Research map / Image generation / Image restoration
Invertible Denoising Network: A Light Solution for Real Noise Removal
Yang Liu, Zhenyue Qin, Saeed Anwar, Pan Ji, Dongwoo Kim, Sabrina Caldwell, Tom Gedeon
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 |
|---|---|---|---|
| replace the sampled clean-detail latent with a prior sample while retaining the low-resolution clean representation | InvDN · SIDD test | PSNR | Improved |
| use InvDN instead of DANet | real-noise image denoising · SIDD test | PSNR: 39.25 -> 39.28 | Improved |
| use InvDN instead of MIRNet | real-noise image denoising · DND | PSNR: 39.88 -> 39.57 | Hurt |
| use Haar wavelet instead of squeeze for down-scaling (ablation) | InvDN · SIDD validation | PSNR | Improved |
| use residual blocks instead of dense blocks for the coupling subnetworks (ablation) | InvDN · SIDD denoising | accuracy | Improved |
| use residual blocks instead of dense blocks for the coupling subnetworks (ablation) | InvDN · SIDD denoising | number of parameters: 4.3M -> 2.6M | Improved |
| increase invertible blocks from 8 to 16 (ablation) | InvDN · SIDD validation | PSNR: 38.40 -> 38.61 | Improved |
| use 2 down-scale blocks instead of 1 or 3 (ablation) | InvDN with 8 invertible blocks · SIDD validation | PSNR: 38.00, 38.40, 37.69 | Mixed |
| use 2 down-scale blocks instead of 1 or 3 (ablation) | InvDN with 16 invertible blocks · SIDD validation | PSNR: 38.30, 38.61, 38.15 | Mixed |
| use L2 for the forward low-resolution objective and L1 for the backward reconstruction objective (ablation) | InvDN · SIDD validation | PSNR: 38.88 | Improved |
| add gradient or SSIM loss (ablation) | InvDN · SIDD validation | PSNR: 38.74, 38.80, 38.78 | No clear effect |
| average outputs from 16 sampled latent variables using Monte Carlo self-ensemble | InvDN · SIDD validation | images with a performance boost: 83.35% | Improved |
Closest papers by meaning
- FastDVDnet: Towards Real-Time Deep Video Denoising Without Flow Estimation (CVPR 2020 · similarity 0.88)
- Variational Denoising Network: Toward Blind Noise Modeling and Removal (NeurIPS 2019 · similarity 0.87)
- Self-Calibrated Variance-Stabilizing Transformations for Real-World Image Denoising (ICCV 2025 · similarity 0.87)
- Universal Denoising Networks : A Novel CNN Architecture for Image Denoising (CVPR 2018 · similarity 0.87)
- Unsupervised Image Restoration Using Partially Linear Denoisers (TPAMI 2021 · similarity 0.87)
- Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration (ICLR 2025 · similarity 0.85)
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.
