What question did this study set out to answer?

The aim is to enhance real-time detection of complex maritime targets through improved accuracy and robustness.

March 5, 2026Open Access

Marine-YOLO: A high-precision object detection algorithm for complex maritime environments

Key Points

The aim is to enhance real-time detection of complex maritime targets through improved accuracy and robustness.
Developed a C3k2-SP module for better feature representation and robustness.
Integrated SAGA attention for optimizing multi-scale target modeling.
Created an AFAE-Head module with a P2 layer to enhance small-target detection under adverse conditions.
Tested Marine-YOLO against the WSODD dataset for performance evaluation.
Achieved 76.7% mAP50 and 44.1% mAP50–95 on WSODD, showing improvements over YOLOv11n.
Marine-YOLO features more parameters and computational costs than YOLOv11n, yet remains efficient compared to other models.
Demonstrated higher accuracy while significantly reducing parameters and computation compared to RT-DETR-X.

Abstract

• Proposed Marine-YOLO, specifically designed for real-time detection of complex maritime targets, achieving a good balance between accuracy and efficiency. • Designed the C3k2-SP module, enhancing feature representation and environmental robustness through the SC-Dual mechanism. • Introduced SAGA attention, combining axial and efficient channel mechanisms to optimize multi-scale target modeling. • AFAE-Head module adds a P2 layer, effectively addressing small-target miss detection under adverse weather conditions. Sea Surface Object Detection is of great significance for intelligent shipping, marine monitoring, and search and rescue. However, in complex sea-surface scenarios, challenges remain, such as severe weather, illumination changes, large variations in object scale, and missed detections of small targets. To address these issues, this paper proposes Marine-YOLO based on YOLOv11 to improve detection accuracy and robustness in complex sea-surface environments. First, Marine-YOLO introduces a C3k2-SP module into the backbone network to enhance feature representation and environmental robustness. Second, a SAGA attention module is added to the neck to strengthen multi-scale modeling capability. Finally, an AFAE-Head module is designed in the detection head, and a P2 layer is incorporated to optimize small-object detection performance. Experimental results on the WSODD dataset show that Marine-YOLO achieves 76.7% and 44.1% on mAP50 and mAP50–95, respectively, representing improvements of 3.5% and 1.3% over YOLOv11n. While achieving higher accuracy, Marine-YOLO (4.2 M parameters and 13.3 GFLOPs) has more parameters and computational cost than YOLOv11n (2.6 M parameters and 6.5 GFLOPs), but still far less than other high-performance models. Compared with RT-DETR-X (67.3 M parameters and 232.4 GFLOPs), Marine-YOLO maintains higher accuracy while reducing parameters and computation by about 94%, and improving mAP50 and mAP50–95 by 12.2% and 13.9%, respectively. The results indicate that Marine-YOLO achieves a good balance between accuracy and efficiency, making it an ideal choice for real-time marine object detection.

Ask AI

Helpful

Bookmark

View Full Paper