Cross-Example Patch Fusion for Face Anti-Spoofing

In this paper, we proposed a novel pipeline for face anti-spoofing (FAS) tasks with re-structure face images based on the cross-example patch fusion (CEPF) mechanism. The effective end-to-end pipeline mainly includes three modules: a face patches mixer, a spoof cue encoder, and a semantic map predictor. Firstly, we generate re-structure face images from the mixer by using multi-scale patches of one sample selected randomly from a mini-batch dataset to overlap corresponding spatial patches of the input example. Then, we capture spoofing cues via an encoder, a designed Convolutional neural network. Last, we achieve the result of anti-spoofing with a semantic map generated from the predictor. Our approach outperforms the existing methods in performance and speed under the same scale architecture and reveals valuable facts on learning spoofing cues. Using random multi-scale CEPF face images can extract the intrinsic spoofing cues more effectively than the original spatial information of none CEPF faces by a simple encoder, which provides inspirable insights for future supervision design.