Towards a physical imaging-driven sparse attention dehazer for Internet of Things-aided Maritime Intelligent Transportation

Yuxuan Tian, Yu Guo, Yuxu Lu, Yuan Gao, Ryan Wen Liu

Abstract: In the field of Maritime Intelligent Transportation Systems (MITS), the integration of Internet of Things (IoT) technologies and intelligent algorithms has revolutionized visual IoT-aided MITS. This integration, enabled by advanced communication technologies, network infrastructures, sensor capabilities, and data science methodologies, has significantly enhanced monitoring, navigation, and collision avoidance systems, thus improving waterway transportation efficiency. However, the performance of these systems can be hampered by atmospheric conditions, leading to degraded imaging quality characterized by contrast reduction, color distortion, and object invisibility. Such challenges impede critical vision-based tasks like object detection, tracking, and scene understanding in MITS. To address the performance gap between clear and hazy scenes, we propose a novel framework called PSDformer. This framework integrates Top-K Sparse Attention with a Physics-Aware Feed-Forward Network to enhance performance under hazy conditions. Additionally, we introduce a novel paired data generation method to reduce the disparity between synthetic and real-world data. Experimental results on synthetic and real-world datasets demonstrate that PSDformer outperforms existing state-of-the-art methods in both qualitative and quantitative evaluations. Importantly, its exceptional dehazing capability significantly improves detection accuracy under adverse hazy conditions, thereby addressing a critical challenge in visual IoT-aided MITS.