Research map / Image generation / Image restoration
Boosting Image De-Raining via Central-Surrounding Synergistic Convolution
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Synergistic Convolution (CSD + CSA, central-surrounding difference and addition) that replaces vanilla convolution in de-raining networks, then re-parameterizes it into a single kernel at inference for zero extra cost.
Problem
Single-image de-raining must jointly model anisotropic rain streaks and isotropic rain accumulation; unified representation learning for both leads to structure distortion and contrast degradation.
Approach
Inject central-surrounding operations into convolution: a difference branch (CSD) captures directional gradients of rain streaks, an addition branch (CSA) captures smooth accumulation using the central pixel as a proxy for local mean; the two parallel branches are merged into one k×k kernel after training via linearity of convolution.
Key results
- SC improves all twelve de-raining networks on nine datasets at zero extra inference cost; e.g., Restormer on Rain12600 rises from 34.04/0.934 to 34.359/0.944 PSNR/SSIM.
- Restormer with SC on SIRR-R, PI/NIQE etc. improves non-reference metrics; SC on PReNet improves Outdoor-Rain from 22.2 to 23.5 PSNR in ablation (SC vs VC).
- SC also gives gains on blind SR (DASR +0.42 dB on Set5), low-light enhancement (+0.67 dB on LOL), and motion deblurring (+0.59 dB on GoPro).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace vanilla convolutions with SC | DRSformer · Rain12 | PSNR: 38.015 -> 38.245 | Helped |
| Replace vanilla convolutions with SC | DRSformer · Rain200H | PSNR: 32.173 -> 32.313 | Helped |
| Replace vanilla convolutions with SC | DRSformer · Rain200L | PSNR: 41.232 -> 41.366 | Helped |
| Replace vanilla convolutions with SC | DRSformer · Rain12600 | PSNR: 35.354 -> 35.419 | Helped |
| Replace vanilla convolutions with SC | DRSformer · Rain1200 | PSNR: 34.354 -> 34.627 | Helped |
| Replace vanilla convolutions with SC | DRSformer · Outdoor-Rain | PSNR: 29.433 -> 29.679 | Helped |
| Use two vanilla convolutions in parallel instead of SC (ablation) | PReNet · Outdoor-Rain | PSNR: 23.5 -> 22.5 | Hurt |
| Use a vanilla convolution in parallel with CSD instead of SC (ablation) | PReNet · Outdoor-Rain | PSNR: 23.5 -> 22.9 | Hurt |
| Use a vanilla convolution in parallel with CSA instead of SC (ablation) | PReNet · Outdoor-Rain | PSNR: 23.5 -> 23.0 | Hurt |
| Replace SC with asymmetric convolution blocks (ablation) | PReNet · Outdoor-Rain | PSNR: 23.5 -> 17.4 | Hurt |
| Replace SC with RepVGG block (ablation) | PReNet · Outdoor-Rain | PSNR: 23.5 -> 15.9 | Hurt |
| Replace SC with central difference convolution (ablation) | PReNet · Outdoor-Rain | PSNR: 23.5 -> 21.9 | Hurt |
Surprising findings
- Simply doubling parameters by running two vanilla convolutions in parallel (VV) can drop performance on Rain200H, showing parallel branches may conflict.
- ACB, RepVGG-style block and CDC substantially degrade PReNet performance on deraining, despite being designed for high-level tasks.
Limitations
- Integration into existing VC-based architectures may not be optimal; the authors note a dedicated architecture would be needed to fully exploit SC.
- Training uses parallel branches, so training cost is higher than the vanilla model even though inference cost is identical.
Experimental setup
Datasets: Rain12 Rain200H Rain200L Rain1200 Rain12600 Outdoor-Rain JORDER-R ID-CGAN-R SIRR-R Set5 LOL GoPro
Benchmarks: Rain12 Rain200H Rain200L Rain1200 Rain12600 Outdoor-Rain JORDER-R ID-CGAN-R SIRR-R Set5 LOL GoPro
Metrics: PSNR SSIM NIQE BRISQUE PIQE PI inference time parameters FLOPs
Baselines: NLEDN RESCAN PReNet UNet Syn2Real SPANet MPRNet HCT-FFN FPNet SPDNet Restormer DRSformer
Builds on: RepVGG ACNet central difference convolution structural re-parameterization
Closest papers by meaning
- Deep Joint Rain Detection and Removal From a Single Image (CVPR 2017 · similarity 0.91)
- Rain Streak Removal via Dual Graph Convolutional Network (AAAI 2021 · similarity 0.91)
- Spatial Attentive Single-Image Deraining With a High Quality Real Rain Dataset (CVPR 2019 · similarity 0.91)
- Multi-Scale Progressive Fusion Network for Single Image Deraining (CVPR 2020 · similarity 0.90)
- Harnessing Joint Rain-/Detail-aware Representations to Eliminate Intricate Rains (ICLR 2024 · similarity 0.90)
- Bidirectional Multi-Scale Implicit Neural Representations for Image Deraining (CVPR 2024 · similarity 0.89)
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.