Adaptive Underwater Image Enhancement Guided by Degradation-Perception Priors

To address the limitations of traditional underwater image enhancement methods, where non-physical-model-based approaches often lack degradation-mechanism constraints and may cause under-enhancement or over-enhancement, while physical-model-based approaches rely on stable estimation of key parameters, this paper proposes an adaptive enhancement method guided by degradation-aware priors. The method constructs physically interpretable priors from local statistical image features, thereby preserving the flexibility of non-physical-model-based enhancement while introducing degradation constraints, without requiring target-dataset network training or explicit inversion of a complete underwater imaging model. Specifically, color deviation is first corrected using a minimum attenuation channel strategy, and adaptive luminance equalization is achieved by estimating a non-uniform ambient light field via guided filtering. Local contrast variance, ambient light field deviation, local energy statistics, and gradient magnitude are then mapped to haze distribution, exposure risk, noise distribution, and structural saliency, respectively, to guide local dehazing and multi-scale texture enhancement. Experiments show that the proposed method performs favorably in color deviation, structural preservation, texture detail enhancement, and computational efficiency.