Research map / Image generation / Image restoration
Synergistic Multiscale Detail Refinement via Intrinsic Supervision 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 SMDR-IS, a multi-degradation encoder-decoder for underwater image enhancement that feeds downscaled inputs into each UNet stage and uses intrinsic-supervised feature selection (ASISF) and bifocal attention (BICA) to refine multi-scale details.
Problem
Underwater image enhancement suffers from scattering and absorption degradations that manifest across scales, but existing methods ignore scale-related characteristics of underwater scenes.
Approach
Uses four resolution stages of inputs in a UNet-style encoder-decoder, with an ASISF module that gates feature transfer from low-resolution stages to the original resolution under intrinsic supervision, and BICA combining a resolution-guided attention branch (ReGIA) with a hierarchical context branch (HCAFE).
Key results
- On UIEB validation, SMDR-IS reaches PSNR 23.710, SSIM 0.922, and ALL 60.466, the best among compared methods (UDA ALL 58.374).
- On LSUI, SMDR-IS achieves PSNR 21.984 and ALL 57.368, best ALL among methods reported.
- Ablation: going from one to four stages raises PSNR from 22.790 to 23.710 on UIEB test.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add En_to_DE (ablation) | SMDR-IS · UIEB testing set | PSNR: 23.697 -> 23.710 | Helped |
| Add En (ablation) | SMDR-IS · UIEB testing set | PSNR: 23.021 -> 23.710 | Helped |
| Add De (ablation) | SMDR-IS · UIEB testing set | PSNR: 23.122 -> 23.710 | Helped |
| Use four stages instead of one (ablation) | SMDR-IS · UIEB testing set | PSNR: 22.790 -> 23.710 | Helped |
| Add S2-S4 to S1 (ablation) | SMDR-IS · UIEB testing set | PSNR: 22.790 -> 23.710 | Helped |
| Add S3-S4 to S1-S2 (ablation) | SMDR-IS · UIEB testing set | PSNR: 22.872 -> 23.710 | Helped |
| Add S4 to S1-S3 (ablation) | SMDR-IS · UIEB testing set | PSNR: 23.248 -> 23.710 | Helped |
| Add CA (ablation) | SMDR-IS BICA · UIEB testing set | PSNR: 21.219 -> 23.710 | Helped |
| Add SA (ablation) | SMDR-IS BICA · UIEB testing set | PSNR: 23.352 -> 23.710 | Helped |
| Add ReGIA (ablation) | SMDR-IS BICA · UIEB testing set | PSNR: 22.673 -> 23.710 | Helped |
| Add HCAFE (ablation) | SMDR-IS BICA · UIEB testing set | PSNR: 23.276 -> 23.710 | Helped |
| Add L1 loss (ablation) | SMDR-IS loss function · UIEB testing set | PSNR: 23.257 -> 23.710 | Helped |
Limitations
- Not the most efficient method; FUnIE-GAN and UWCNN are faster.
- Trained only on UIEB (800 images), with generalization shown only via testing on other datasets.
- Several components (CFA, HCAFE) are only described in supplementary material.
Experimental setup
Datasets: UIEB
Benchmarks: UIEB U45 LSUI
Metrics: PSNR MSE SSIM VSI FSIM FSIMc UIQM UCIQE CCF CEIQ
Baselines: ULAP IBLA GDCP WaterNet FUnIE-GAN UWCNN UColor UDA U-shape UIEC^2 USUIR UIEPTA
Builds on: U-Net
Closest papers by meaning
- Adaptive Dual-domain Learning for Underwater Image Enhancement (AAAI 2025 · similarity 0.87)
- UVEB: A Large-scale Benchmark and Baseline Towards Real-World Underwater Video Enhancement (CVPR 2024 · similarity 0.87)
- NAUTILUS: A Large Multimodal Model for Underwater Scene Understanding (NeurIPS 2025 · similarity 0.86)
- Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration (CVPR 2024 · similarity 0.86)
- WaterDiffusion: Learning a Prior-involved Unrolling Diffusion for Joint Underwater Saliency Detection and Visual Restoration (AAAI 2025 · similarity 0.86)
- Atlantis: Enabling Underwater Depth Estimation with Stable Diffusion (CVPR 2024 · 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.