Lightweight model-based sheep face recognition via face image recording channel.

Area of Science:

• Computer Vision and Machine Learning
• Agricultural Technology
• Animal Science

Background:

• Accurate individual sheep identification is essential for digital and precision livestock farming.
• Current deep learning models for sheep face recognition are often too large and computationally expensive for practical use.
• Existing methods lack efficiency in data collection and model optimization for real-world applications.

Purpose of the Study:

• To develop a lightweight and efficient sheep face recognition model for digital sheep farming.
• To address the limitations of existing models regarding size, computational cost, and data acquisition.
• To improve the accuracy and speed of sheep identity recognition in livestock management.

Main Methods:

• Developed a streamlined face image recording channel for efficient data collection.
• Created a sheep face dataset of 22,000 images from 50 Small-tailed Han sheep using data augmentation.
• Introduced a lightweight model, YOLOv7-Sheep Face Recognition (YOLOv7-SFR), incorporating attention modules (shuffle attention, Dyhead), depthwise separable convolutions, and knowledge distillation with YOLOv7 as the teacher model.

Main Results:

• The YOLOv7-SFR model achieved a mean average precision@0.5 of 96.9% on the sheep face dataset.
• The model boasts a small size (11.3 MB) and fast average recognition time (3.6 ms).
• YOLOv7-SFR demonstrated a 2.1% higher precision, 5.8% smaller size, and 42.9% faster recognition compared to YOLOv7-tiny.

Conclusions:

• The proposed YOLOv7-SFR model offers a significant advancement in lightweight and efficient sheep face recognition.
• The developed model meets the practical requirements for digital sheep farms and precision livestock farming.
• This research is expected to accelerate the adoption of sheep face recognition technology in the agricultural sector.