Parallel Attention Groups and Dynamic Channel Fusion for Image Dehazing in Port Security
Jingqi Zhang, Xianqiao Chen, Boshi Li, Yan Chun Zhou · 2025
In the context of port safety management, dense fog formed by severe weather often makes ship images blurry, posing significant safety hazards. We address the limitations of existing image dehazing methods by introducing a dynamic color correction technique to mitigate residual haze, detail loss, and color distortion. We incorporate parallel attention module groups (PAMG) in the decoder to enhance feature selection and preserve image details. Additionally, an adaptive color-correction (ACC) branch based on a differentiable dark channel prior is used to guide color restoration. The network is trained with a hybrid loss combining the Smooth L1 and MS-SSIM to balance pixel fidelity and structural similarity. Experimental results on both synthetic and real-world hazy datasets demonstrate that the proposed method significantly improves dehazed image clarity and color fidelity. However, the model’s dynamic convolution and attention modules introduce increased computational complexity, which may limit real‑time performance under extremely dense or heterogeneous haze conditions.