FEB-YOLOv8: A multi-scale lightweight detection model for underwater object detection.

Area of Science:

• Computer Vision
• Robotics
• Marine Technology

Background:

• Underwater object detection is vital for marine resource management.
• Underwater robots face challenges with limited storage and computational power.
• Existing detection models often struggle with efficiency and accuracy in marine environments.

Purpose of the Study:

• To develop a novel lightweight object detection model for underwater robots.
• To enhance the efficiency and accuracy of underwater object detection systems.
• To address the constraints of limited storage and computational power in underwater robotic applications.

Main Methods:

• Proposes FEB-YOLOv8, a lightweight model based on the YOLOv8 framework.
• Enhances the backbone network with refined C2f and novel P-C2f modules.
• Incorporates the EMA module to improve multi-scale feature extraction and a Bi-FPN-inspired feature pyramid network for balanced performance.

Main Results:

• FEB-YOLOv8 achieved a 1.2% and 1.3% increase in mAP on DUO and URPC2020 datasets, respectively.
• Reduced computational load with 6.2G GFLOPs (24.39% decrease) and 1.64M parameters (45.51% decrease) compared to the baseline.
• Demonstrated a significant improvement in the balance between model lightness and detection precision.

Conclusions:

• FEB-YOLOv8 offers an advantageous solution for underwater object detection.
• The model effectively harmonizes lightness with accuracy, suitable for resource-constrained underwater robots.
• The proposed modifications enhance feature extraction and detection performance in marine environments.