An enhanced YOLOv8 model for accurate detection of solid floating waste.

Area of Science:

• Environmental Science
• Computer Vision
• Artificial Intelligence

Background:

• Detecting floating waste on water surfaces presents challenges due to small object sizes, irregular shapes, and background interference.
• Existing methods struggle with accurate identification and localization in complex aquatic environments.

Purpose of the Study:

• To develop an enhanced object detection model for improved floating waste identification.
• To address limitations in detecting small-scale, irregularly shaped floating debris in aquatic environments.

Main Methods:

• Proposed ES-YOLOv8, an enhanced model based on YOLOv8s, optimizing feature fusion with a "160-80-40-20" framework.
• Integrated the Efficient Multiscale Attention (EMA) module for enhanced feature extraction of small objects.
• Employed a novel Shape-IoU loss function for improved bounding box regression accuracy of irregular targets.

Main Results:

• ES-YOLOv8 demonstrated significant improvements over baseline YOLOv8s, with a 5.4% increase in mAP@0.5 and a 6.1% increase in mAP@0.5:0.95 on a custom dataset.
• Comparative experiments confirmed the model's superiority against state-of-the-art methods.
• Validation on public datasets confirmed the robustness and generalizability of the ES-YOLOv8 algorithm.

Conclusions:

• ES-YOLOv8 offers a high-precision, low-power solution for intelligent water governance and floating waste management.
• The model's enhancements in feature extraction and localization provide a promising technological advancement for ecological and engineering applications.