This paper presents PAF-Net, an image dehazing framework that integrates physical priors and adaptive feature modeling. Its aims to enhance image restoration quality in complex non-uniform haze scenes, while maintaining moderate model complexity and computational cost. Built upon the classical FFA-Net backbone, PAF-Net enhances dehazing performance from three complementary perspectives: spatial attention decision, physics-consistent feature representation in feature space, and operator-level adaptability. These perspectives are realized by three corresponding modules. Specifically, a Fog-aware Concentration Attention (FCA) module is introduced. It embeds haze-related spatial guidance into attention-weight generation, reducing attention–allocation imbalance under spatially non-uniform haze and strengthening the model’s focus on dense-haze regions. To improve physical feature consistency under non-uniform haze, a Physics-aware Feature Dehazing Unit (PFDU) is further designed to explicitly model and reorganize transmission-related and atmospheric-light-related feature components under atmospheric scattering priors. In addition, a Dynamic Convolution Module (DCM) is incorporated to adapt convolutional responses at the sample level according to global degradation patterns, enhancing robustness across diverse haze conditions. Experiments on RESIDE SOTS and real-world benchmarks (Dense-Haze and NH-Haze) demonstrate that PAF-Net achieves higher PSNR/SSIM and yields more natural visual results than representative methods. Further experiments and evaluations based on paired remote sensing datasets have verified the applicability of the designed PAF-Net algorithm in remote sensing scenes with spatially varying haze, as well as the algorithm’s generalization ability.
Yang et al. (Sun,) studied this question.