SGLE-Net: polarization-aware enhancement in low light via an intensity-Stokes domain information fusion network

In low-light conditions, polarization camera imaging quality deteriorates, compromising both light intensity capture and polarization information extraction, such as the degree of polarization (DoP) and angle of polarization (AoP). Unlike RGB images, polarized images contain additional physical information, making general enhancement methods unsuitable. To address this, we propose SGLE-Net, a Stokes-guided enhancement network specialized for low-light polarized images, leveraging both intensity-domain and polarization-domain information during enhancement. It consists of two sub-networks: (1) the intensity-enhanced network, which initially enhances brightness for global feature improvement, and (2) the polar-merged network, which fuses intensity and polarization domain information for comprehensive enhancement. To guide intensity-Stokes feature fusion, we introduce a Stokes-based fusion-enhanced module, leveraging complete intensity information and edge details less affected by environmental noise. Additionally, we design a polar-joint objective function combining intensity loss and polarization loss to balance the network's focus on total intensity and polarization information during training. Evaluated on two state-of-the-art low-light polarized datasets, SGLE-Net outperforms previous methods, effectively enhancing total intensity and producing higher-quality DoP and AoP.