Research map / Image generation / Image restoration
Complementary Advantages: Exploiting Cross-Field Frequency Correlation for NIR-Assisted Image Denoising
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 FCENet, a two-stage frequency-domain network that uses a measured NIR-RGB frequency correlation prior to dynamically select and fuse complementary NIR and RGB features for NIR-assisted image denoising.
Problem
NIR-assisted RGB image denoising is hard because NIR and RGB images are inconsistent in color, brightness, and structure, and existing fusion methods either fuse indiscriminately or discard differential information (e.g., NIR sharp textures and RGB colors).
Approach
A cross-field frequency correlation prior (NIR is more similar to clean RGB at high frequencies, noisy RGB at low frequencies) motivates a frequency dynamic selection mechanism (FDSM) that filters each field's features with input-conditioned frequency filters, and a frequency exhaustive fusion mechanism (FEFM) combining frequency-domain cross-attention for common features with differential cross-a
Key results
- On DVD at sigma=4, FCENet reaches 30.37 dB PSNR / 0.950 SSIM versus 29.81 dB for NIR-Restormer, the best prior NIR-assisted method.
- On IVRG at sigma=50, FCENet achieves 29.08 dB PSNR versus 28.77 dB for NIR-Restormer.
- Ablation on DVD shows FDSM adds 0.51 dB and FEFM adds 0.83 dB over a two-stage U-Net baseline with addition fusion.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FCENet instead of NIR-Restormer | NIR-assisted denoising · DVD test, σ = 4 | PSNR: 29.81 -> 30.37 | Helped |
| Add FDSM to the two-stage U-Net baseline (ablation) | FCENet · DVD test, σ = 4 | PSNR: 29.13 -> 29.64 | Helped |
| Add FEFM to the two-stage U-Net baseline (ablation) | FCENet · DVD test, σ = 4 | PSNR: 29.13 -> 29.96 | Helped |
| Add both FDSM and FEFM (ablation) | FCENet · DVD test, σ = 4 | PSNR: 29.13 -> 30.37 | Helped |
| Use frequency dynamic selection instead of spatial dynamic selection (ablation) | FCENet FDSM · DVD test, σ = 4 | PSNR: 30.11 -> 30.37 | Helped |
| Add FDSM rather than remove it (ablation) | FCENet · DVD test, σ = 4 | PSNR: 29.96 -> 30.37 | Helped |
| Replace direct feature summation with cross-attention (ablation) | FCENet FEFM · DVD test, σ = 4 | PSNR: 29.64 -> 29.85 | Helped |
| Use DFR instead of direct feature summation (ablation) | FCENet FEFM · DVD test, σ = 4 | PSNR: 29.64 -> 30.15 | Helped |
| Add CFR to DFR (ablation) | FCENet FEFM · DVD test, σ = 4 | PSNR: 30.15 -> 30.37 | Helped |
Surprising findings
- Frequency-domain dynamic filters learned on the NIR side emphasize mid-to-high frequencies while RGB-side filters focus on low frequencies, matching the correlation prior.
- Spatial dynamic selection is less effective than frequency dynamic selection in the ablation (30.11 vs 30.37 dB).
Limitations
- The authors note existing methods, including this one, may fail on significantly misaligned NIR-RGB pairs.
- Evaluation relies on simulated noise for DVD and IVRG, with only qualitative results on real low-light data.
Experimental setup
Datasets: DVD IVRG
Benchmarks: DVD IVRG
Metrics: PSNR SSIM FLOPs Params
Baselines: Uformer Restormer MPRNet Xformer CUNet SANet MNNet DVN NIR-Restormer
Builds on: Restormer SAM (supervised attention module from MPRNet) Charbonnier loss frequency domain loss FSAS frequency self-attention
Closest papers by meaning
- Efficient RAW Image Deblurring with Adaptive Frequency Modulation (NeurIPS 2025 · similarity 0.85)
- Zero-Shot Blind-spot Image Denoising via Implicit Neural Sampling (CVPR 2025 · similarity 0.85)
- Robust Image Denoising through Adversarial Frequency Mixup (CVPR 2024 · similarity 0.85)
- APR-RD: Complemental Two Steps for Self-Supervised Real Image Denoising (AAAI 2025 · similarity 0.85)
- Interpretable Unsupervised Joint Denoising and Enhancement for Real-World low-light Scenarios (ICLR 2025 · similarity 0.85)
- FDS: Frequency-Aware Denoising Score for Text-Guided Latent Diffusion Image Editing (CVPR 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.