Face anti-spoofing via jointly modeling local texture and constructed depth

To solve the security problem posed by fake face presentation attacks in face recognition systems, many detection methods have been proposed in recent years. Among these detection methods, the methods by analyzing a single image have achieved good real-time performance. However, their detection performances are often limited by the effectiveness of the extracted features. Compared with the real face, the fake face always shows structure and texture differences. Based on this, we propose a detection method that fuses local texture and structural depth information, where the Swin-Transformer is invoked to generate both local and constructed features. More specifically, the face image is first divided into non-overlapped small patches, which can further extract the features in micro-clues. Then, the Swin-Transformer with four stages is invoked to encode these patches, and the outputs from different stages are fused to generate constructed depth descriptors and local texture features. Finally, the local texture features and constructed depth map are combined to classify whether the input face image is captured from a real face. Extensive experiments about intra-dataset and cross-dataset evaluation indicate that our proposed approach significantly outperforms the state-of-the-art methods and show promising generalization performance.