Individual Cow Identification Using Non-Fixed Point-of-View Images and Deep Learning

Monitoring and traceability are crucial for ensuring efficient and financially beneficial cattle breeding in contemporary animal husbandry. While most farmers rely mainly on ear tags, the development of computer vision and machine learning methods opened many new noninvasive opportunities for the identification, localization, and behavior recognition of cows. In this paper, a series of experimental analyses are presented aimed at investigating the possibility of identification of cows using non-fixed point-of-view images and deep learning. 14 objects were chosen and a photo session was made for each one, which provides training/validation images with different viewing angles of the animals. Next, a darknet-53-based convolutional neural network (CNN) was trained using YOLOv3, capable of identifying the investigated objects. The optimal model achieved 92.2% accuracy when photos of single or grouped non-overlapping animals were used. On the other hand, the trained CNN showed poor performance with group images, containing overlapping cows. The obtained results showed that cows could be reliably recognized using non-fixed point-of-view images, which is the main novelty of this study; however, certain limitations exist in the usage scenarios.

@article{Kalmukov2024,
  title     = {Individual Cow Identification Using Non-Fixed Point-of-View Images and Deep Learning},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {15},
  number    = {10},
  year      = {2024},
  publisher = {The Science and Information Organization},
  author    = {Yordan Kalmukov and Boris Evstatiev and Seher Kadirova},
  doi       = {10.14569/IJACSA.2024.0151066},
  url       = {https://doi.org/10.14569/IJACSA.2024.0151066}
}