Dynamic Large-Kernel Detail Enhancement Network for Underwater Image Super-Resolution

Underwater imaging is degraded by light absorption, scattering, and suspended particles, which reduce clarity, suppress brightness, and diminish fine details.To address these challenges, we propose a lightweight Dynamic Large-Kernel Detail Enhancement Network (DLKDENet).The proposed framework integrates a Dynamic Global Feature Adaptation Module (DGFAN) and a Detail Perception Enhancement Module (DPEN), enabling coordinated optimization between global structural modeling and local detail restoration.DGFAN leverages large-receptive-field convolutions and a two-stage channel-attention strategy to attenuate noise and promote global structural consistency.In complement, DPEN enhances fine-grained texture reconstruction through overlapping patch decomposition and a gated convolution mechanism.The experimental results show that DLKDENet achieves consistent gains in PSNR and SSIM over several baseline models on the USR-248 and UFO-120 datasets.Furthermore, it delivers high-quality reconstruction with fewer parameters, demonstrating its efficiency and strong adaptability to underwater imaging conditions.