Low Visibility Enhancement for Intelligent Marine Vehicles in Hazy Environments⁎

Abstract

Visible-light vision sensors have emerged as a cost-effective and easily deployable solution for enhancing marine vessel navigation safety. However, their effectiveness is substantially compromised in hazy marine environments, where captured images often exhibit severe contrast reduction, color distortion, and interference from sea surface waves and specular reflections. These degradation effects collectively impair visual perception accuracy and range, posing risks to navigational safety. To address these challenges, this paper proposes MDNet, a Maritime Dehazing Network designed to improve visual clarity for maritime vehicles in low-visibility hazy conditions. The MDNet framework integrates two core components: a Multi-receptive-field feature enhancement block (MFEB) and a multi-expert collaborative interaction block (MCIB). The MFEB leverages parallel multi-branch processing to perform multi-scale feature extraction, enabling the network to simultaneously capture global contextual information and fine-grained texture details. The MCIB further enhances adaptability by dynamically selecting optimal expert modules tailored to learn degradation features at different network depths, thereby effectively modulating multi-granularity contextual information in degraded images. Extensive experiments validate that MDNet achieves state-of-the-art performance, outperforming existing methods in both quantitative metrics and qualitative visual assessments. Additionally, the proposed method demonstrates practical utility by significantly boosting the accuracy and reliability of target detection in hazy marine scenarios, thereby advancing the safety and stability of autonomous navigation systems.