AMGN-RUNet: A multi-scale attention guided U-Net for non-homogeneous image dehazing
Shaohui Jin, Zhengguang Qin, Pengfei Zhao, Fuqiang Wang, Yang Lu, Mingliang Xu · Journal of Visual Communication and Image Representation · 2026
The spatial uncertainty of non-homogeneous haze distribution leads to localized degradations, including uneven blurring, color distortion, and the loss of contrast and texture, posing severe challenges for image dehazing. To address this challenge, we propose AMGN-RUNet, a dehazing network that integrates multi-scale gated convolutions with attention-guided feature fusion. The network proposes an Attention-Guided Feature Extractor (AMGN), which generates attention map under the guidance of ground truth. The attention map enables the network to focus on the most severely degraded regions and guides the reconstruction process in the RUNet decoder. Multi-scale receptive field convolutions, as an integral part of the efficient U-Net backbone, improve the perception of regions with different haze densities. Together, AMGN and the RUNet collaborate to achieve robust non-homogeneous dehazing. Extensive experiments on both synthetic and real-world datasets demonstrate that AMGN-RUNet achieves superior performance in recovering structural details, color fidelity, and overall visual quality.