YOLO-AR: An Improved Artificial Reef Segmentation Algorithm Based on YOLOv11.

Area of Science:

• Marine ecology
• Artificial intelligence
• Sonar imaging

Background:

• Artificial reefs are vital for marine habitat restoration and fishery resource management.
• Accurate identification of artificial reefs is crucial for ecological conservation.
• Multibeam sonar images are essential for mapping seafloor structures.

Purpose of the Study:

• To develop a precise artificial reef segmentation model for multibeam sonar images.
• To improve the efficiency and accuracy of artificial reef detection in marine environments.
• To create a lightweight yet high-performing deep learning model for marine applications.

Main Methods:

• An improved YOLOv11-based model, named YOLO-AR, was proposed.
• Incorporated Dynamic Convolution Coordinate Attention (DCCA) to enhance robustness in complex seafloor environments.
• Integrated a small-object detection layer with Dynamic Sampling (DySample) and Advanced Downsampling (ADown) for improved segmentation and reduced complexity.

Main Results:

• YOLO-AR achieved a mean Average Precision (mAP@0.5) of 0.912, Intersection-over-Union (IOU) of 0.832, and F1 score of 0.908.
• The model boasts a lightweight structure with 2.67 million parameters and a size of 5.58 MB.
• Demonstrated superior segmentation performance compared to other advanced models in real-world sonar images.

Conclusions:

• YOLO-AR accurately segments artificial reefs in multibeam sonar images.
• The model's lightweight design and high performance make it highly effective for practical marine applications.
• This technology supports marine ranching development and effective fishery resource management.