LGD-Net: a local-global feature interaction network for image dehazing
Guobo Xie, Qiao Huang, Zhiyi Lin · The Imaging Science Journal · 2025
Image dehazing, crucial for enhancing image recognizability and preserving information integrity, is a pivotal research area in computer vision. Despite deep learning advancements, challenges persist in detail restoration, non-uniform haze identification, and tone recovery. To address these challenges, this paper proposes a Local and Global feature encoding-interaction Dehazing Network (LGD-Net). First, a Channel-Pixel Joint Attention Module (CPJAM) is designed in the encoder to extract local features from hazy images. Second, a Transformer-based encoding branch is constructed, incorporating Residual DehazeFormer Blocks (RDB) to integrate global information into featuresFinally, a Dynamic Fusion Module (DFM) is added during decoding, utilizing PConv to guide the fusion of skip-connection features into the main branch. Experiments on synthetic and real-world datasets show LGD-Net outperforms state-of-the-art methods in metrics and visual quality. On the O-HAZE dataset, it achieved 25.14 dB PSNR and 0.791 SSIM, at least 1.69 dB and 0.042 higher than others.