Research map / Image generation / Image restoration
Bidirectional Multi-Scale Implicit Neural Representations for Image Deraining
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 NeRD-Rain, an end-to-end bidirectional multi-scale Transformer that adds intra-scale implicit neural representation (INR) branches and inter-scale bidirectional feedback units for single-image deraining.
Problem
Single image deraining with single-scale Transformer methods that fail to exploit multi-scale rain representations and common degradation features across spatially varying rain streaks.
Approach
Combines a multi-scale Transformer (unequal UNet branches per scale) with coordinate-based INR branches inserted between adjacent scales via a shared encoder in a closed loop, plus a BFPU that propagates features coarse-to-fine and fine-to-coarse.
Key results
- NeRD-Rain reaches 41.71 dB PSNR on Rain200L, above DRSformer's 41.23 dB.
- On the real SPA-Data benchmark, NeRD-Rain reports 49.58 dB PSNR, 1.04 dB above DRSformer.
- Ablation shows the BFPU gives a 0.10 dB PSNR gain over unidirectional propagation on Rain200H.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the INR branch (ablation) | bidirectional multi-scale Transformer · Rain200L | PSNR: 41.40 -> 41.71 | Helped |
| place INR between adjacent scales rather than within the Transformer branch (ablation) | NeRD-Rain · Rain200L | PSNR: 41.51 -> 41.71 | Helped |
| use a multi-scale rather than fixed-scale feature grid (ablation) | NeRD-Rain INR · Rain200L | PSNR: 41.47 -> 41.71 | Helped |
| add position encoding (ablation) | NeRD-Rain INR · Rain200L | PSNR: 41.50 -> 41.71 | Helped |
| add interpolation (ablation) | NeRD-Rain INR · Rain200L | PSNR: 41.58 -> 41.71 | Helped |
| use a shared rather than separate encoder (ablation) | NeRD-Rain INR · Rain200L | PSNR: 41.62 -> 41.71 | Helped |
| use 1, 2, and 3 UNets at scales S1, S2, and S3 rather than 2 at every scale (ablation) | multi-scale Transformer · Rain200L | PSNR: 41.42 -> 41.71 | Helped |
| use 1, 2, and 3 UNets at scales S1, S2, and S3 rather than 3, 2, and 1 (ablation) | multi-scale Transformer · Rain200L | PSNR: 40.70 -> 41.71 | Helped |
| add one UNet at scale S1 while keeping 2 and 3 at S2 and S3 (ablation) | multi-scale Transformer · Rain200L | PSNR: 41.63 -> 41.71 | Helped |
| add bidirectional feedback propagation rather than remove it (ablation) | NeRD-Rain · Rain200H | PSNR: 32.30 -> 32.40 | Helped |
| use BFPU rather than direct feature concatenation (ablation) | NeRD-Rain · Rain200H | PSNR: 32.31 -> 32.40 | Helped |
| use a CNN-based U-Net architecture in the proposed framework (ablation) | NeRD-Rain · Rain200L | PSNR: 39.47 -> 41.34 | Helped |
Surprising findings
- Bilinear-interpolation-based implicit neural interpolation in the INR branch naturally suppresses white rain streaks, acting as a deraining effect without extra operations.
Limitations
- Training time is longer than other multi-scale architectures due to INR optimization.
- Fails to handle veiling effect (mist) in complex rainy environments.
Experimental setup
Datasets: Rain200L Rain200H DID-Data DDN-Data SPA-Data RE-RAIN
Benchmarks: Rain200L Rain200H DID-Data DDN-Data SPA-Data RE-RAIN
Metrics: PSNR SSIM FLOPs parameters
Baselines: DSC GMM DDN RESCAN PReNet MSPFN RCDNet MPRNet DualGCN SPDNet Uformer Restormer
Builds on: Restormer NeRF LIIF MPRNet MSPFN
Closest papers by meaning
- Rethinking Multi-Scale Representations in Deep Deraining Transformer (AAAI 2024 · similarity 0.91)
- Multi-Scale Progressive Fusion Network for Single Image Deraining (CVPR 2020 · similarity 0.91)
- Harnessing Joint Rain-/Detail-aware Representations to Eliminate Intricate Rains (ICLR 2024 · similarity 0.91)
- Hybrid CNN-Transformer Feature Fusion for Single Image Deraining (AAAI 2023 · similarity 0.91)
- Boosting Image De-Raining via Central-Surrounding Synergistic Convolution (AAAI 2025 · similarity 0.89)
- FMRNet: Image Deraining via Frequency Mutual Revision (AAAI 2024 · similarity 0.87)
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.