Lightweight Sheep Face Recognition Model Combining Grouped Convolution and Parameter Fusion.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Agricultural Technology

Background:

• Sheep face recognition is vital for individual identification and behavior monitoring.
• Existing models demand high computational resources, hindering deployment on mobile or embedded devices.
• This leads to reduced accuracy and increased recognition time in practical applications.

Purpose of the Study:

• To develop a lightweight and efficient sheep face recognition model.
• To overcome the computational and accuracy limitations of current models on resource-constrained devices.
• To introduce an improved model based on YOLOv8n architecture.

Main Methods:

• Proposed the Parameter Fusion Lightweight You Only Look Once (PFL-YOLO) model, an enhancement of YOLOv8n.
• Integrated Efficient Hybrid Conv (EHConv) and Residual C2f (RC2f) modules to improve feature extraction and multi-scale fusion.
• Developed a Parameter Fusion Detection (PFDetect) module to reduce model parameters and computational complexity.

Main Results:

• PFL-YOLO achieved a performance-efficiency balance with mAP@50 of 99.5% and mAP@50:95 of 87.4%.
• The model boasts only 1.01 M parameters and a size of 2.1 MB, significantly reducing computational load.
• Parameter count and model size were reduced by up to 83.7% and 82.5% compared to various lightweight models.

Conclusions:

• The PFL-YOLO model offers high accuracy and efficiency for sheep face recognition.
• Its lightweight nature makes it suitable for deployment on resource-constrained devices.
• PFL-YOLO presents a viable new solution for advanced sheep monitoring systems.