Ocular Biometrics

A distant or visible-light camera often cannot resolve iris fibers, especially for dark irides. The eye region still carries identity in the sclera vessels, eyelid shape, and periocular skin. The group studies these signals on their own and as a fusion with iris features.

Sclera-TransFuse combines a Swin transformer and a convolutional network for sclera segmentation, and received the Best Student Paper Award at IJCB 2023. Earlier models, including ScleraSegNet, targeted accurate vessel masks on mobile images. Multi-modal ocular recognition reconstructs spatial features and estimates image quality without extra labels.

CASIA-Iris-Africa and CASIA-Face-Africa were built so that algorithms are not evaluated only on East Asian and Caucasian cohorts. The iris collection contains 28,717 images of 1,023 African subjects, with age, gender, and ethnicity labels.