Research map / Image generation / Image restoration
Rethinking Nighttime Image Deraining via Learnable Color Space Transformation
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.
Introduces HQ-NightRain, an illumination-aware synthetic nighttime rain dataset, and CST-Net, a two-stage network that derains in a learnable YCbCr-like Y channel with implicit illumination guidance.
Problem
Nighttime image deraining suffers from synthetic datasets that add rain uniformly and ignore how rain visibility depends on artificial illumination, and existing methods operate in RGB space.
Approach
Synthesize rain masks modulated by a background V-channel illumination matrix (with thresholding and defocus blur), and perform rain removal in a Y luminance channel produced by a learnable color space converter, with an implicit neural illumination guidance module between two stages.
Key results
- CST-Net achieves 39.07 PSNR average on HQ-NightRain/GTAV-NightRain versus 38.51 for NeRD-Rain (the second-best average in the table).
- On GTAV-NightRain, CST-Net reaches 38.94 dB PSNR, 1.1 dB above NeRD-Rain as reported in the text.
- Ablation: replacing learnable CSC with fixed YCbCr drops PSNR from 40.50 to 39.60 on HQ-NightRain.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the learnable CSC instead of using fixed color-space conversion (ablation) | CST-Net · HQ-NightRain | PSNR: 39.5959 -> 40.4984 | Helped |
| use YCbCr rather than RGB, HSV, HSL, or YUV (ablation) | CST-Net without IIG · HQ-NightRain | PSNR: 38.7507; 39.0317; 39.1613; 39.1932 -> 39.5959 | Helped |
| add IIG (ablation) | CST-Net · HQ-NightRain | PSNR: 39.8767 -> 40.4984 | Helped |
| use both network stages instead of either one alone (ablation) | CST-Net with YCbCr, learnable CSC, and no IIG · HQ-NightRain | PSNR: 35.0858; 36.4385 -> 39.8767 | Helped |
| use the standard stage module combinations rather than swapping them (ablation) | CST-Net · HQ-NightRain | PSNR: 39.0333 -> 40.4984 | Helped |
| use LN rather than BN (ablation) | CST-Net · HQ-NightRain | PSNR: 38.8534 -> 40.4984 | Helped |
| use the standard module counts rather than larger counts (ablation) | CST-Net · HQ-NightRain | PSNR: 39.7594; 40.1056 -> 40.4984 | Helped |
| use illumination-fused dataset synthesis instead of linear addition (ablation) | CST-Net trained on the corresponding synthesized data · HQ-NightRain | PSNR: 25.38 -> 36.55 | Helped |
| use defocus blur in the synthesis pipeline (ablation) | CST-Net trained on synthesized data · HQ-NightRain | PSNR: 36.55 -> 31.91 | Hurt |
| set illumination thresholds to (0.2, 0.8) rather than other tested settings (ablation) | CST-Net dataset synthesis · HQ-NightRain RS | PSNR: 40.78 / 41.91 / 42.13 -> 42.89 | Helped |
| train the same model on HQ-NightRain rather than GTAV-NightRain or RoadScene-rain | IDT · RealRain-1k-L | PSNR: 26.47; 25.63 -> 26.94 | Helped |
| apply CST-Net rather than the strongest listed baseline | Nighttime deraining · GTAV-NightRain | PSNR: 38.1271 -> 38.9378 | Helped |
Surprising findings
- Restormer and SFNet are competitive with the proposed method on several subsets, and the dataset-pipeline ablation shows a variant D3 (no defocus blur) scoring higher PSNR (36.55) than the full pipeline (31.91) despite the full pipeline being chosen.
Limitations
- Synthetic training data relies on hand-designed illumination thresholds (0.2, 0.8) chosen empirically.
- Real-world evaluation uses non-reference metrics and small subsets (20 natural-scene pairs, 512 real images).
- Downstream detection gains are measured on a single detector (YOLOv8) with modest absolute improvements.
Experimental setup
Datasets: HQ-NightRain GTAV-NightRain RealRain-1k RainDS-real BDD100K BDD350-Night Muti-Weather6k
Benchmarks: HQ-NightRain GTAV-NightRain RealRain-1k RainDS-real BDD350-Night Muti-Weather6k
Metrics: PSNR SSIM LPIPS PaQ-2-PiQ MANIQA precision recall IoU
Baselines: PReNet RCDNet SPDNet MPRNet IDT Restormer SFNet DRSformer RLP MSGNN NeRD-Rain
Builds on: Restormer IDT NeRD-Rain
Closest papers by meaning
- NightRain: Nighttime Video Deraining via Adaptive-Rain-Removal and Adaptive-Correction (AAAI 2024 · similarity 0.91)
- NightCC: Nighttime Color Constancy via Adaptive Channel Masking (CVPR 2024 · similarity 0.88)
- Exploring the Common Appearance-Boundary Adaptation for Nighttime Optical Flow (ICLR 2024 · similarity 0.87)
- See through the Dark: Learning Illumination-affined Representations for Nighttime Occupancy Prediction (NeurIPS 2025 · similarity 0.84)
- DNIT: Enhancing Day-Night Image-to-Image Translation through Fine-Grained Feature Handling (Student Abstract) (AAAI 2024 · similarity 0.84)
- Harnessing Joint Rain-/Detail-aware Representations to Eliminate Intricate Rains (ICLR 2024 · 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.