Research map / Image generation / Image restoration
Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration
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 WF-Diff, a two-stage underwater image enhancement framework that combines a wavelet/Fourier-based network (WFI2-net) with a wavelet-space residual diffusion module (FRDAM) that refines high- and low-frequency components.
Problem
Underwater image enhancement mostly operates in raw pixel space, underexploiting frequency characteristics and the representational power of deep and diffusion models, leading to poor detail and color restoration.
Approach
Decompose images with Haar DWT, restore low-frequency amplitude spectra (Fourier) and high-frequency sub-bands (Transformer) in a parallel U-Net, then refine residuals of each sub-band with separate diffusion models in wavelet space, linked by a cross-frequency conditioner.
Key results
- On UIEBD, WF-Diff reaches 23.86 PSNR and 0.8730 SSIM, versus 21.88 PSNR for the next-best DM-water.
- On LSUI, WF-Diff achieves 27.26 PSNR and 0.9437 SSIM, with the best SSIM and FID (26.75).
- Inference time is 0.2816 s for WF-Diff versus 0.1441 s for DM-water, so the method is slower than recent approaches.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use WF-Diff instead of DM-water | Underwater image restoration · UIEBD | PSNR: 21.88 -> 23.86 | Helped |
| Use WF-Diff instead of U-shape | Underwater image restoration · LSUI | LPIPS: 0.1028 -> 0.1096 | Hurt |
| Use WF-Diff instead of DM-water | Underwater image restoration · LSUI | PSNR: 27.65 -> 27.26 | Hurt |
| Use WF-Diff instead of U-shape | Underwater image restoration · U45 | UIQM: 3.151 -> 3.181 | Helped |
| Add self-attention to WTB (ablation) | WFI2-net · UIEBD | PSNR: 20.94 -> 21.87 | Helped |
| Add channel attention to WTB (ablation) | WFI2-net · UIEBD | PSNR: 20.82 -> 21.87 | Helped |
| Add the spatial domain unit to SFFB (ablation) | WFI2-net · UIEBD | PSNR: 21.11 -> 21.87 | Helped |
| Add the frequency domain unit to SFFB (ablation) | WFI2-net · UIEBD | PSNR: 20.23 -> 21.87 | Helped |
| Add high-frequency loss Lh (ablation) | WFI2-net · UIEBD | PSNR: 20.46 -> 21.87 | Helped |
| Add reconstruction loss Lrec (ablation) | WFI2-net · UIEBD | PSNR: 20.65 -> 21.87 | Helped |
| Add amplitude loss La (ablation) | WFI2-net · UIEBD | PSNR: 19.81 -> 21.87 | Helped |
| Add the cross-frequency conditioner (ablation) | WFI2-net · UIEBD | PSNR: 20.97 -> 21.87 | Helped |
Surprising findings
- Swapping only low-frequency wavelet amplitudes (S2) gave worse PSNR (27.10) than swapping amplitude in the raw Fourier domain (S1, 28.68), while swapping both low and high sub-band amplitudes (S3) gave the best result (29.97).
Limitations
- Two diffusion models increase inference time, making it slower than recent approaches.
- Evaluation is limited to UIEBD, LSUI and U45 with full-reference metrics on a small set of images.
- Reported inference uses 10 implicit sampling steps, and the paper does not explore faster sampling.
Experimental setup
Datasets: UIEBD LSUI
Benchmarks: UIEBD LSUI U45
Metrics: PSNR SSIM LPIPS FID UIQM UCIQE
Baselines: UIEWD UWCNN UIEC^2-Net Water-Net SCNet U-color U-shape DM-water
Builds on: DDPM Palette Wasserstein GAN Haar wavelet transform
Closest papers by meaning
- Adaptive Dual-domain Learning for Underwater Image Enhancement (AAAI 2025 · similarity 0.90)
- UVEB: A Large-scale Benchmark and Baseline Towards Real-World Underwater Video Enhancement (CVPR 2024 · similarity 0.87)
- WaterDiffusion: Learning a Prior-involved Unrolling Diffusion for Joint Underwater Saliency Detection and Visual Restoration (AAAI 2025 · similarity 0.87)
- Synergistic Multiscale Detail Refinement via Intrinsic Supervision for Underwater Image Enhancement (AAAI 2024 · similarity 0.86)
- Atlantis: Enabling Underwater Depth Estimation with Stable Diffusion (CVPR 2024 · similarity 0.83)
- AdaIR: Adaptive All-in-One Image Restoration via Frequency Mining and Modulation (ICLR 2025 · similarity 0.83)
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.