Siamese Networks for identification of Holstein cattle during growth and across different physiological stages

In dairy production, accurate individual animal identification is essential for farm management and addressing concerns related to food security and consumer trust. While computer vision systems (CVS) have been proposed for non-invasive animal recognition, limited research has explored their capability to identify the same individual across various life stages. This study aims to bridge this gap by developing a CVS using Siamese Neural Networks (SNN), designed for open-set identification of Holstein calves based on images captured during their initial weeks of life, with the capability to recognize these same individuals after a year of growth. The training dataset consisted of top-down view infrared images of 51 calves aged 1 to 6 weeks, collected on six separate days, resulting in 300 images per calf. These images were used to train a SNN for individual identification with 12,000 image pairs, implemented in Python using Tensorflow and Keras. The trained model was tested on 10 infrared images of each animal after one year (60 weeks of age) using 5 support images captured during their initial weeks of life. Furthermore, the network's ability to recognize individuals not seen in the training set was assessed using 14 additional animals. The results show an F1-score of 73% for identifying individual calves after a year and 83% for recognizing individuals outside the 51-animal training group. These findings highlight the effectiveness of SNN in open-set animal identification and across life stages, demonstrating the potential of reliable traceability systems for animals throughout their lifespan using CVS.