Identification of Individual Hanwoo Cattle by Muzzle Pattern Images through Deep Learning

Area of Science:

• Animal Science
• Computer Science
• Artificial Intelligence

Background:

• Individual identification is crucial for livestock management and traceability.
• Traditional methods for cattle identification can be labor-intensive and prone to errors.
• Deep learning offers a potential automated solution for accurate animal identification.

Purpose of the Study:

• To develop and evaluate a deep-learning model for identifying individual Hanwoo cattle using muzzle images.
• To assess the effectiveness of different Efficientnet v2 model sizes and optimizers for this task.
• To determine the potential of muzzle patterns as a unique identifier for Hanwoo cattle.

Main Methods:

• Collected 9230 muzzle images from 336 Hanwoo cattle, with multiple images per individual.
• Utilized YOLO v8 for muzzle image cropping and prepared data for training, validation, and testing.
• Applied transfer learning with Efficientnet v2 (tiny, small, medium) and various optimizers (SGD, RMSProp, Adam, Lion).

Main Results:

• The Efficientnet v2 small model with the Lion optimizer achieved the highest validation accuracy (0.981).
• The Efficientnet v2 small model with the Adam optimizer demonstrated the best test accuracy (0.970).
• The model with RMSProp optimizer showed the lowest repeated error, indicating robustness.

Conclusions:

• Deep learning models utilizing muzzle images can accurately identify individual Hanwoo cattle.
• Muzzle patterns serve as a reliable biometric key for cattle identification.
• This technology has significant potential for enhancing livestock traceability and management systems.