Automated Identification of Individuals in Wildlife Population Using Siamese Neural Networks
Nkosikhona Dlamini, Terence L. van Zyl
Abstract
Similarity learning coupled with semi-hard pair mining has been applied successfully in human individual identification using images of faces. This approach is coupled with innovative training data sampling techniques, trained to optimise a ranking loss function, aimed at increasing model performance at a minimal additional computational cost. We demonstrate that similarity learning coupled with semi-hard negative pair mining, minimising a triplet loss function, can be applied in the identification of wild animals: Lions, Zebra, Nyalas, and Chimpanzees. There is varying performance depending on the dataset being studied and the network architecture. There is improved performance on models trained using semi-hard triplets on the Chimpanzees hold out test-set data; VGG-19 achieves a 96% accuracy and DenseNet-201 90.1% accuracy. Mean average precision was measured for the different network architectures, varying performances were obtained depending on dataset and network depth.