Perturbed Attention-Assisted Siamese Network for Cross-Spectral Periocular Recognition

Person identification in heterogeneous conditions finds application in surveillance and security systems, where probe and gallery images are associated with cross-spectral visual characteristics. Cross-spectral periocular recognition has proven to be more feasible than face and iris recognition systems when full-face information or high-resolution iris scans are unavailable. We propose a perturbed attention-assisted Siamese network to accurately match the periocular images captured in cross-spectral environments. We introduce two novel attention modules, namely, perturbed axialized spatial attention (PASA) and two-fold channelized spatial attention (2-FCSA), to generate finer discriminative feature maps. The PASA module considers the width and height axes independently for computing the partial attention over one axis at a time. The perturbation information is infused within the PASA module to drive the focus of our network more into the relevant foreground regions. The 2-FCSA module explicitly models the channel correlations of the perturbed feature maps using independent global context descriptors to generate the discriminative features. We created a new cross-spectral periocular dataset containing 12,584 visible and near-infrared periocular images from 100 subjects with pose and accessory variations. Extensive experiments and ablation studies on four datasets show the superior performance of the proposed network over state-of-the-art methodologies.