Adaptive cross-resolution representation learning for generalizable face anti-spoofing

Since diverse Presentation Attacks arise under unseen scenarios, face anti-spoofing (FAS) based on domain generalization has drawn growing attention due to its robustness. Most existing FAS methods rely on a chain of CNN to extract features, where the receptive field is fixed in each layer. Such methods may be not enough to handle the spoof cues with various scales in FAS. We propose a novel perspective that introduces cross-resolution presentation learning for generalizable FAS. Specifically, we propose a novel framework, Adaptive Cross-resolution Representation Learning, combined with two specially designed modules, to capture spoof cues with various scales and adaptively select the features for each sample to fuse the generalizableY9+wRrNd72R4qRcxANAOJcprepresentations. Concretely, Cross Resolution Interaction Module is introduced to capture multi-resolution features and encourage the interactions among these features for more comprehensive representations. Moreover, Adaptive Representation Selecting Module is proposed to construct generalizable representations for each sample based on a meta-learning strategy. Extensive experiments and visualizations are presented to demonstrate the effectiveness and interpretability of our method against the state-of-the-art competitors.