PGF-Net: fusing physical imaging model with self-attention for robust underwater feature detection

Robust feature detection in underwater environments is severely impeded by image degradation from light absorption and scattering. Traditional algorithms fail in these low-contrast, blurred conditions, while deep learning methods suffer from the domain gap between terrestrial and underwater imagery and a scarcity of annotated data. To address these challenges, this paper introduces PGF-Net, a systematic framework that fuses physical imaging principles with deep learning. The framework leverages a dual-fusion strategy: First, a parametric underwater imaging model is proposed to guide the synthesis of a large-scale, physically realistic training dataset, effectively injecting prior knowledge of the degradation process into the data domain. Second, a novel detection network architecture is designed, which incorporates a self-attention mechanism to fuse local features with global contextual information, enhancing robustness against detail loss. This end-to-end network is trained on the synthesized data using a curriculum learning strategy, progressing from mild to severe degradation conditions. Extensive experiments on public datasets demonstrate that PGF-Net significantly outperforms classic and state-of-the-art deep learning methods in both keypoint detection and matching, particularly in turbid water. The proposed framework validates the efficacy of integrating physical priors with data-driven models for challenging computer vision tasks and provides a robust solution for underwater visual perception.