IGFD-Net: Illumination-guided frequency decoupling for polarization image fusion

The intensity image (S₀) mainly captures the scene radiance information, while the degree of linear polarization (DoLP) reflects the material properties of objects. The effective fusion of the two modalities contributes to robust visual perception. However, existing polarization image fusion methods struggle to simultaneously maintain photometric consistency and saliency preservation. Particularly under complex illumination conditions, most methods fuse polarization information indiscriminately, leading to brightness drift and structural artifacts. To address this issue, we propose an illumination-guided frequency-decoupled fusion network (IGFD-Net). The network innovates with an illumination-guided attention mechanism that dynamically adjusts the weights of different modalities according to the illumination prior. Complementing this, a frequency-decoupled fusion module disentangles and reorganizes frequency components to ensure visual naturalness. Furthermore, we introduce a structure-aware saliency injection loss that selectively injects discriminative polarization details into S₀, which can constrain photometric uniformity. Experimental evaluations across multiple datasets indicate that the proposed method achieves competitive performance on both subjective and objective quality metrics and improves accuracy in downstream target detection tasks.