TL;DR
WaveFlow-UIE is a novel, single-pipeline restoration model that uses flow matching in the wavelet domain to clean up underwater images. By combining frequency decomposition with physics-inspired priors, it achieves state-of-the-art perceptual quality while being significantly faster than iterative diffusion-based alternatives—offering a $4\times$ speedup with virtually no loss in restoration quality.
What Problem Does This Paper Address?
Restoring underwater images is notoriously difficult because water absorbs and scatters light in wavelength-dependent ways. This leads to three main issues: color distortion (often a green/blue cast), haze (loss of contrast), and loss of fine detail.
Existing deep learning methods often struggle to balance these effects. Spatial-domain models might miss the distinct ways that low-frequency (color/lighting) and high-frequency (texture/edges) components degrade. Recent diffusion-based models offer high quality but are computationally expensive due to their multi-stage pipelines and hundreds of iterative refinement steps. The authors sought a middle ground: a model that is frequency-aware like a wavelet transform but efficient enough for practical use.
Methods
The core of WaveFlow-UIE is a conditional flow model operating in the wavelet domain. Here is the breakdown of the architecture:
- Wavelet Decomposition: The input image is processed using a Haar Discrete Wavelet Transform (DWT), splitting it into four frequency sub-bands (LL, LH, HL, HH). This allows the model to treat structural and textural information separately.
- Physics-Prior Conditioning: A lightweight branch estimates two key parameters from the underwater formation model: a per-pixel transmission map ($t$) and a global ambient light value ($A$). These provide the restoration network with explicit cues about the haze and color cast.
- Wavelet-Domain Flow Matching: Instead of the complex noise-prediction used in diffusion, the authors use "Flow Matching." The model learns a "velocity field" that describes the direct path from a degraded wavelet representation to a clean one.
- Velocity U-Net: A U-Net architecture predicts this velocity, conditioned on the degraded image, the physics priors, and the current "flow time."
- Multi-Term Loss: The training objective is a weighted combination of flow-matching loss, frequency-weighted reconstruction, perceptual loss (LPIPS), color consistency (in Lab space), and physics consistency.
Datasets
The model was trained and evaluated on several standard underwater benchmarks:
- UIEB (Underwater Image Enhancement Benchmark): 800 training pairs, 90 test pairs.
- LSUI (Large-Scale Underwater Image dataset): ~3,900 training pairs, 400 test pairs.
- UFO-120 & EUVP: Used as additional paired test sets to evaluate generalization.
- U45 & C60: Reference-free challenge sets used to test performance on real-world, unpaired data.
Key Insights and Ideas
- The Power of Single-Step Inference: One of the most surprising findings is that the model's performance barely drops when reducing the inference process from five flow steps to just one step. This suggests that in the wavelet domain, the restoration path is nearly linear, allowing for massive speedups without sacrificing quality.
- Frequency Weighting Matters: By applying a higher weight to high-frequency wavelet residuals during training, the model preserves fine textures and edges better than models that treat all pixels equally.
- Flow vs. Diffusion: Flow matching provides a more direct mathematical framework for restoration than diffusion. Because the target "velocity" is simply the difference between the clean and degraded states, the model effectively learns a sophisticated form of residual prediction.
Key Results
WaveFlow-UIE demonstrates a superior quality-efficiency tradeoff compared to previous state-of-the-art (SOTA) methods:
- Perceptual Superiority: On the UIEB dataset, it achieved the best LPIPS score (0.1390) and the second-best PSNR and SSIM, indicating results that are perceptually very close to the ground truth.
- Speedup: Compared to the wavelet-based diffusion baseline (WF-Diff), WaveFlow-UIE is roughly $4\times$ faster at $256\times256$ resolution (79.31 ms vs 301.30 ms). When compared to the iterative DiffWater, the speedup is orders of magnitude higher.
- Qualitative Clarity: As seen in Figure 2, WaveFlow-UIE tends to recover more natural colors and sharper details than feed-forward baselines like Water-Net or UWCNN.
Limitations and Future Work
- Inference Speed vs. Feed-Forward: While much faster than diffusion models, WaveFlow-UIE is still slower than simple feed-forward CNNs (like UIE-WD) because it still involves a U-Net pass and wavelet transforms.
- Reference-Free Performance: The model's performance on fully reference-free datasets (U45, C60) was "mixed." It remains competitive but does not always lead the rankings, suggesting that the model generalizes best when the degradation follows the patterns seen in paired training data.
- Physics Branch Interpretation: The authors noted that the physics branch acts more like a "learned conditioning input" than a strictly accurate physical estimator. Future work could involve tighter integration of more complex underwater light transport models.
Summary updated