Deep Attention Networks for Periocular Recognition in Cross Spectral Environments

Abstract

nificant newlineattention due to its advantages over face and iris traits in challenging newlinescenarios where it is difficult to acquire either full face images or high-resolution newlineiris scans. Recent advancements in surveillance applications require deployment newlineof infra-red sensing equipments to capture activities occurring in low-illumination newlineconditions. However, most of the existing recognition systems are enrolled using newlineimages captured in constrained imaging setup in presence of visible light. This newlinegives rise to the problem of cross-spectral recognition where probe and gallery images newlineare captured in distinct wavelength ranges. Specifically, in case of periocular newlineimages, alleviating the wide appearance gap and learning to extract illuminationinvariant newlinefeatures become more challenging. This thesis introduces several deep newlinelearning-based frameworks to address periocular recognition in cross-spectral environments. newlineA new dataset, namely, IITBBS cross-spectral periocular dataset, newlinecontaining 12,584 visible and near-infrared periocular images is created where the newlineimages are captured in unconstrained acquisition setup involving unsupervised newlinepose and accessory variations. The new dataset is made publicly available and newlinededicated to research community for academic and research purpose. newlineWe propose a twin deep convolutional neural network (TCNN) with shared parameters newlineto match periocular images captured in cross-spectral scenario. It finds newlinesemantic similarity between heterogeneous image pairs applied at its input rather newlinethan classifying them into a certain class. During training, the distance between newlineimages corresponding to genuine pairs is reduced and that of imposter pairs is newlinemaximized. Then, we introduce a dual-spectrum network to explore the relationship newlinebetween deep features and hand-crafted visual attributes where deep features newlineare mapped into the attribute space using the mapping network.