Research map / Image generation / Image restoration
Adaptive Dual-domain Learning for Underwater Image Enhancement
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 SS-UIE, a spatial-spectral dual-domain underwater image enhancement network built from parallel Mamba-based spatial (MCSS) and FFT-based spectral (SWSA) blocks, plus a frequency-wise loss (FWL).
Problem
Learning-based underwater image enhancement treats all spatial regions and spectral bands equally despite inconsistent degradation, and quadratic-cost Transformers are expensive for dense restoration.
Approach
Combine a Mamba-based multi-scale cycle selective-scan spatial branch and a learnable global-filter FFT spectral branch in parallel (SS-block), both with linear complexity, and add a frequency-weighted loss that emphasizes hard high-frequency content.
Key results
- On LSUI, SS-UIE reaches 28.87 PSNR and 0.952 SSIM, versus 27.65 PSNR for the next best (DM-Water) and 27.26 for WF-Diff.
- On UIEB, SS-UIE achieves 24.19 PSNR and 0.897 SSIM, outperforming WF-Diff (23.86 PSNR, 0.873 SSIM).
- On the non-reference U45 set, SS-UIE achieves the best UIQM (3.246) and UCIQE (0.703).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SS-UIE instead of WF-Diff | UIE methods · UIEB test | PSNR: 23.86 -> 24.19 | Helped |
| Use SS-UIE instead of WF-Diff | UIE methods · UIEB test | SSIM: 0.873 -> 0.897 | Helped |
| Use SS-UIE instead of DM-Water | UIE methods · LSUI test | PSNR: 27.65 -> 28.87 | Helped |
| Use SS-UIE instead of WF-Diff | UIE methods · U45 | UCIQE: 0.619 -> 0.703 | Helped |
| Add MCSS instead of using SS2D (ablation) | BL on LSUI · LSUI test | PSNR: 23.88 -> 24.47 | Helped |
| Add MCSS to the baseline (ablation) | BL on LSUI · LSUI test | PSNR: 22.67 -> 24.47 | Helped |
| Add SWSA to the baseline (ablation) | BL on LSUI · LSUI test | PSNR: 22.67 -> 24.03 | Helped |
| Add SWSA alongside MCSS (ablation) | BL + MCSS on LSUI · LSUI test | PSNR: 24.47 -> 28.87 | Helped |
| Combine MCSS and SWSA in parallel instead of serially (ablation) | SS-block on LSUI · LSUI test | PSNR: 26.87 -> 28.13 | Helped |
| Add FWL (ablation) | parallel MCSS + SWSA model on LSUI · LSUI test | PSNR: 28.13 -> 28.87 | Helped |
| Add FWL (ablation) | parallel MCSS + SWSA model on LSUI · LSUI test | SSIM: 0.938 -> 0.952 | Helped |
| Combine MCSS and SWSA in parallel instead of serially (ablation) | SS-block on LSUI · model parameters | parameter count: 5.78M -> 4.25M | Helped |
Surprising findings
- The ablation reports the parallel SS-block design is cheaper (4.25M params, 19.37G FLOPs) than the serial design (5.78M, 28.95G) while performing better.
Limitations
- Evaluation is limited to UIEB, LSUI and U45 with paired training on small datasets.
- Frequency loss weight lambda is set heuristically to match magnitudes of L1 and FWL terms.
Experimental setup
Datasets: UIEB LSUI
Benchmarks: UIEB LSUI U45
Metrics: PSNR SSIM LPIPS FID UIQM UCIQE LFD FLOPs Params
Baselines: Fusion MLLE Water-Net UGAN U-Color U-Shape DM-Water WF-Diff
Builds on: Mamba VMamba Vision-Mamba Swin transformer F-Principle
Closest papers by meaning
- Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration (CVPR 2024 · similarity 0.90)
- UVEB: A Large-scale Benchmark and Baseline Towards Real-World Underwater Video Enhancement (CVPR 2024 · similarity 0.89)
- Synergistic Multiscale Detail Refinement via Intrinsic Supervision for Underwater Image Enhancement (AAAI 2024 · similarity 0.87)
- WaterDiffusion: Learning a Prior-involved Unrolling Diffusion for Joint Underwater Saliency Detection and Visual Restoration (AAAI 2025 · similarity 0.85)
- Atlantis: Enabling Underwater Depth Estimation with Stable Diffusion (CVPR 2024 · similarity 0.85)
- NAUTILUS: A Large Multimodal Model for Underwater Scene Understanding (NeurIPS 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.