Research map / Image generation / Image restoration
Underwater Organism Color Fine-Tuning via Decomposition and Guidance
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 CECF, a disentangled (color code / content code) underwater image enhancement network whose color codes can be fused with guidance images to fine-tune organism color in a controllable way.
Problem
Underwater organisms' colors are distorted by wavelength-dependent attenuation, and existing UIE methods output a single fixed-color result because paired ground truth is hard to obtain, so they cannot produce diverse or controllable colors.
Approach
Decompose the enhancement network into a color code and a content code, learn content invariance via self- and cross-reconstruction (style/content disentanglement after Huang et al. 2018), and fine-tune color by fusing the color code of a guidance image (long-wavelength colored image) with the distorted image's color code using a controllable parameter alpha.
Key results
- On UIEB, CECF reaches 21.819 PSNR and 0.894 SSIM, slightly above SGUIE (21.757 / 0.894) and the best UIQM among the compared methods on UIEB (3.122 vs. UGAN 3.276 is higher, so not best overall).
- On EUVP-I, CECF reaches 25.156 PSNR and 0.856 SSIM, the best PSNR and SSIM in Table 1.
- Color fine-tuning: UIQM changes with alpha are reported, with a reasonable range found in [0, 0.3]; the ablation shows structural loss improves UIEB SSIM from 0.883 to 0.894.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Structural Loss (ablation) | CECF · UIEB test | SSIM: 0.883 -> 0.894 | Helped |
| add Structural Loss (ablation) | CECF · EUVP-I test | SSIM: 0.856 -> 0.856 | No clear effect |
| increase color code length from 8 to 16 (ablation) | CECF · UIEB and EUVP-I test | PSNR: -0.1148; -0.0294 | Mixed |
| increase color code length from 8 to 16 (ablation) | CECF · UIEB and EUVP-I test | SSIM: 0.0017; 0.0002 | No clear effect |
| increase color code length from 8 to 16 (ablation) | CECF · UIEB and EUVP-I test | UIQM: -0.0185; 0.0318 | Mixed |
| vary color guidance and increase α (ablation) | CECF color controller · underwater and natural guidance images | UIQM | Mixed |
| vary the ratio λ1:λ2 from 10 to 2, 4, 6, or 8 (ablation) | CECF · UIEB test | PSNR: -0.282; -0.307; -0.065; -0.288 | Hurt |
Surprising findings
- UIQM does not increase monotonically with alpha; it first rises then falls, so the useful range of alpha is limited.
- Color code length (8, 16, 24, 32) does not significantly affect quantitative results.
Limitations
- Ground-truth references are scarce, so color fine-tuning quality is evaluated only with no-reference UIQM and visual inspection.
- Guidance selection relies on the assumption that long-wavelength hues are invariant; no strict criteria are given.
- Segmentation masks are assumed available; organism segmentation is out of scope.
Experimental setup
Datasets: UIEB EUVP-I (Underwater ImageNet subset of EUVP) EUVP-S (Underwater Scenes subset of EUVP)
Benchmarks: UIEB EUVP-I
Metrics: PSNR SSIM UIQM UCIQE
Baselines: ADMNet FUnIE MLLE URes HLRP WaterNet PhyNN SMBL UWNet UGAN SGUIE
Builds on: MUNIT (Huang et al. 2018) Structural Loss (Zhao et al. 2016) GAN
Closest papers by meaning
- Wavelet-based Fourier Information Interaction with Frequency Diffusion Adjustment for Underwater Image Restoration (CVPR 2024 · similarity 0.81)
- Adaptive Dual-domain Learning for Underwater Image Enhancement (AAAI 2025 · similarity 0.80)
- UVEB: A Large-scale Benchmark and Baseline Towards Real-World Underwater Video Enhancement (CVPR 2024 · similarity 0.79)
- ChromaFusionNet (CFNet): Natural Fusion of Fine-Grained Color Editing (AAAI 2024 · similarity 0.78)
- Synergistic Multiscale Detail Refinement via Intrinsic Supervision for Underwater Image Enhancement (AAAI 2024 · similarity 0.78)
- Color Shift Estimation-and-Correction for Image Enhancement (CVPR 2024 · similarity 0.78)
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.