Facial StO2-based personal identification: dataset construction, feasibility study, and recognition framework

Zheyuan Zhang, Xinyu Liu, Yingjuan Jia, Ju Zhou, Hanpu Wang, Jiaxiu Wang, Tong Chen

Abstract: Biometrics have been extensively utilized in the realm of identity recognition. However, each biometric method has its inherentlimitations in specific scenarios. For example, identity recognition based on facial images is contactless but can be forged;finger vein recognition is very secure but generally requires contact collection to ensure accurate identification. In somescenarios with high security requirements, there is often a need for contactless acquisition of biometric features that cannot beforged to recognize identity. Therefore, a novel biometric, facial tissue oxygen saturation (StO2) with the advantages of robustanti-spoofing capabilities and non-contact measurement, is proposed for identity recognition. To more comprehensively verifythe feasibility of facial StO2 for identity recognition, a Facial StO2 Identity Dataset (FSID148) containing 148 identities iscollected and the feasibility of facial StO2 identity recognition is validated by performing verification, close-set identification,and open-set identification tasks. In order to enhance the performance of facial StO2 identity recognition, an attention-guidedcontrastive learning framework that enables backbones to derive discriminative identity representations from both local andglobal facial StO2 regions is proposed. The method proposed has achieved accuracies of 96.11%, 94.60%, and 88.51% in theaforementioned tasks, positioning facial StO2 as a promising biometric for a wide array of application scenarios.