Dairy DigiD – AI-Driven Facial and Video-Based Recognition for Monitoring Physiological and Reproductive States in Dairy Cattle

Abstract

Precision livestock farming increasingly depends on non-invasive, high-fidelity systems capable of monitoring cattle with minimal disruption to behavior or welfare. Conventional identification methods, including ear tags and wearable sensors, often compromise animal comfort and generate inconsistent data under real-world farm conditions. This study introduces Dairy DigiD, a deep learning-based biometric classification framework that categorizes dairy cattle into four physiologically and age-defined groups—young, top-yielding, late gestation, and non-lactating cows. It described a new classification method based on full-body classification using DenseNet121, the use of facial landmark detection from Detectron2, and the real-time detection capabilities of YOLOv11. Dairy DigiD is a multi-model pipeline that is flexible and can operate from either static images or video data. It started with assembling a dataset of more than 8,700 annotated images from commercial dairy farms in Nova Scotia and New Brunswick. DenseNet121 was trained for a baseline for full-image classification, but background variation and occlusion led to considerations of using keypoint-based facial recognition with Detectron2 to improve identification by having model stability on a 30-keypoint landmarks (e.g., eyes, nostrils, and muzzle), resulting in less sensitivity to posture, lighting, and weight variation. For video-based detection, the novelty of using YOLOv11 (latest architecture) for multi-class physiological cattle classification. Roboflow-annotated data and deployed through Gradio on Hugging Face. Built as an interactive interface for real-time streaming, detection overlays, and roles for farm users, YOLO was developed as a useful system. The analysis showed that DenseNet121 had great baseline accuracy (97%); however, in contrast, Detectron2 demonstrated superior adaptability in uncontrolled farm environments, attaining classification accuracies between 93% and 98% on test datasets, and YOLOv11 provided the best performance as a real-time capable method (20 FPS and a maximum latency of 230 ms). The thesis also assesses model explainability on the basis of Grad-CAM, cross-validation accuracy and the impact of dataset imbalance. Ethical considerations, limitations, and future integration possibilities with multimodal systems—such as gait analysis and 3D biometrics—are also acknowledged. Overall, Dairy DigiD represents a critical innovation in precision livestock monitoring and data-based welfare, traceability, and farm automation with minimal animal distress.