A dual-branch feature fusion learning framework for waterway image dehazing
Ziqiang Huang, Wei Liu, Hangyu Nie, Miao Qi, Guoping Qiu · Autonomous Transportation Research · 2025
Haze severely degrades image quality in waterway environments, affecting the reliability of vision-based navigation systems. To address this challenge, we propose LG-DBNet, a Local-Global Dual-Branch Network tailored for waterway image dehazing. The network integrates four key innovations: a Star-enhanced multi-scale convolutional block, a global modeling module based on Swin Transformer, a gated modulation fusion module, and a color enhancement module based on wavelet transform. LG-DBNet adopts a dualbranch architecture to capture both local texture details and global structural information, which is crucial for handling water reflections and large-scale sky–water distributions in waterway scenes. Specifically, the local branch uses convolutional operations and the Star operation to enhance boundary restoration and fine-grained features, especially in low-texture or highly reflective areas. Meanwhile, the global branch, built upon Swin Transformer, models long-range dependencies to preserve scene consistency and restore distant targets. The gated fusion module then adaptively balances local and global features, ensuring effective integration of complementary representations. In addition, the color enhancement module refines image appearance and improves overall restoration quality by restoring natural color tones and reducing over-enhancement. Furthermore, we construct SeaShip-Haze, a synthetic dataset based on real-world waterway images with simulated haze at three density levels. Extensive experiments on both synthetic and real-world hazy images demonstrate that LG-DBNet consistently outperforms existing state-of-the-art methods in both quantitative metrics and visual quality.