MAPF-Net: lightweight network for dehazing via multi-scale attention and physics-aware feature fusion

Single image dehazing is an important preprocessing step in advanced computer vision tasks. Existing deep learning models often have complex network structures, leading to low computational efficiency, while lightweight models usually compromise on dehazing performance. To address these issues, this paper proposes a lightweight network for dehazing with multi-scale attention and physics-aware feature fusion (MAPF-Net), aiming to achieve significant improvement in dehazing performance while maintaining a lightweight model. MAPF-Net includes two novel blocks: the multi-scale attention aggregation (MAA) module and the physics-aware feature fusion (PAFF) module. Specifically, the MAA module utilizes multi-scale large-kernel convolutions to extract features and aggregates multi-scale features through dual-residual parallel attention. This design not only enriches feature representation by fusing multi-scale contextual information but also mitigates gradient degradation via residual pathways, thereby maintaining discriminative capabilities for both dense and sparse haze regions. The PAFF module integrates atmospheric scattering priors into a learnable feature fusion framework, where multi-head self-attention captures global dependencies of haze patterns while pixel attention refines local texture details. It adopts a frequency-aware feature-weighted fusion strategy that balances low-frequency structural preservation and high-frequency texture recovery through learnable channel-wise coefficients. Comprehensive experiments conducted on multiple benchmark datasets, along with ablation studies, validate the effectiveness of the proposed method. The outcomes show that MAPF-Net uses just 0.687M parameters and 7.547G FLOPs to deliver state-of-the-art performance.