Livestock Identification Using Deep Learning for Traceability.

Area of Science:

• Agricultural Technology
• Computer Vision
• Deep Learning

Background:

• Non-invasive digital technology for livestock identification and welfare assessment is gaining traction in modern agriculture.
• Accurate traceability is crucial for farm management and food safety.
• Existing identification methods may be invasive or less efficient for large herds.

Purpose of the Study:

• To develop a novel face recognition system for dairy cows using advanced deep learning and computer vision.
• To create a non-invasive method for individual animal identification and potential welfare monitoring.
• To adapt human face recognition pipelines for agricultural applications.

Main Methods:

• A video analysis pipeline comprising face detection, cropping, encoding, and lookup was implemented.
• Three deep learning models (face detector, landmark predictor, face encoder) were fine-tuned using transfer learning.
• A custom dairy cow dataset was collected from a robotic dairy farm at The University of Melbourne.

Main Results:

• The developed system achieved an overall accuracy of 84% across videos of 89 distinct dairy cows.
• The computer program was successfully tested on an NVIDIA Jetson Nano edge device with a live camera stream.
• The system's non-invasive nature makes it suitable for integration with existing welfare assessment tools.

Conclusions:

• Deep learning and computer vision offer a viable solution for non-invasive dairy cow identification.
• The developed face recognition system provides accurate traceability and has potential for welfare assessment.
• The system's adaptability to edge computing enhances its practical applicability in real-world farm settings.