A physics-informed prior-guided network for maritime low-light image enhancement and visual perception

Reliable maritime visual perception is important for vessel monitoring and maritime surveillance, but low-light images are often affected by non-uniform illumination, atmospheric scattering, and sea-surface reflections. To address these degradations, we propose the Decoupled Prior-Guided Linkage Network (DPL-Net), which separates enhancement into physics-aware prior estimation and conditional texture refinement. A Structural Prior Subnetwork (StructureNet) extracts illumination, edge, and contrast priors from the input image. The Texture Refinement Subnetwork (RefineNet) incorporates these priors through Feature-wise Linear Modulation (FiLM) to adjust the restoration features according to spatially varying degradation. Within RefineNet, a HybridBlock combines shifted-window self-attention with convolution in parallel, allowing the network to model broad illumination variations while preserving local structures important for vessel identification. Experiments on the Low-Light (LOL-v1) dataset and self-built Maritime Low-light (MLOW) dataset show that DPL-Net achieves competitive or improved performance compared with existing methods. The proposed model delivers a significant performance gain, achieving approximately 6% improvement in reconstruction quality (measured by Peak Signal-to-Noise Ratio, PSNR) and a notable reduction in color deviation compared to existing methods on the maritime-specific MLOW dataset. On real-world images, DPL-Net achieves a Natural Image Quality Evaluator (NIQE) score of 3.93. The enhanced images also improve the visual quality of inputs used for downstream maritime ship detection under challenging environmental conditions.