关键词: 车辆工程；挡风玻璃除雾；透射率估计；曼哈顿自注意力；频域动态聚合

Abstract: Fogging on vehicle windshields can severely obstruct the driver’s field of vision and endanger driving safety. Traditional automatic defogging systems rely on temperature and humidity sensors to determine fogging risk, making it difficult to accurately perceive the complex and non-uniform microclimate distribution inside the vehicle, which easily causes delayed defogging response or control redundancy. To address this issue, a lightweight transmission estimation method was proposed from a visual perception perspective. The pixel-level transmission distribution on the windshield surface was directly estimated from a single image to characterize the fogging state. Aiming at the spatial non-uniformity of fog distribution and the directional nature of its physical diffusion, a lightweight network named MSF-Net was designed, which integrated a Manhattan self-attention mechanism with a frequency-domain dynamic aggregation module. Specifically, the Manhattan self-attention mechanism models had horizontal and vertical feature dependencies, incorporating physical diffusion path prior to enhance the ability to characterize fog structures. The frequency-domain dynamic aggregation module adaptively decoupled background texture and fog features in the frequency domain, effectively mitigating the interference of high-frequency background details on transmission estimation. Furthermore, to overcome the scarcity of real-world fog data, a physics-optics coupled data synthesis method based on COMSOL multi-physics field simulation was proposed, constructing the synthetic windshield fog image with physical consistency and the corresponding transmittance true values. Research results demonstrate that MSF-Net outperforms mainstream methods such as DCP and DehazeNet in terms of PSNR and SSIM on the synthetic dataset, meanwhile it also exhibits significant lightweight advantages. The proposed model has been successfully deployed on a Raspberry Pi 5 embedded platform, and its inference speed can meet the application requirements of periodic state updates and closed-loop regulation in onboard defogging control systems, thereby providing a viable visual perception solution for intelligent defogging systems.

Key words: vehicle engineering; windshield defogging; transmittance estimation; Manhattan self-attention; frequency-domain dynamic aggregation