusion Prior for Solving Inverse Problems (NeurIPS 2024 · similarity 0.89)
- Stochastic Deep Restoration Priors for Imaging Inverse Problems (ICML 2025 · similarity 0.89)
- Weakly Convex Regularisers for Inverse Problems: Convergence of Critical Points and Primal-Dual Optimisation (ICML 2024 · similarity 0.89)
- A Variational Perspective on Solving Inverse Problems with Diffusion Models (ICLR 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
