A Deep Learning based Approach for 3D Mask Face Anti-Spoofing Enhancement
Qiushi Guo, Shisha Liao, Yifan Chen, Shihua Xiao, Jin Ma · 2023
Face recognition system is widely used in recent years, but it is vulnerable to attacks via print paper or 3D masks. By diving into the attack process, we noticed that detecting face occlusion of small stick-like objects benefits the face antispoofing. In this paper, we propose a segmentation-based deep learning method as an enhancement to protect the system being attacked by high-fidelity 3D masks. However, manually annotation is labor intensive and easily involves extra errors, We generated a novel high quality dataset called TinyOcc combining CelebA-HQ and only 50 self-annotated stick-like objects images. Several popular segmentation models are fine-tuned using above novel datasets, achieving the mIoU 0.92(intra),0.87(cross) and precision 0.94, which proved the robustness and effectiveness of our approach.