Underwater image enhancement by jointly exploiting RGB and polarization modalities

Underwater images often suffer from severe quality degradation due to light absorption and scattering in the water medium. Most existing underwater image enhancement (UIE) methods rely solely on RGB inputs, which lack the capability to distinguish between scattered and reflected light, thus limiting their performance. In contrast, polarization imaging offers the potential to disentangle physical components and preserve surface structures by capturing the polarization state of light. This paper thus proposes a novel RGB-polarization multimodal fusion framework for UIE tasks. Specifically, we first present a Polarization Feature Extractor (PFE) to capture direction-dependent polarization responses via multi-dimensional interaction modeling. In addition, a cross-modal fusion module is introduced to effectively and adaptively combine meaningful cues from both RGB and polarization domains. The effectiveness is enforced by the channel attention mechanism and the spatial attention mechanism to improve feature representation; the adaptiveness is facilitated by a specifically designed weighting scheme that balances the contributions of the two domains. Extensive experiments show that the proposed approach outperforms state-of-the-art underwater image enhancement methods in terms of both full-reference and non-reference metrics. Furthermore, the contribution of each key component is validated through comprehensive ablation study.