Enhanced Water Surface Object Detection with Dynamic Task-Aligned Sample Assignment and Attention Mechanisms.

Area of Science:

• Computer Vision and Machine Learning
• Robotics and Autonomous Systems
• Environmental Monitoring

Background:

• Object detection on water surfaces is critical for Unmanned Surface Vehicles (USVs).
• Existing systems face challenges with indistinct bottom boundaries and foggy imagery.
• The YOLOv8s model serves as a baseline for real-time detection.

Purpose of the Study:

• To develop a novel real-time target detection system for USVs.
• To enhance object detection accuracy in challenging aquatic environments.
• To improve the delineation of indistinct bottom boundaries in imagery.

Main Methods:

• Enhanced the YOLOv8s model with Convolutional Block Attention Module (CBAM) and self-attention mechanisms.
• Implemented a dynamic sample assignment strategy for improved precision and convergence.
• Utilized a two-strategy approach (threshold filter and FFN) for precise bottom boundary refinement.

Main Results:

• Achieved a mean average precision (mAP) of 47.1%, a 1.7% increase over the baseline YOLOv8.
• Dynamic sample assignment improved AP^0.5 by 1.0%; FFN strategy improved AP^0.75 by 0.8%.
• Ablation studies confirmed the versatility and potential integration into various detection frameworks.

Conclusions:

• The proposed enhanced YOLOv8s model significantly improves real-time object detection on water surfaces.
• The integration of attention modules and novel strategies effectively addresses challenges like fog and indistinct boundaries.
• The system demonstrates robust performance and adaptability for USV perceptual systems.