Dynamic Residual Distillation Network for Face Anti-Spoofing With Feature Attention Learning
Dynamic Residual Distillation Network for Face Anti-Spoofing With Feature Attention Learning
复制标题
DOI：
10.1109/tbiom.2023.3312128
复制
发表时间：
2023-10
期刊：
影响因子：
--
通讯作者：
Yan He;Fei Peng;Min Long
中科院分区：
文献类型：
--
作者：
Yan He;Fei Peng;Min Long
目前，大多数Face反欺骗方法依赖于辅助信息(如附加注释和通道)来解决泛化问题。然而，这些辅助信息在实际场景中是不可用的，这可能会阻碍这些方法的应用。同时，预先确定或固定的特性限制了它们的泛化能力。为了解决这些问题，提出了一种带特征注意学习的动态残差精馏网络，无需获取任何辅助信息即可自适应地搜索判别表示和嵌入空间。具体地说，首先设计了像素级残差蒸馏模块，通过同时抑制高层语义和低频光照因素来获得与域无关的活跃度表示，从而自适应地缓解源域和目标域之间的域差异。其次，提出了一种特征级别的注意力对比学习方法，构造了一个距离感知的非对称嵌入空间，避免了类边界的过度拟合。最后，设计了一个结合注意块的注意力增强主干，用于在特征提取中自动捕获重要区域和通道。实验结果和分析表明，该方法在单源和多源域泛化场景中的性能均优于现有的反欺骗方法。
Currently, most face anti-spoofing methods target the generalization problem by relying on auxiliary information such as additional annotations and modalities. However, this auxiliary information is unavailable in practical scenarios, which potentially hinders the application of these methods. Meanwhile, the predetermined or fixed characteristics limit their generalization capability. To countermeasure these problems, a dynamic residual distillation network with feature attention learning (DRDN) is developed to adaptively search discriminative representation and embedding space without accessing any auxiliary information. Specifically, a pixel-level residual distillation module is first designed to obtain domain-irrelevant liveness representation by suppressing both the high-level semantic and low-frequency illumination factors, thus the domain divergence between the source and target domains can be adaptively mitigated. Secondly, a feature-level attention contrastive learning is proposed to construct a distance-aware asymmetrical embedding space to avoid the class boundary over-fitting. Finally, an attention enhancement backbone incorporated with attention blocks is designed for automatically capturing important regions and channels in feature extraction. Experimental results and analysis demonstrate that the proposed method outperforms the state-of-the-art anti-spoofing methods in both single-source and multi-source domain generalization scenarios.