结合对比学习与空间上下文的人脸活体检测

Face anti-spoofing based on spatial context-aware contrastive learning

人脸活体检测是面部识别应用的先决条件。现有方法利用多种特征提升检测精度,针对特征优化的研究较少,一些能起有效区分作用的特征未被提取,提出结合空间上下文特征与对比学习的人脸活体检测(SAC),分两个阶段:①对比学习得到高级语义特征,扩大活脸和攻击脸之间的特征差距。此外,提出两种定义样本的策略以增强特征的可辨性。②上下文判别器:自我注意学习输入的极端贡献性空间上下文。两阶段用跳跃连接保持特征表示一致性,分别用特征相似度和交叉熵损失训练。在公开数据集SiW、CAISA FASD和Replay Attack上测试,SAC在SiW的3种测试协议下均取得了与先进算法可比较的结果,在CAISA FASD和Replay Attack的跨数据集测试结果分别提升4%、11%,表明SAC能够有效准确判别欺骗人脸。

Face anti-spoofing is a prerequisite for facial recognition applications.Existing methods prefer to utilize multiple clues to improve liveness detection accuracy,and few of them focus on refining feature representation.We propose a Spatial Context-Aware Contrastive Learning(SAC)for Face Anti-Spoofing.SAC is divided into two phases:①Representation learning:enhance features representation by high-level semantic features,which enlarges the feature distance between live and attack faces.And we further propose two strategies for sampling.②Spatial context-aware discrimination:learning the critical spatial context based on self-attention.And skip-connection is equipped to maintain the consistency of feature representation between two phases.The two phases are trained by feature similarity and cross-entropy loss respectively.Experiments are conducted on datasets SiW,CASIA FASD,and Replay Attack,and achieve comparable results compared with state-of-arts methods under three protocols on SiW.On the cross-dataset testing with CASIA FASD and Replay Attack,SAC improved 4%,11%respectively.These demonstrate the effectiveness of the proposed model.