PDFormer: a polarization-driven dual-branch transformer for dense fog removal

Foggy scenes pose significant challenges to image restoration due to complex light scattering, which reduces contrast and obscures fine details. Existing defogging methods that rely solely on intensity inputs struggle in dense fog conditions, lacking both physical priors and global reasoning capabilities. A novel dual-branch transformer-based architecture, termed PDFormer, is proposed to integrate physics-based and data-driven insights, aiming to jointly capture global context and recover fine-grained details: the intensity branch employs a hierarchical multiscale window-based Transformer to capture long-range dependencies and structural coherence, while the polarization branch leverages multi-angle polarization images to extract scattering-aware local features guided by physical priors. A lightweight fusion module adaptively integrates both branches, enabling comprehensive restoration across varying fog densities. Extensive experiments on both synthetic and real-world datasets demonstrate that PDFormer achieves state-of-the-art performance in terms of perceptual quality and quantitative metrics, effectively restoring visibility even under challenging atmospheric conditions.