Recent advances in deep learning have substantially improved single image dehazing performance, yet most existing approaches remain limited in modeling the complex spatial variability of haze in real-world scenes. Such limitations often result in inaccurate estimation of haze density, leading to local over-dehazing or residual haze artifacts, particularly under non-uniform atmospheric conditions. To address these challenges, this paper proposes a density-aware dehazing network that integrates physical priors with deep feature learning to achieve more accurate haze modeling and feature restoration. Specifically, a Grid-Aware Atmospheric Attention module is designed to divide the image into local grids and adaptively learn spatial variations in haze density through channel-wise weighting, effectively preventing regional over-dehazing. In addition, a Multi-Scale Density-Aware module is introduced, which employs dilated convolutions with different rates to capture multi-scale contextual information and enhance feature representation in complex haze scenes. Furthermore, a Spatial Detail Enhancement module combines wavelet decomposition and spatial attention to restore fine-grained texture and structural details suppressed by haze. Extensive experiments on four widely used dehazing datasets, including ITS, O-HAZE, NH-HAZE, and NH-HAZE2, validate the effectiveness of the proposed network, which achieves 40.25 dB PSNR and 0.996 SSIM on the ITS dataset while maintaining computational efficiency with 13.94M parameters and 40.43G MACs. The results demonstrate that our method significantly improves visual quality and quantitative metrics compared with existing state-of-the-art approaches, providing a robust solution for image enhancement in low-visibility environments such as UAV vision, traffic monitoring, and optical imaging systems.

Image Dehazing, Deep Learning, Physical Prior, Attention Mechanism, Multi-scale Feature Fusion, Density Awareness