Sheep biometric identification based on multiple body parts

Precision livestock farming technologies require individual animal identification to integrate specific animal data with farm processes, including health, milk recording, and feeding management. Identification methods, such as physical marks and ear tags, require manual handling and are prone to loss or wear. Radio Frequency Identification (RFID) tags enable contactless identification, but still face limitations under farm conditions, including tag loss, maintenance costs, and proximity to the sensor. Recent advances in computer vision and deep learning have led to alternative approaches for animal identification. Earlier studies in small ruminants reported face or tag-based identification, while cattle studies explored multifeature fusion from different body parts. Our study developed an automated system for individual identification of sheep during drinking visits without human intervention, based on computer vision and deep learning techniques. The aims were to evaluate whether sheep could be identified using visual features from separate body regions (face, back, legs), and whether combining predictions from multiple regions with visual tag recognition improved visit-level identification. Data were collected at the Ivry Dairy Farm in Azarya, central Israel, using a water trough with an overhead Intel RealSense D435 RGB-D (red, green, blue and depth) camera and a Jetson Orin device. A total of 287 visits from 85 sheep were recorded. YOLOv8, an object detection model, was used to detect body parts (face, back, leg) and ear tags. ResNet50, a convolutional neural network (CNN) used for visual feature extraction, was applied to generate image embeddings, and the GLASS text spotting and recognition algorithm was used for visual tag text recognition. Leg-based features alone reported a peak accuracy of 0.64, which was lower than other body regions, and their inclusion in combinations yielded only minor and inconsistent effects. Back-based features alone reported a peak accuracy of 0.79, while combining back and face features reported 0.91, and integration of back, face, and visual tag features reported 0.93. The results suggest that sheep could be identified using multiple body regions, and that combining predictions from these regions provided complementary information and higher accuracy. Adding visual tag recognition further improved prediction results under typical farm conditions without human involvement.