TIDE: Two-Stage Inverse Degradation Estimation with Guided Prior Disentanglement for Underwater Image Restoration

TL;DR

TIDE tackles underwater image restoration by explicitly modeling multiple degradations as spatially varying factors. It introduces inverse degradation mapping with specialized priors to generate hypothesis-specific restorations, followed by a two-stage progressive refinement that targets residual artifacts with expert-guided corrections. The framework employs adaptive fusion, safety gating, and a diversified loss design to promote specialization and robust performance across standard and turbid-water datasets. Empirical results show improvements in perceptual quality and color/contrast restoration while maintaining real-time or near-real-time efficiency on consumer GPUs, highlighting its practical potential for marine imaging tasks.

Abstract

Underwater image restoration is essential for marine applications ranging from ecological monitoring to archaeological surveys, but effectively addressing the complex and spatially varying nature of underwater degradations remains a challenge. Existing methods typically apply uniform restoration strategies across the entire image, struggling to handle multiple co-occurring degradations that vary spatially and with water conditions. We introduce TIDE, a $\underline{t}$wo stage $\underline{i}$nverse $\underline{d}$egradation $\underline{e}$stimation framework that explicitly models degradation characteristics and applies targeted restoration through specialized prior decomposition. Our approach disentangles the restoration process into multiple specialized hypotheses that are adaptively fused based on local degradation patterns, followed by a progressive refinement stage that corrects residual artifacts. Specifically, TIDE decomposes underwater degradations into four key factors, namely color distortion, haze, detail loss, and noise, and designs restoration experts specialized for each. By generating specialized restoration hypotheses, TIDE balances competing degradation factors and produces natural results even in highly degraded regions. Extensive experiments across both standard benchmarks and challenging turbid water conditions show that TIDE achieves competitive performance on reference based fidelity metrics while outperforming state of the art methods on non reference perceptual quality metrics, with strong improvements in color correction and contrast enhancement. Our code is available at: https://rakesh-123-cryp.github.io/TIDE.

Figures (10)

• Figure 1: Architecture of the TIDE framework. The top shows the overall flow: the input image is first processed by the base model to generate an initial restoration, which along with the original image feeds into the residual degradation estimator (RDE) and refinement fusion (RF) to produce the final output. The bottom left shows the base model details, where a feature extractor $\mathcal{E}$ feeds specialized decoders $\mathcal{D}_1$-$\mathcal{D}_4$ that generate restoration hypotheses $\mathcal{H}_1$-$\mathcal{H}_4$, while also estimating degradation maps. These hypotheses are combined through a learned mapping function for initial restoration. The bottom right illustrates the refinement stage, where the original and initially restored images are processed by shared convolutional processing before being fed to specialized refinement experts $E_1$-$E_4$ that generate targeted corrections for color, contrast, detail, and noise, which are then adaptively scaled and combined.
• Figure 2: Visual comparison of input degradations, initial restoration, and expert refinement stages, illustrating TIDE’s progressive restoration process.
• Figure 3: Qualitative restoration results across EUVP, UIEB, and SUIM-E datasets.
• Figure 4: Qualitative comparison on naturally turbid underwater images from the RUIE dataset. The left group corresponds to the UCCS dataset, while the right group corresponds to the UIQS dataset. For each result, the two metrics displayed in the top corners are UICM (left, measuring color correction) and UIConM (right, measuring contrast enhancement). Higher values indicate better performance for both metrics.
• Figure 5: Performance analysis of TIDE across different resolutions and batch sizes on RTX 4070 Ti SUPER. Bubble size indicates processing speed (inverse latency), while position represents the throughput-memory tradeoff. The horizontal red line marks the 30 FPS real-time processing threshold. Even at $256\times256$ resolution, TIDE achieves real-time performance (33.7 FPS) with minimal memory usage (0.44 GB), while smaller resolutions enable substantial throughput improvements with negligible overhead.