Research map / Detection and segmentation / CNN features
WaterDiffusion: Learning a Prior-involved Unrolling Diffusion for Joint Underwater Saliency Detection and Visual 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 WaterDiffusion, a prior-involved unrolling diffusion model that jointly performs underwater saliency detection and underwater image restoration using a single network.
Problem
Underwater salient object detection suffers from object mislocalization and imprecise boundaries due to light attenuation and scattering; existing pipelines cascade UIE and USOD, ignoring that restoration could help saliency detection.
Approach
A conditional diffusion model whose denoiser jointly predicts noise and saliency masks, injects an underwater medium transmission prior into the decoder, uses gate-guided selection between instance-normalized and raw features, and unrolls sampling with half-quadratic splitting to iteratively refine masks and restored images.
Key results
- On UIEB/UWScene/LSUI average, WaterDiffusion reports UIF 0.467 and PSNR 25.231, best UIF among compared UIE methods.
- On USOD10K, WaterDiffusion achieves F-mean 0.767, S-measure 0.827 and MAE 0.053 (best MAE), versus TC-USOD and RMFormer baselines.
- Removing the medium transmission prior drops saliency F-mean on UFO-120 from 0.654 to 0.610.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add iterative mask refinement (IMR) (ablation) | WaterDiffusion · UIEB | PSNR: 22.864 -> 24.938 | Helped |
| Add iterative mask refinement (IMR) (ablation) | WaterDiffusion · USOD | saliency detection | Helped |
| Add medium transmission prior (MTP) (ablation) | WaterDiffusion · UIEB | PSNR: 20.338 -> 24.938 | Helped |
| Add medium transmission prior (MTP) (ablation) | WaterDiffusion · UFO-120 | F-measure: 0.610 -> 0.654 | Helped |
| Add gate-guided feature selection (GFS) (ablation) | WaterDiffusion · training | SSIM | Helped |
| Use WaterDiffusion instead of HCLR-Net | UIE methods · UIEB | SSIM: 0.881 -> 0.865 | Hurt |
| Use WaterDiffusion instead of U-Shape | UIE methods · UWScene | PSNR: 25.242 -> 25.751 | Helped |
| Use WaterDiffusion instead of RMFormer | USOD methods · USOD10K | F-measure: 0.788 -> 0.767 | Hurt |
| Use WaterDiffusion instead of TC-USOD | USOD methods · T-SUIM | MAE: 0.072 -> 0.056 | Helped |
Surprising findings
- Iterative mask refinement is reported as necessary for the USOD task itself, since saliency detection ceases to function without it.
Limitations
- Slow convergence and oscillations in the first 100 epochs after introducing the GFS module.
- Relies on a dark-channel-prior transmission estimate, which may be unreliable for some underwater scenes.
Experimental setup
Datasets: SUIM USOD10K UIEB UWScene LSUI UFO-120
Benchmarks: UIEB UWScene LSUI UFO-120 SUIM USOD10K
Metrics: UIF SSIM PSNR S-measure E-measure F-measure MAE
Baselines: Haze-Lines MLLE WWPF PUIE-Net U-Shape Semi-UIR SMDR-IS HCLR-Net U2-Net SUIM-Net SVAM-Net RMFormer
Builds on: denoising diffusion probabilistic models half-quadratic splitting instance normalization
Closest papers by meaning
- Diving into Underwater: Segment Anything Model Guided Underwater Salient Instance Segmentation and A Large-scale Dataset (ICML 2024 · similarity 0.89)
- Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration (CVPR 2024 · similarity 0.87)
- Atlantis: Enabling Underwater Depth Estimation with Stable Diffusion (CVPR 2024 · similarity 0.86)
- NAUTILUS: A Large Multimodal Model for Underwater Scene Understanding (NeurIPS 2025 · similarity 0.86)
- Synergistic Multiscale Detail Refinement via Intrinsic Supervision for Underwater Image Enhancement (AAAI 2024 · similarity 0.86)
- Adaptive Dual-domain Learning for Underwater Image Enhancement (AAAI 2025 · 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.