An offline deep learning model for bovine muzzle identification using the variance of the laplacian, YOLOv8, EfficientNetV2-S, and ArcFace

Abstract Precise animal identification underpins livestock management, disease control, and traceability. Traditional methods such as tattoos and ear tags can fade or be lost, motivating a shift toward biometric-based systems. This study proposed an offline Python-based deep learning biometric identification model for bovine using muzzle images. YOLOv8 detected and cropped the muzzle region, EfficientNetV2-S architecture was integrated with ArcFace (angular margin loss) for muzzle pattern extraction and matching. The model was evaluated under two conditions: a controlled dataset (Kaggle) and a preliminary, small-scale field dataset using smartphone-acquired images from smallholder farmers. The model was validated using the Kaggle dataset and tested on both Kaggle and field-collected muzzle images. Model performance was evaluated using the cosine similarity score, processing time, F1 score, recall, accuracy, and precision. The model showed shorter processing times and higher recognition performance on the Kaggle images than on the field images. A domain gap was observed between Kaggle dataset and field dataset. The current model is a controlled prototype with preliminary field testing for offline muzzle recognition; however, it has not yet met pre-defined field readiness measures. Overall, the proposed offline muzzle biometric model provides a non-invasive prototype for bovine identification under the tested conditions.