Current dehazing algorithms often rely on small-kernel ( \(e.g., 3\times 3\) ) convolutional neural networks (CNNs) to extract local features, but their limited receptive fields hinder the capture of global dependencies. While Transformers excel at modeling long-range features, they suffer from high computational complexity. Moreover, most existing methods lack guidance from physical models, making the dehazing process more like a “black box” and lacking interpretability. To address these challenges, we propose the Global and Local-Physics-aware Fusion Network (GLPF-Net), which balances local and global feature extraction while maintaining computational efficiency and interpretability. GLPF-Net integrates two key modules: the physics-aware multi-scale feature extraction (PMFE) module and the dehazing enhanced Transformer (DEformer) module. PMFE combines physics-aware multi-branch attention (PMA) and multi-scale convolutional feature extraction (MCFE) to address the issues of lacking interpretability and limited local feature extraction. PMA integrates the atmospheric scattering model into the network learning process to better adapt to non-uniform haze scenarios. MCFE has designed targeted receptive fields for dehazing tasks by effectively combining multi-scale convolution and activation functions. Meanwhile, DEformer introduces multi-head cross-covariance self-attention and dynamic layer selection to achieve global modeling while reducing complexity. Extensive experiments demonstrate that GLPF-Net achieves state-of-the-art performance in both objective metrics and subjective visual quality.
