Hap-Net: Single Image Dehazing Based on Hybrid Attention and Physical Prior Model
Shihao Chen, Yanduo Zhang, Tao Lü, Yunxiao Hu, Jiaming Wang · 2024
Existing dehazing methods use spatial attention to tackle the uneven distribution of haze at the image level and channel attention to model channel-wise differences, which ignore information interaction between them. Therefore, we introduce a hybrid attention module to solve this issue. Moreover, we incorporate a pixel attention module afterward to enhance image details. In addition, we introduce a physical prior module, based on the Atmospheric Scattering Model, to improve the interpretability of the dehazing process. By combining the abovementioned components, we propose our hybrid attention and physical prior network (HAP-Net) for recovering haze-free images. Extensive experimental results demonstrate the effectiveness of our HAP-Net.