Learning Efficient Deep Feature Extraction For Mobile Ocular Biometrics

Abstract

Ocular biometrics uses physical traits from eye regions such as iris, conjunctival vasculature, and periocular for recognizing the person. Ocular biometrics has gained popularity amongst research and industry alike for its identification capabilities, security, and simplicity in the acquisition, even using a mobile phone's selfie camera. With the rapid advancement in hardware and deep learning technologies, better performances have been obtained using Convolutional Neural Networks(CNN) for feature extraction and person recognition. Most of the early works proposed using large CNNs for ocular recognition in subject-dependent evaluation, where the subjects overlap between the training and testing set. This is difficult to scale for the large population as the CNN model needs to be re-trained every time a new subject is enrolled in the database. Also, many of the proposed CNN models are large, which renders them memory intensive and computationally costly to deploy on a mobile device. In this work, we propose CNN based robust subject-independent feature extraction for ocular biometric recognition, which is memory and computation efficient. We evaluated our proposed method on various ocular biometric datasets in the subject-independent, cross-dataset, and cross-illumination protocols.