Attention-Based Lightweight YOLOv8 Underwater Target Recognition Algorithm.

Area of Science:

• Computer Vision
• Robotics
• Marine Technology

Background:

• Underwater object detection presents significant challenges due to complex environments.
• Existing models often struggle with high computational complexity, slow speeds, and low accuracy.

Purpose of the Study:

• To propose an efficient and accurate underwater target detection model, SPSM-YOLOv8.
• To address limitations in computational complexity, detection speed, and accuracy of current underwater detection systems.

Main Methods:

• Utilized SPDConv module in the backbone for efficient feature extraction.
• Integrated Polarized Self-Attention (PSA) mechanism to enhance feature polarization and pixel-level prediction accuracy.
• Introduced spatial-channel decoupled downsampling (SCDown) to reduce computational cost while preserving information.
• Employed Minimum Point Distance-based IoU (MPDIoU) loss function for faster convergence and improved bounding box regression.

Main Results:

• SPSM-YOLOv8 achieved 87.3% accuracy on the ROUD dataset and 76.4% on the UPRC2020 dataset.
• Demonstrated a 4.3% reduction in parameters and a 4.9% decrease in computation compared to the YOLOv8n baseline.
• Achieved a detection frame rate of 189 frames per second on the ROUD dataset.

Conclusions:

• SPSM-YOLOv8 significantly improves detection accuracy and speed for underwater objects.
• The model's lightweight and fast nature facilitates efficient deployment on edge devices.
• The proposed enhancements meet the high accuracy and speed requirements for advanced underwater object detection.