Automated classification of mislabeled near-infrared left and right iris images using convolutional neural networks
Yixin Du, Thirimachos Bourlai, Jeremy Michael Dawson · 2016
In this paper, we propose a Convolutional Neural Network (CNN) with unified architecture (no need to re-design it for each unique iris database used) that operates well in a diverse set of iris databases. The CNN is designed to automatically recognize mislabeled left and right iris images by iris recognition system operators, and thus, extend the capabilities of a conventional iris recognition system. Our proposed approach is composed of three steps. First, for each iris database used as input, a CNN is trained using part of the database. Second, an empirical parameter optimization study is conducted so that classification performance is acceptable. Finally, the proposed classifier is tested on the remaining images of the same database used for training. The performance of the proposed network is evaluated on small- and large-scale iris databases, including the NIST's Iris Challenge Evaluation (ICE), LG ICAM 4000 iris, the CASIA Lamp, and the Pupil Light Reflex (PLR) databases. Experimental results show that independent of the databases used or whether the classification performance is tested on either a left-or right-side dataset, our approach results in a classification performance ranging from 97.5 to 100%.