Snapshot Spectral Imaging for Face Anti-Spoofing: Addressing Data Challenges with Advanced Processing and Training

Hui Li, Yaowen Xu, Zhaofan Zou, Zhixiang He; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 1005-1012

Although considerable research progress has been made in the field of face anti-spoofing(FAS) it still faces continuous threats from ultra-realistic face mask attacks. Facing the challenge of highly realistic flexible masks spectral sensors show great potential in enhancing the safety of FAS systems. However the application of snapshot spectral imaging (SSI) in FAS is still in its infancy and faces two major challenges: data scarcity and data content differences. To this end we introduce a data processing and model training method for SSI images. In terms of data processing we use RandomBorderMask technology and RandomDropChannels strategy to avoid misjudgment of material information reduce interference from redundant information and learn from RGB image preprocessing methods to enhance data diversity. Regarding model training we propose a model integration strategy and semi-supervised learning technology which combines the prediction results of multiple models and uses pseudo labels to expand the training data to solve the over-fitting problem caused by data scarcity. These innovative methods achieved first-class results in the 5th Face Anti-Spoofing Challenge @CVPR2024 verifying their effectiveness in improving the accuracy and robustness of the FAS system.