FADR-Net: fog-aware deformable and Retinex-guided network for image dehazing

Haze degrades image quality and undermines the performance of vision systems in detection and recognition tasks, especially in complex scenes. Although lightweight methods such as AOD-Net are computationally efficient, they exhibit two critical limitations: limited adaptability to spatially varying haze and insufficient preservation of fine image details. To overcome these issues, we propose an enhanced AOD-Net-based dehazing algorithm. Specifically, we introduce a Fog-Aware Dynamically Constrained Deformable Convolution to better handle spatially varying haze. Inspired by Retinex theory, we further design a lightweight Detailed Enhancement Module that decomposes an image into illumination and reflectance components, enabling separate enhancement of global brightness and local details. In addition, we employ a composite loss that combines MS-SSIM and L1 losses to better preserve structural information during training. Experiments on the RESIDE benchmark show that our method achieves higher PSNR and SSIM than AOD-Net, while producing dehazed images with improved naturalness and visual quality.