A method for dairy cow identification based on multiview feature fusion of the cow's back

Area of Science:

• Computer Vision
• Animal Science
• Machine Learning

Background:

• Single-view cow back identification faces challenges like occlusion and posture variation.
• Accurate dairy cow identification is crucial for efficient farm management.

Purpose of the Study:

• To develop a robust dairy cow identification method using multi-view feature fusion.
• To overcome limitations of single-view identification in dynamic milking environments.

Main Methods:

• Constructed a multi-view dataset of cow backs (31,356 images, 1954 cows).
• Employed YOLO11n and ByteTrack for adaptive image acquisition (central, left, right views).
• Evaluated 7 feature extraction backbones, selected top 3, and optimized single-view models before feature concatenation and cosine distance matching.

Main Results:

• Achieved 99.69% overall accuracy and 99.83% mAP.
• Demonstrated 94% accuracy on samples unidentifiable by single-view models.
• Successfully identified cows in complex milking scenarios.

Conclusions:

• The proposed multi-view feature fusion method significantly enhances dairy cow identification accuracy.
• This approach provides an effective solution for cow identification in real-world milking scenarios.
• The study offers a valuable tool for improving dairy farm management and efficiency.