| 摘要: |
| 人脸活体检测是人脸识别技术安全落地的有力保障. 但活体攻击方式的不断变换, 给现有检测方法带来了极大的挑战. 为应对层出不穷的未知场景和攻击方式, 提出一种基于视觉注意力和域特征融合的双流人脸活体检测模型. 首先, 提出基于视觉注意力的特征提取模块, 增强模型提取基于全局信息的内容特征的能力. 接着, 构建一种新型的风格特征融合模块, 将内容特征和浅层纹理表达的风格特征相融合来优化样本的特征表示. 此外, 设计基于孪生网络的特征映射策略并修正对比损失函数, 分别强化模型的鲁棒性和规避训练过程中梯度易振荡的问题. 还采用对抗训练来降低模型对样本数据域之间分歧的敏感性, 进一步增强其泛化性. 多项实验结果表明, 所提方法在主流数据集上跨域表现均优于现有模型, 验证其泛化性和强鲁棒性. |
| 关键词: 人脸活体检测 域泛化 特征融合 |
| DOI：10.13328/j.cnki.jos.007285 |
| 分类号:TP391 |
| 基金项目:国家重点研发计划(2023YFF1105102, 2023YFF1105105); 国家社科基金重大项目(21&ZD166); 江苏省自然科学基金(BK20221535); 江苏省研究生研究与实践创新计划(KYCX23_2438) |
|
| Face Anti-spoofing Method Based on Visual Attention and Domain Feature Fusion |
|
ZHU Jian-Qiu, HUA Yang, SONG Xiao-Ning
|
|
School of Artificial Intelligence and Computer Science, Jiangnan University, Wuxi 214122, China
|
| Abstract: |
| Face anti-spoofing is a powerful guarantee for the practical security of facial recognition technology. However, the constant evolution of live attack methods poses significant challenges to existing detection methods. To address the increasing number of unknown scenarios and attack methods, a two-stream face anti-spoofing model based on visual attention and domain feature fusion is proposed. First, a visual attention-based feature extraction module is proposed to strengthen the model’s capacity to extract content features based on global information. Second, a novel style feature fusion module is designed to optimize the feature representation of the sample by fusing content features with low-level textural style features. Third, a feature mapping strategy based on the Siamese network is developed and the contrast loss function is modified to improve the model robustness and avoid easy gradient oscillation during training, respectively. Furthermore, domain adversarial training (DAT) is used to reduce the sensitivity of the model to differences between sample data domains and further improve its generalization. Extensive experimental results verify the generality and strong robustness of the proposed method, demonstrating that it outperforms existing models in cross-domain performance on mainstream datasets. |
| Key words: face anti-spoofing domain generalization feature fusion |
