Research map / Image generation / Image restoration
Beyond Spatial Domain: Cross-domain Promoted Fourier Convolution Helps Single Image Dehazing
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 JSFC-Net, a dehazing network whose FS-PPL blocks run a Fourier-domain learnable-kernel branch promoted by spatial features in parallel with a physics-based spatial branch, fused by a receptive field selection module, plus a frequency loss.
Problem
Single image dehazing is limited by small receptive fields of convolution and window attention, and by poor learning of high-frequency details in spatial-only models.
Approach
Applies learnable multi-scale convolution kernels (generated from spatial features) in the FFT domain to capture global frequency features, fuses them with a physics-aware spatial branch via per-pixel receptive field selection, and adds an L1 loss on the Fourier spectrum to narrow frequency gaps.
Key results
- JSFC-Net reaches 41.32 dB PSNR on SOTS-Indoor and 37.84 dB on SOTS-Outdoor, the highest among compared methods (0.50 dB above FocalNet indoors).
- On O-Haze, JSFC-Net achieves 25.72 dB PSNR / 0.804 SSIM, 0.24 dB above the second-best; on NH-HAZE it reaches 21.42 dB PSNR.
- Adding frequency loss improves PSNR of nine SOTA dehazing methods, e.g., FFA-Net +1.19 dB and AOD-Net +1.94 dB, without extra inference parameters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the Frequency Promoted Learning Branch (ablation) | JSFC-Net · RESIDE SOTS-Indoor | PSNR: 38.68 -> 41.32 | Helped |
| Add the Spatial Physical Learning Branch (ablation) | JSFC-Net · RESIDE SOTS-Indoor | PSNR: 21.44 -> 41.32 | Helped |
| Use spatially generated learnable kernels instead of randomly initialized kernels (ablation) | JSFC-Net · RESIDE SOTS-Indoor | PSNR: 40.58 -> 41.32 | Helped |
| Use multi-scale rather than single-scale learnable kernels (ablation) | JSFC-Net · RESIDE SOTS-Indoor | PSNR: 41.21 -> 41.32 | Helped |
| Add Mixture-of-Experts fusion (ablation) | JSFC-Net · RESIDE SOTS-Indoor | PSNR: 41.17 -> 41.32 | Helped |
| Add the Receptive Field Selection Module (ablation) | JSFC-Net · RESIDE SOTS-Indoor | PSNR: 40.99 -> 41.32 | Helped |
| Add frequency loss during training | SOTA dehazing methods · RESIDE | PSNR: ↑1.94; ↑0.47; ↑0.09; ↑0.22; ↑1.19; ↑0.30; ↑0.19; ↑0.08; ↑0.11 | Helped |
Surprising findings
- Frequency-domain features alone perform poorly (removing the spatial physical branch drops SOTS-Indoor PSNR to 21.44 dB), so frequency features work best as a supplement to spatial features.
Limitations
- Frequency-domain branch alone is insufficient, so performance depends on the spatial branch.
- Evaluated only on synthetic and generated haze datasets with few real-world images (qualitative only).
Experimental setup
Datasets: RESIDE NH-HAZE O-Haze
Benchmarks: RESIDE SOTS-Indoor RESIDE SOTS-Outdoor NH-HAZE O-Haze
Metrics: PSNR SSIM parameters (M) inference time (ms)
Baselines: DCP DehazeNet GridDehazeNet MSBDN FFA-Net AECR-Net MAXIM-2S DeHamer DehazeFormer-M MITNet FocalNet MB-TaylorFormer-B
Builds on: PDU (physics-aware dual-branch unit) Fourier Unit (Chi et al. 2020) FFA-Net DehazeFormer Atmospheric Scattering Model
Closest papers by meaning
- Dual-Domain Attention for Image Deblurring (AAAI 2023 · similarity 0.89)
- Intriguing Findings of Frequency Selection for Image Deblurring (AAAI 2023 · similarity 0.89)
- FourierMamba: Fourier Learning Integration with State Space Models for Image Deraining (ICML 2025 · similarity 0.86)
- Exploiting Diffusion Prior for Real-World Image Dehazing with Unpaired Training (AAAI 2025 · similarity 0.86)
- DeblurDiff: Real-Word Image Deblurring with Generative Diffusion Models (NeurIPS 2025 · similarity 0.85)
- Fourier Priors-Guided Diffusion for Zero-Shot Joint Low-Light Enhancement and Deblurring (CVPR 2024 · similarity 0.85)
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.