CoRe-UIE: Rethinking Coexisting and Region-wise Degradation for Underwater Image Enhancement

CoRe-UIE: Rethinking Coexisting and Region-wise Degradation for Underwater Image Enhancement proposes a degradation-oriented expert collaboration framework to address spatially diverse and locally coexisting underwater degradations, including color distortion, scattering haze, texture attenuation, and uneven illumination. The paper states: "Underwater images often suffer from diverse and coexisting degradations, including color distortion, scattering haze, texture attenuation, and uneven illumination. These degradations vary across regions and may coexist locally, making conventional uniform restoration difficult to adapt to different degradation patterns."

The proposed CoRe-UIE framework combines a content-preserving shared expert with four shared-backbone routed experts for color correction, scattering suppression, texture recovery, and illumination protection. The routed experts share the same architecture but have independent parameters, and are assigned to different regions through input-derived degradation cues and region-adaptive Top-k routing. The paper further introduces a Hilbert–Schmidt Independence Criterion (HSIC)-based representation constraint to reduce statistical dependence among expert features and alleviate redundant expert responses.

The main contributions are summarized as: (1) proposing CoRe-UIE, a mechanism-aware adaptive restoration framework for underwater image enhancement, which addresses spatially diverse and locally coexisting underwater degradations through shared-backbone expert collaboration; (2) designing "a region-adaptive Top-k routing strategy that selects different expert combinations for different image regions according to local degradation characteristics, enabling adaptive restoration under mixed underwater degradation patterns; and (3) introducing a mechanism-cue-guided expert specialization objective that combines response alignment and HSIC-based representation disentanglement, encouraging experts with the same architecture to learn less redundant and more complementary restoration representations."

The methodology includes a shallow encoder that extracts base features, a content-preserving shared expert that captures degradation-invariant content information, and four routed experts (Ec, Esc, Et, El) corresponding to color, scattering, texture, and illumination roles. The routed experts use a residual restoration block: Ei (F) = F + Ci2 σ Ci1 (F), where parameters are independent across experts. The region-adaptive Top-k router predicts spatially varying expert weights from the encoded feature F and input-derived mechanism cues P, including color imbalance, scattering-related contrast degradation, texture structure, and illumination risk. The router selects the Top-k experts (k=2 by default) and renormalizes their weights for feature fusion.

The loss function combines reconstruction fidelity, response alignment, and expert representation disentanglement: L = Lrec + λalign Lalign + λhsic Lhsic. The response alignment loss softly aligns expert responses with their corresponding degradation-related regions without requiring pixel-level degradation annotations. The HSIC loss reduces statistical dependence among expert representations, computed empirically as HSIC(Zi, Zj) = 1/(m−1)2 tr (Ki HKj H).

Experiments were conducted on UIEB, LSUI, and U45 datasets. On UIEB, CoRe-UIE achieves the best PSNR, SSIM, FSIM, FSIMC, and VSI scores, reaching 26.88 dB, 0.9103, 0.9666, 0.9554, and 0.9855, respectively. On LSUI, CoRe-UIE also achieves the best results across all metrics. On U45, CoRe-UIE achieves the best results on all four metrics (NIQE, BRISQUE, CEIQ, PIQE), suggesting favorable perceptual quality and robustness on reference-free underwater scenes.

Ablation studies show that removing any routed expert leads to performance degradation, with removing the scattering-related expert causes the largest drop. The variant without the shared expert performs worse, and the variant without HSIC also degrades performance. For Top-k routing, Top-2 achieves the best balance between collaboration and selectivity, and is therefore adopted as the default setting.

The conclusion states: CoRe-UIE achieves competitive quantitative performance and visually balanced enhancement, especially for scenes with coexisting color distortion, scattering haze, texture degradation, and illumination variation.

Improvements to AI Systems:

1. Region-Adaptive Expert Routing for Multi-Degradation Handling: Implement a Top-k routing mechanism that dynamically selects specialized sub-models (experts) per image region based on input-derived degradation cues (color imbalance, contrast loss, texture structure, illumination risk). This allows the AI to handle spatially heterogeneous degradation in a single forward pass, avoiding uniform processing that fails on mixed distortions.

2. Shared-Backbone Expert Collaboration with Disentangled Representations: Use a content-preserving shared expert to maintain degradation-invariant features, while four independent-parameter experts (color, scattering, texture, illumination) process routed regions. Add an HSIC-based loss to minimize statistical dependence between expert features, forcing complementary rather than redundant representations—improving generalization to unseen degradation combinations.

3. Mechanism-Cue-Guided Soft Alignment for Weakly Supervised Training: Replace pixel-level degradation annotations with soft response alignment loss that matches expert outputs to region-level degradation cues (e.g., color imbalance maps, contrast maps). This enables training on datasets without dense labels, reducing annotation cost while preserving restoration accuracy.

4. Residual Restoration Blocks with Independent Parameters: Adopt the residual block formulation E i(F) = F + C i2(σ(C i1(F))) for each expert, where parameters are independent. This allows efficient fine-tuning of individual degradation types without retraining the whole network, and enables incremental addition of new degradation experts (e.g., for low-light or motion blur) without catastrophic forgetting.

5. Adaptive Fusion via Renormalized Top-k Weights: After selecting Top-k experts per region, renormalize their predicted weights before feature fusion. This prevents over-suppression of less dominant degradations and ensures balanced contribution, improving performance on images where multiple degradations coexist with varying severity.

What the Improved AI System Can Do:

• Enhance underwater images with mixed, spatially varying degradations (e.g., color cast on left, haze on right, texture loss in center) in one pass, achieving state-of-the-art PSNR (26.88 dB on UIEB) and perceptual quality (NIQE, BRISQUE) without manual region segmentation.

• Adapt to unseen degradation patterns by routing to the most relevant experts based on input cues, even if the exact combination was not seen during training.

• Train efficiently with weak supervision—only needing global image-level cues or rough degradation maps, not pixel-level annotations.

• Scale to new degradation types by adding a new expert with the same architecture and routing logic, without retraining existing experts.

• Produce visually balanced outputs that avoid over-enhancing one aspect (e.g., over-saturating color while leaving haze), as demonstrated by best FSIM, VSI, and CEIQ scores on benchmark datasets.