Abstract
Marine pollution has become an increasingly severe environmental issue, with oil spills, marine debris, and turbid water significantly impacting ecosystems and human health. The You Only Look Once (YOLO) series of target detection has been widely applied in Marine pollution monitoring. However, in complex underwater environments, challenges such as irregular pollutant shapes, varying scales, and background interference limit detection accuracy and robustness. To address these issues, this study proposes an improved YOLOv11 model that integrates Deformable Convolutional Networks version 4 (DCNv4) to enhance adaptability to deformable pollutants, improving detection precision. The Marine Fusion Loss (MFL) mechanism optimizes detection weight allocation among different pollutant categories, reducing false positives. Additionally, Multi-scale Feature Fusion (MFF) combines Convolutional Neural Networks (CNN) and Transformer-based feature extraction to enhance robustness in complex environments. Furthermore, instance segmentation is incorporated to refine boundary detection of pollutants. Experiments show that the improved YOLOv11 model outperforms the most advanced methods such as YOLOv8 and YOLOv10, with an average accuracy of 90.2% when 50% intersection exceeds union (mAP50) and an inference speed of 3.5ms, ensuring high precision and high efficiency. The results validate the effectiveness of the proposed method in enhancing marine pollution detection, providing a high-performance solution for intelligent environmental monitoring.