Research map / Image generation / Image restoration
FMRNet: Image Deraining via Frequency Mutual Revision
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 FMRNet, a wavelet-domain network that jointly models low-frequency rain residue, full rain residue, and background through multi-level mutual revision for single-image deraining.
Problem
Single image deraining suffers from aliasing between rain residue and background, which spatial-domain and existing wavelet methods handle poorly, causing structure and texture loss.
Approach
Learns mutual representations of rain residue and background in the Haar wavelet frequency domain, using predicted low-frequency rain distribution as a degradation prior and feeding refined rain residue back to update that prior, via cascaded multi-level mutuality fusion modules (MMFM).
Key results
- Surpasses ELFormer by 1.14 dB PSNR on Rain100L (R100L) with similar computation cost (91.23 GFLOPs vs 66.39).
- Achieves 33.09 average PSNR across four synthetic datasets versus 32.74 for ELFormer and 32.12 for DAWN.
- Ablation: removing the multi-level mutuality module (w/o MMF) drops Test1200 PSNR by 0.87 dB to 32.34 from 33.21.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply FMRNet instead of ELFormer | image deraining · R100L | PSNR: 36.67 -> 37.81 | Helped |
| apply FMRNet instead of ELFormer | image deraining · Test1200 | PSNR: 33.38 -> 33.21 | Hurt |
| apply FMRNet instead of ELFormer | image deraining · R100H | PSNR: 30.48 -> 30.69 | Helped |
| remove self-attention from FEM (ablation) | FMRNet · Test1200 | PSNR: 32.98 -> 33.21 | Helped |
| add low-frequency prior guidance (ablation) | FMRNet · Test1200 | PSNR: 33.01 -> 33.21 | Helped |
| add mutual learning between rain residue and background (ablation) | FMRNet · Test1200 | PSNR: 32.63 -> 33.21 | Helped |
| add low-frequency prior guidance and mutual learning (ablation) | FMRNet · Test1200 | PSNR: 32.34 -> 33.21 | Helped |
| use depth-wise separable convolutions instead of standard convolutions (ablation) | FMRNet · Test1200 | PSNR: 32.96 -> 33.21 | Helped |
| add the Y-channel structure loss (ablation) | FMRNet · Test1200 | PSNR: 33.06 -> 33.21 | Helped |
Limitations
- Poor generalization to hybrid degradation such as heavy rain veiling effect.
- Performance on heavy rain relative to the strongest baselines is not uniformly best on real-world NIQE/SSEQ metrics.
Experimental setup
Datasets: Test100 Test1200 R100H R100L RID RIS Real127 13,700 clean/rain pairs (Zhang, Sindagi, and Patel 2020; Fu et al. 2017b)
Benchmarks: Test100 Test1200 R100H R100L RID RIS Real127
Metrics: PSNR SSIM NIQE SSEQ parameters inference time GFLOPs
Baselines: IADN MSPFN DRDNet MPRNet SWAL DANet ELFormer DAWN PreNet
Builds on: DANet ELFormer DAWN MPRNet Haar wavelet transform
Closest papers by meaning
- FourierMamba: Fourier Learning Integration with State Space Models for Image Deraining (ICML 2025 · similarity 0.89)
- Bidirectional Multi-Scale Implicit Neural Representations for Image Deraining (CVPR 2024 · similarity 0.87)
- Rethinking Multi-Scale Representations in Deep Deraining Transformer (AAAI 2024 · similarity 0.87)
- Harnessing Joint Rain-/Detail-aware Representations to Eliminate Intricate Rains (ICLR 2024 · similarity 0.85)
- Towards Universal Rainy Image Restoration: Benchmark and Baseline (AAAI 2025 · similarity 0.84)
- MODEM: A Morton-Order Degradation Estimation Mechanism for Adverse Weather Image Recovery (NeurIPS 2025 · similarity 0.84)
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.