Development of an artificial intelligence deep neural network for the identification of individual animals

Abstract

The literature review highlighted a clear need for a system capable of identifying individual animals. Such a system will enable a farmer to feed specific livestock animals, such as cows, to prevent under- and overfeeding, and to track cattle for stock theft prevention and identification if found. Other current identification methods, such as ear tags and RFID tags, were investigated and considered, but not used. This research further explored the application of deep learning, specifically the You Only Look Once (YOLO) family of models, for real-time animal identification and monitoring in automated feeding systems. Traditional animal recognition methods, such as RFID tags and manual observation, are limited by accuracy, maintenance costs, and scalability. To address these challenges, this study investigated the potential of YOLOv3 and YOLOv4 architectures on portable hardware for detecting and identifying individual cattle using image and video data. A comprehensive review of object detection frameworks, including R-CNN and SSD, was conducted to establish YOLO’s advantages in terms of speed, accuracy, and ease of deployment. The methodology involved dataset creation, preprocessing, augmentation, and training of YOLO models using Python and GPU-accelerated platforms such as the NVIDIA® Jetson Nano. Model performance was evaluated using precision, recall, and mean average precision (mAP) metrics, with results demonstrating YOLOv4’s superior real-time detection capability and robustness under varying environmental conditions. The YOLO-based detection system demonstrates strong overall performance across most classes, with the majority achieving near-perfect detection accuracy. Fourteen classes attain an Average Precision (AP) of 100% alongside consistent Area Under the Precision–Recall Curve (AUC) values of 0.75, indicating highly reliable and stable detection behaviour. Several additional classes also perform well, recording AP values above 90% with only a modest reduction in AUC, which suggests robust generalisation with minimal precision–recall degradation. However, performance declines noticeably for a small number of classes, where AP falls to 50% and below, accompanied by significantly reduced AUC values. Three classes show no successful detections, yielding zero AP and AUC scores. Overall, these results highlight the effectiveness of the YOLO system for most target classes, while indicating that limited or challenging classes may benefit from additional training data, improved class balance, or further model optimisation. The findings confirmed that YOLO-based vision systems can significantly enhance the automation and efficiency of livestock management by enabling accurate, non-invasive identification and behavioural monitoring. The study contributes to the advancement of intelligent feeding systems and provides a foundation for future research in precision agriculture and AI-based animal welfare monitoring.