The field of image-based biometry focuses on recognition of individuals
based on their unique biological characteristics with the usage of images.
Most research is based on human recognition, while less work is focused on
the remaining animal kingdom, especially wild species. We wanted to explore
how well recognition works on one of these animal families. In the master’s
thesis, we explore the possibilities of using image-based biometrics and deep
neural networks, specifically Siamese networks, for the recognition of members
of the wild cat family Felidae. We construct several datasets for teaching the
models, which include different biometric modalities – the entire eye region,
the iris, fur patterns and fur patterns of specific body parts. All datasets
are trained with multiple models and hyperparameters using contrastive loss.
We evaluate the accuracy of image pairs predicted as genuine matches. The
best performance is achieved with recognition based on the entire eye region,
followed by the iris, and finally fur patterns. All datasets reach accuracy
above 85%, with the eye region performing best, achieving 92.76% accuracy
using the ResNet18 model without data augmentation.
|
