Physics-based model and transformer for single image dehazing

Convolutional Neural Networks (CNNs) struggle with global dependencies due to their constrained receptive fields, whereas Transformers excel at modeling long-range relationships at higher computational cost. In addition, most dehazing methods lack physical model guidance, making their processes less interpretable. To overcome these limitations, we propose PhysFormer, which mainly includes two key modules: the Physics-aware Local Feature Enhancement (PLFE) module and the Dehazing Transformer Module (DTM). PLFE integrates the atmospheric scattering model with multi-scale convolution enhancement into the network. DTM combines the Multi-head Global Attention (MGA) block and the Dynamic Feature Enhancement (DFE) block to build dynamic global dependencies. Extensive experiments have demonstrated the effectiveness of our method.