Twin Deep Convolutional Neural Network-based Cross-spectral Periocular Recognition

Abstract

Recognition of individuals using periocular information has received significant importance due to its advantages over other biometric traits such as face and iris in challenging scenarios where it is difficult to acquire either full facial region or iris images. Recent surveillance applications give rise to a challenging research problem where individuals are recognized in cross-spectral environments in which a probe infra-red (IR) image is matched with a gallery of visible (VIS) images and vice versa. Cross-spectral recognition has been studied mostly for face and iris traits over the past few years; however, the performance of periocular biometric in the cross-spectral domain still needs to be improved. In this paper, we propose a twin deep convolutional neural network (TCNN) with shared parameters to match VIS periocular images with those of near IR (NIR) ones. The proposed TCNN finds the similarity between the VIS and NIR image pairs applied at its input rather than classifying them into a certain class. The learning mechanism involved in this network is such that the distance between the images corresponding to the genuine pairs is reduced and that of the imposter pairs is maximized. Based on the experimental results and analysis on three publicly available cross-spectral periocular databases, the TCNN achieves the state-of-the-art recognition results.