Research map / Image generation / Image restoration
NightHaze: Nighttime Image Dehazing via Self-Prior Learning
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 NightHaze, a nighttime dehazing method that pretrains an MAE-style encoder-decoder on severely augmented clear images (glow blending plus strong noise) and refines it with a teacher-student scheme gated by no-reference IQA scores.
Problem
Nighttime image dehazing on real-world haze images, where synthetic-to-real domain gaps and the absence of paired real data lead existing methods to produce poor results.
Approach
Adapts MAE's "severe augmentation" idea to nighttime dehazing by replacing masking with light-map blending and strong Gaussian noise, so the network learns clear-background priors. A self-refinement stage uses a teacher-student setup on unlabeled real haze images, accepting student updates only when a no-reference IQA score improves.
Key results
- On RealNightHaze, NightHaze reaches 52.87 MUSIQ and 76.88 TRES, versus 45.76 and 65.83 for the best prior nighttime dehazing method NightDeFog.
- Ablation: increasing the fraction of severely augmented training data from 0% to 100% raises MUSIQ from 44.67 to 52.87 and CLIPIQA from 0.5364 to 0.6360.
- The paper reports 15.5% gain in MUSIQ and 23.5% gain in ClipIQA over existing nighttime dehazing methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use 0% severe augmentation instead of severe augmentation (ablation) | NightHaze self-prior learning · RealNightHaze | MUSIQ: 52.87 -> 44.67 | Hurt |
| Use 25% severe augmentation instead of 100% severe augmentation (ablation) | NightHaze self-prior learning · RealNightHaze | MUSIQ: 52.87 -> 48.64 | Hurt |
| Use 50% severe augmentation instead of 100% severe augmentation (ablation) | NightHaze self-prior learning · RealNightHaze | MUSIQ: 52.87 -> 49.12 | Hurt |
| Use 75% severe augmentation instead of 100% severe augmentation (ablation) | NightHaze self-prior learning · RealNightHaze | MUSIQ: 52.87 -> 49.31 | Hurt |
| Increase the proportion of severe augmentation in training (ablation) | NightHaze self-prior learning · RealNightHaze | TRES: 62.18 -> 67.96 -> 68.74 -> 68.92 -> 76.88 | Helped |
Surprising findings
- Backbones trained on synthetic daytime-style data with the same architecture perform worse than the original haze images on several no-reference metrics (e.g., MUSIQ 44.25 vs 42.12 for haze images is better, but syn-training lags the clear-image statistics substantially).
Limitations
- Evaluation relies on no-reference IQA metrics rather than ground truth, which may not reflect true dehazing quality.
- Over-suppression artifacts (e.g., light sources) remain after self-prior learning.
- Self-refinement relies on IQA scores as a gate, and the choice of thresholds (0.005, 0) is set empirically.
Experimental setup
Datasets: RealNightHaze
Benchmarks: RealNightHaze
Metrics: MUSIQ TRES Hyperiqa CLIPIQA MANIQA contrast
Baselines: NightDeFog NightVDM NightEnhance Restormer Uformer WeatherDiffusion DiT
Builds on: masked autoencoder mean teacher
Closest papers by meaning
- VQCNIR: Clearer Night Image Restoration with Vector-Quantized Codebook (AAAI 2024 · similarity 0.87)
- NightRain: Nighttime Video Deraining via Adaptive-Rain-Removal and Adaptive-Correction (AAAI 2024 · similarity 0.86)
- PromptHaze: Prompting Real-world Dehazing via Depth Anything Model (AAAI 2025 · similarity 0.86)
- SelfPromer: Self-Prompt Dehazing Transformers with Depth-Consistency (AAAI 2024 · similarity 0.86)
- See through the Dark: Learning Illumination-affined Representations for Nighttime Occupancy Prediction (NeurIPS 2025 · similarity 0.86)
- Learning Hazing to Dehazing: Towards Realistic Haze Generation for Real-World Image Dehazing (CVPR 2025 · similarity 0.86)
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.