PAC-Net: Physics-Aware Constraint Network for Underwater Image Enhancement

Optical sensors in underwater environments typically suffer from significant challenges in capturing high-quality underwater images due to complex oceanographic conditions, including light scattering, non-uniform illumination, and noise from aquatic media. These issues inevitably lead to a reduction in contrast and color cast in underwater images, which severely affects the quality of subsequent downstream tasks. To this end, this paper proposes an innovative Physics-Aware Constraint Network (PAC-Net) to achieve underwater image enhancement. Specifically, the salient properties of our PAC-Net are: 1) a Physics-Aware Constraint Module (PACM) is developed, which utilizes the self-attention mechanism in the Transformer architecture to imitate the underwater light transmission process, providing sufficient prior knowledge for the model to facilitate enhancement of underwater images, 2) an Asymmetric Feature Cross Excitation Module (AFCEM) is designed, which can achieve sufficient feature extraction and representation through an innovative feature cross excitation mechanism, and 3) a Cascade Feature Recalibration Module (CFRM) is constructed, which aims to further facilitate adaptive feature recalibration by constructing a multi-stage progressive refinement structure. Experimental results demonstrate the effectiveness of our model on several real-world datasets acquired by underwater optical sensors, and the ablation studies further confirm the effectiveness of each component in our network. The source code is available at https://github.com/XXX.