A Lightweight and High-Precision Passion Fruit YOLO Detection Model for Deployment in Embedded Devices

Area of Science:

• Computer Vision
• Machine Learning
• Agricultural Technology

Background:

• Accurate and efficient detection of passion fruit in complex environments is crucial for agricultural automation.
• Existing models often struggle with challenges like backlighting, occlusion, and varying weather conditions, limiting their real-world application.

Purpose of the Study:

• To develop a lightweight and high-accuracy object detection model for passion fruit in challenging orchard settings.
• To improve detection speed and average precision on embedded devices for real-time agricultural tasks.

Main Methods:

• Replaced the YOLOv5 backbone with a lightweight GhostNet model to reduce parameters and computational load.
• Introduced a new feature branch and reconstructed the neck network's feature fusion layer to enhance feature integration.
• Employed knowledge distillation to transfer knowledge from a teacher model to the G-YOLO-NK student model, boosting accuracy.

Main Results:

• The G-YOLO-NK model achieved 96.00% average accuracy, surpassing the original YOLOv5s by 1.00%.
• Model size was reduced by half to 7.14 MB, with a real-time detection rate of 11.25 FPS on Jetson Nano.
• Demonstrated superior performance compared to state-of-the-art models in average precision and detection capabilities.

Conclusions:

• The G-YOLO-NK model offers an effective solution for real-time passion fruit detection in complex orchard environments.
• Provides valuable technical support for developing intelligent orchard picking robots and enhancing overall orchard intelligence.