Feasibility of using convolutional neural networks for individual-identification of wild Asian elephants

Individual identification is a basic requirement for research in behavior, ecology and conservation. Photographic records are commonly used situations where individuals visually distinct. However, keeping track of identities becomes challenging with increasing population sizes corresponding datasets. There growing interest the potential deep-learning methods computer vision to assist automating this task. Here we apply Convolutional Neural Networks, popular architecture Artificial Networks image classification, problem identifying individual Asian elephants through photographs. We evaluate performance five different types CNN models facial recognition (VGG16, ResNet50, InceptionV3, Xception, Alexnet), on datasets representing three feature regions (the full body, face, ears), trained two techniques (transfer learning vs. training from scratch) n = 56 elephants. tested accuracy matching top candidate as well candidates. found that VGG16 transfer-learning technique outperformed other body face accuracies 21.34% 42.35%, respectively, candidate. Nevertheless, best was achieved by an Xception model scratch ear dataset, 89.02% 99.27% including correct among five. impressive level obtained dataset 3816 labeled images more than 1000 wild under observation, requiring extensive human effort skill initially annotate data. Therefore, consider approach impractical monitoring large populations. Nevertheless it could be very useful record fraud prevention captive elephant populations, animals have been rehabilitated released or moved management purposes.