Individual identification of dairy cows with occluded camera views using open-set contrastive learning model

Area of Science:

• Agricultural technology
• Computer vision
• Deep learning
• Livestock management

Background:

• Individual identification of dairy cows is crucial for effective livestock management and precision farming.
• Traditional identification methods can be labor-intensive, costly, and prone to errors.
• Challenging visual conditions in free stall barns (occlusions, lighting, poses) hinder accurate identification.

Purpose of the Study:

• To develop and evaluate computer vision techniques for individual dairy cow identification in free stall barns.
• To compare the performance of two deep learning models: Xception (classification) and contrastive learning (similarity).
• To assess the models' robustness and generalization capabilities in both closed-set and open-set scenarios.

Main Methods:

• Collected a dataset of 500 images per cow per day for 49 cows over 4 days.
• Utilized YOLOv8 for initial animal detection and annotation, combined with RFID data.
• Trained and evaluated two deep learning models: Xception for multi-class classification and a Siamese network for contrastive learning.

Main Results:

• The Xception model achieved high accuracy in a closed-set scenario (e.g., 88% accuracy).
• The contrastive learning model demonstrated superior performance in an open-set scenario (100% accuracy), successfully identifying unknown individuals.
• Computer vision offers a scalable, non-invasive alternative to traditional methods, improving efficiency and data collection.

Conclusions:

• Both Xception and contrastive learning models show promise for automated dairy cow identification.
• Contrastive learning excels in real-world, dynamic herd compositions due to its open-set identification capabilities.
• Computer vision integration enhances precision livestock farming by enabling automated phenotyping, tracking, and informed decision-making.