Bandpass Filter Based Dual-stream Network for Face Anti-spoofing

Abstract

Face Attack Detection (PAD) technology is crucial for protecting facial recognition systems. At present, methods for Face Anti-spoofing (FAS) mainly focus on short-distance applications, and algorithm performance can sharply decline when facing challenges such as low resolution, pedestrian obstruction, and blurriness in long-distance scenarios. To address these issues, we propose a dual-stream architecture that combines information from the images and its bandpass filtered image to distinguish attacks. Specifically, one branch extracts detailed facial structure and texture information from the original spatial domain of images. The other branch take the Gaussian bandpass filtered image as input to learn the complementary discriminative features. The filtering process was done in frequency domain by FFT/IFFT. We proposed a cross-attention fusion module to fuse the features extracted by the two network branches. Additionally, to further improve the model's generalization ability to data quality, we use automatic correction and lion optimizer. Finally, our method achieved a result of 6.22% on the ACER metric and ranked third in the 4th Face Anti-Spoofing Challenge @CVPR2023.