On Assessing the Impact of Ocular Pathologies on the Performance of Deep Learning Ocular Based Recognition Systems in the Visible and NIR Bands

Deep learning techniques can extract highly discriminative iris features for the task of human recognition. The success of these techniques on benchmark iris datasets has generated interest in other unique iris applications, including anti-spoofing, post-mortem, and non-healthy recognition. In this paper, we study the effect of ocular pathologies on datasets with healthy enrolled identities by adding diverse groups of non-healthy identities. We include irises captured in the NIR (near-infrared) spectrum and periocular images captured in the visible spectrum. Non-healthy eyes are selected randomly and by a specific disease from the Warsaw-BioBase-Disease-Iris v2.1 dataset and added to a healthy baseline dataset. ResNet-50 architectures with pretrained weights are fine-tuned on healthy data from the NIR and visible spectrums and used to match healthy and non-healthy subjects for the identification and verification tasks. We find that our trained models work well in all scenarios for verification and that periocular biometrics, in the presence of specific ocular pathologies, can be better suited for the identification task. To our knowledge, we find no other work investigating deep learning techniques on the Warsaw-BioBase-Disease-Iris v2.1 dataset, the largest of its kind in the open literature.