Effects of Distance and Gaze Angle on CNN-based Standoff Iris Recognition

Abstract

Biometrics has been a popular topic for nearly 50 years now, with accuracy and use growing at an exponential rate. Iris recognition has become one of the most distinctive and accurate methods of biometric authentication, but it presents some challenges. Traditional systems are only trained with iris images that are taken in very conducive circumstances, which control the dilation, distance, and angle of the eye being captured. This however is not practical in the fast-moving world that is evolving around us, especially during a global pandemic, thus the creation of standoff recognition systems. Standoff recognition systems allow for the iris to be photographed in less restrictive conditions, but the question is do they yield recognitions that are just as accurate as the traditional systems. In this paper, we present a study that not only compares the results of systems that were trained with frontal images vs. off-angle images, but also the results of images with high numbers of pixels in iris texture vs. images with low numbers of pixels in iris texture (i.e., periocular vs. ocular). We investigate the effects gaze angle and distance to camera on recognition using different size and angles of images. Using convolutional neural networks, we train and test the proposed method with an off-angle iris dataset with 11,000 images.