Improved YOLOv5 s and transfer learning for floater detection.

Area of Science:

• Environmental Science
• Computer Vision
• Machine Learning

Background:

• Floating objects like plastic bottles and dead fish degrade water quality and harm ecosystems.
• Manual detection and cleanup are inefficient, costly, and pose risks.

Purpose of the Study:

• To develop an efficient, accurate, and real-time system for detecting and classifying floating objects on water surfaces.
• To improve upon existing object detection models for environmental monitoring applications.

Main Methods:

• An improved YOLOv5s model was developed using a curated dataset of floating object images.
• Optimizations included integrating EfficientNetv2, advanced upsampling modules, bidirectional feature pyramid networks, attention mechanisms (SE, EMA), and SIoU loss.
• Transfer learning and data augmentation (SAGAN) were employed to enhance model performance.

Main Results:

• The enhanced YOLOv5s model achieved a 5.27 percentage point increase in accuracy compared to the original model.
• The improved model reduced parameter count by 53.9%, computational load by 21.3%, and weight size by 54%.
• Ablation experiments confirmed the effectiveness of individual improvements.

Conclusions:

• The proposed model offers an efficient, accurate, and real-time solution for floating object detection on water surfaces.
• This methodology is crucial for monitoring aquatic environments and managing floating debris effectively.
• The study provides valuable insights for precise and efficient detection and classification of aquatic surface pollutants.