High Boost 3-D Attention Network for Cross-Spectral Periocular Recognition

Abstract

Recognition of individuals in a heterogeneous surveillance environment is a challenging problem as it involves matching images associated with different visual characteristics. These differences mainly arise due to the involvement of various wavelength ranges and sensing equipments. This letter presents a novel high boost 3-D attention mechanism embedded within a deep Siamese framework to address periocular recognition in heterogeneous wavelength domains. The proposed high boost attention module (HBAM) generates 3-D attended features by calculating the importance of each feature location. The HBAM works by enhancing significant components of the input feature map through a boost coefficient without any additional parameter overhead. We have created a new cross-spectral periocular dataset comprising 12 584 visible and near-infrared images from 200 classes by considering random pose (eye and head movement) and accessory (mask and eyeglass) variations. Experiments and ablation studies on five cross-spectral periocular datasets demonstrate efficacy of the proposed 3-D attention network and significance of its design choice. � 2017 IEEE.