USF-Net: U-shaped Mamba Dehazing Network Enhanced by Wavelet Features and Spatial Position Fusion

In this paper, we employ wavelet transform analysis to reveal the wavelet degradation prior to haze. Through this prior, we demonstrate that haze- related information primarily resides in low-frequency components, while its effect on high-frequency components manifests mainly as edge blurring and tex- ture detail attenuation. Leveraging this insight, we propose a novel dehazing framework, USF-Net, which decouples the ill-posed image dehazing problem into two subtasks: position-guided channel selective enhancement and spatial- aware multi-scale feature extraction. Specifically, we integrate Mamba blocks with spatially-informed channel-weighted attention modules, achieving global structure reconstruction with linear complexity while effectively fusing spatial and channel information. Additionally, we incorporate position-aware multi- receptive-field attention modules to efficiently aggregate multi-scale spatial features, enhancing the network’s perception of spatial structures under varying haze concentrations. This design significantly improves both local detail fidelity and global semantic understanding. Extensive experimental results demonstrate that our method achieves state-of-the-art performance on both synthetic datasets and real-world hazy images.