Collaborative Optimization of Model Pruning and Knowledge Distillation for Efficient and Lightweight Multi-Behavior...

Area of Science:

• Agricultural Technology
• Computer Vision
• Animal Science

Background:

• Accurate piglet behavior monitoring is vital for intensive pig farming, but current models are computationally intensive.
• Existing piglet behavior recognition models face limitations in practical farming applications due to high resource demands.

Purpose of the Study:

• To develop a computationally efficient and accurate piglet multi-behavior recognition model for intensive farming.
• To balance model compression with high detection accuracy and speed for real-time applications.

Main Methods:

• Utilized the LAMP pruning algorithm to create a lightweight YOLOv8-Prune model.
• Integrated the AIFI module and Gather-Distribute mechanism into YOLOv8, forming YOLOv8-GDA.
• Employed knowledge distillation with YOLOv8-GDA as the teacher and YOLOv8-Prune as the student to create the YOLOv8-Piglet model.

Main Results:

• YOLOv8-Piglet achieved a 6.3% increase in precision, 11.2% increase in recall, and 91.8% mAP@0.5.
• Inference time was reduced by 53.8% (from 353.9 ms to 163.2 ms) on an NVIDIA Jetson Orin NX platform.
• Demonstrated significant model compression without compromising recognition accuracy.

Conclusions:

• The YOLOv8-Piglet model offers an effective balance between detection accuracy and processing speed for piglet behavior recognition.
• This approach, combining pruning and knowledge distillation, optimizes model performance for edge computing in agricultural settings.
• The developed model enhances the feasibility of advanced behavior monitoring in real-world intensive pig farming environments.