KLDD: Underwater image enhancement via KL-divergence-driven color cast correction and multistage enhancement fusion

Underwater images often suffer from severe color casts, nonuniform illumination, reduced contrast, and blurred details because of wavelength-dependent absorption and light scattering. To address these degradations, this study proposes KLDD, an underwater image enhancement method based on KL-divergence-driven color cast correction and multistage enhancement fusion. First, a KL-divergence-driven adaptive color cast correction module measures the discrepancy between the two-dimensional joint chrominance distributions of an underwater image and a natural reference image in the CIELAB color space. The resulting divergence is mapped to a bounded compensation coefficient that adaptively controls the strength of statistical color transfer. The color-corrected image is subsequently processed in parallel by a multistage adaptive luminance and contrast enhancement module and a multiscale adaptive detail enhancement module. The former combines complementary gamma–logarithmic mapping, noise-aware multiscale local reconstruction, confidence-guided fusion, and luminance-anchor regularization to improve dark-region visibility and local contrast while preserving highlights. The latter integrates multiscale decomposition, adaptive denoising, and edge-saliency-based weighting to enhance reliable textures and structural details while suppressing noise amplification. Finally, a cross-scale consistency-guided adaptive DWT fusion module integrates the complementary information from the two enhancement branches using reliability-aware low-frequency fusion and high-frequency fusion incorporating a cross-scale structural prior, noise-aware weighting, and sign-consistency constraints. Experiments on the UCCS, UIQS, and UIEB datasets demonstrate that KLDD achieves the best or near-best performance on several no-reference image-quality metrics while producing natural color restoration, balanced luminance, and clear structural details. Ablation studies, reference-image analysis, parameter sensitivity analysis, and runtime comparisons further verify the effectiveness and stability of the proposed method.