Research map / Label-efficient and robust learning / Semi-supervised learning
A Semi-supervised Nighttime Dehazing Baseline with Spatial-Frequency Aware and Realistic Brightness Constraint
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 SFSNiD, a semi-supervised nighttime dehazing network with a spatial-frequency information interaction module, pseudo-label retraining, and a local-window brightness loss for realistic brightness on real hazy images.
Problem
Nighttime image dehazing is hard because haze, glow and noise are localized, coupled and frequency-inconsistent, and models trained on simulated data produce unrealistic brightness on real-world nighttime images.
Approach
Combine spatial local attention with dynamic amplitude/phase spectrum filtering in a bidomain interaction block (SFII), and train semi-supervised with pseudo labels from game-engine-trained model plus a window-based brightness loss that ties dehazed brightness to real hazy images scaled by a coefficient kappa.
Key results
- On UNREAL-NH, the method reports SSIM 0.862 and PSNR 25.907, versus 0.827/25.680 for the next best MSBDN.
- On the real-world RWNH set, BRISQUE 30.975 is best (lower is better), while MUSIQ 32.120 is below DeHamer's 26.788 and GD's 33.433 among shown methods only partially; the method is best on BRISQUE.
- Ablation shows the frequency-domain loss raises SSIM from 0.816 to 0.862 on UNREAL-NH.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add frequency domain projection to SFII (ablation) | SFSNiD · UNREAL-NH | SSIM: 0.848 -> 0.862 | Helped |
| Add local perception to SFII (ablation) | SFSNiD · UNREAL-NH | SSIM: 0.858 -> 0.862 | Helped |
| Add frequency processing in BNM (ablation) | SFSNiD · UNREAL-NH | SSIM: 0.851 -> 0.862 | Helped |
| Add spatial processing in BNM (ablation) | SFSNiD · UNREAL-NH | SSIM: 0.845 -> 0.862 | Helped |
| Use hierarchical training at scales s∈{0,1,2} instead of s∈{0} (ablation) | SFSNiD · UNREAL-NH | SSIM: 0.854 -> 0.862 | Helped |
| Use hierarchical training at scales s∈{0,1,2} instead of s∈{0,1} (ablation) | SFSNiD · UNREAL-NH | SSIM: 0.851 -> 0.862 | Helped |
| Add frequency-domain loss (ablation) | SFSNiD · UNREAL-NH | SSIM: 0.816 -> 0.862 | Helped |
| Add realistic brightness loss to pseudo-label retraining (ablation) | SFSNiD · RWNH | BRISQUE: 34.210 -> 30.975 | Helped |
| Add realistic brightness loss to pseudo-label retraining (ablation) | SFSNiD · RWNH | MUSIQ: 32.373 -> 32.120 | Hurt |
| Set brightness intensity coefficient κ to 1.3 rather than 1.0 (ablation) | SFSNiD · RWNH | average pixel value: 0.193 -> 0.225 | Mixed |
Surprising findings
- Retraining with pseudo labels alone yields unrealistic brightness on real images; the brightness loss is needed to fix it.
Limitations
- Real-world brightness target relies on a global brightness statistic and a manually chosen kappa (1.3).
- Real-world evaluation uses no-reference metrics, which the authors note are less reliable for dehazing.
Experimental setup
Datasets: GTA5 UNREAL-NH NHR NHM NHCL NHCM NHCD NightHaze YellowHaze RWNH
Benchmarks: NHR NHM NHCL NHCM NHCD NightHaze YellowHaze GTA5 UNREAL-NH RWNH
Metrics: PSNR SSIM BRISQUE MUSIQ
Baselines: MRP GD OSFD MSBDN 4KDehazing AECRNet DeHamer FSDGN DF MITNet Fourmer
Builds on: Fourier transform Swin-style window attention atmospheric scattering model
Closest papers by meaning
- NightHaze: Nighttime Image Dehazing via Self-Prior Learning (AAAI 2025 · similarity 0.81)
- NightCC: Nighttime Color Constancy via Adaptive Channel Masking (CVPR 2024 · similarity 0.80)
- The Parables of the Mustard Seed and the Yeast: Extremely Low-Budget, High-Performance Nighttime Semantic Segmentation (AAAI 2025 · similarity 0.80)
- See through the Dark: Learning Illumination-affined Representations for Nighttime Occupancy Prediction (NeurIPS 2025 · similarity 0.80)
- NightAdapter: Learning a Frequency Adapter for Generalizable Night-time Scene Segmentation (CVPR 2025 · similarity 0.80)
- NightRain: Nighttime Video Deraining via Adaptive-Rain-Removal and Adaptive-Correction (AAAI 2024 · similarity 0.79)
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.