YOLOv10n-CF-Lite: A Method for Individual Face Recognition of Hu Sheep Based on Automated Annotation and Transfer...

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Individual recognition of Hu sheep is crucial for precision livestock management.
• Traditional machine vision methods for sheep recognition are time-consuming and require extensive retraining for new individuals.

Purpose of the Study:

• To develop an efficient sheep face recognition algorithm integrating automatic annotation and transfer learning.
• To address challenges in manual annotation time costs and adaptability to new sheep individuals.

Main Methods:

• Created a dataset of 6055 images from 82 Hu sheep with multi-view video collection and data augmentation.
• Designed a sheep face detection and automatic annotation algorithm, reducing annotation time to 0.014 min/image.
• Employed the YOLOv10n-CF-Lite model and applied transfer learning to adapt the model to new sheep individuals.

Main Results:

• Achieved 92.3% recognition precision and 96.2% mAP@0.5 for Hu sheep faces using the YOLOv10n-CF-Lite model.
• Transfer learning improved recognition precision to 94.9% and mAP@0.5 to 97% in the target domain (new sheep).
• Reduced training epochs from 43 to 14, significantly improving model convergence speed.

Conclusions:

• The proposed algorithm enhances annotation efficiency, recognition precision, and convergence speed for Hu sheep face recognition.
• Automatic annotation and transfer learning enable rapid adaptation to new sheep individuals, supporting intelligent livestock management.
• This provides an efficient and reliable technical solution for precision livestock farming.