Deep learning-based holstein face recognition in real-world farming conditions

Abstract

Non-contact biometric identification of cattle using visual facial features is now feasible due to recent advances in computer vision and deep learning. However, the lack of publicly available datasets continues to prevent the development and fair evaluation of robust models. This study attempts to address this gap by proposing Holstein2025, a benchmark dataset of individual cow face images captured using fixed-position side-view surveillance cameras in a commercial dairy farm. Holstein2025 reflects real-world environmental variability and supports both face detection and identification tasks. A systematic evaluation of state-of-the-art backbones and face recognition loss functions was conducted. The final pipeline integrates YOLO11n with an oriented bounding box head for cow face detection and alignment (average precision of 0.995, processing time of 11.2 ms), and a ConvNeXt-Tiny-based face identification network trained with an optimised ArcFace-based loss incorporating two auxiliary terms (accuracy of 0.97, processing time of 0.6 ms per face). Overall, the system runs in real time and achieved a precision of 0.98 in open set testing at a similarity threshold of 0.807. To promote reproducibility and practical application, all code and data are open-sourced and released with this paper.