Dual-branch underwater image enhancement network based on fusion of polarization and color information

Underwater images suffer from degradation due to light scattering and absorption, hindering high-level vision tasks. We propose an end-to-end multimodal network that integrates polarization and color information by exploiting their distinct polarization characteristics. Our method features a dual-branch architecture to extract polarization and color features. In the polarization branch, the angle of linear polarization is decomposed into sine/cosine components, and a dedicated module captures geometric cues. A gated cross-modal attention mechanism enables adaptive feature fusion, followed by reconstruction via a Restormer backbone. Evaluated on a polarized dataset across turbidity levels, our network outperforms existing methods in both quantitative metrics and visual quality, offering reliable support for underwater vision applications.