高安全性人脸识别系统中的唇语识别算法研究  

Lip-reading algorithm in face recognition systems for high security

作　　者：任玉强[1] 田国栋[1] 周祥东[1] 吕江靖 周曦[1] 

机构地区：[1]中国科学院重庆绿色智能技术研究院智能多媒体技术研究中心,重庆400714

出　　处：《计算机应用研究》2017年第4期1221-1225,1230,共6页Application Research of Computers

基　　金：中国科学院战略性先导科技专项基金资助项目(XDA06040103);国家自然科学基金资助项目(61472386);国家自然科学基金青年基金资助项目(61502444)

摘　　要：针对目前人脸识别系统面临的图片和视频攻击,构建了一种将人脸识别与口令密码相结合,并采用唇语识别技术进行活体检测的高安全性身份认证系统。首先由于汉语唇语数据的缺少,建立了CNLIP1和CNLIP2两个较大的汉语唇语数据库;其次,为了保留唇语的时序性,采用堆叠卷积独立子空间分析(ISA)深度神经网络模型来实现唇动时序特征的提取;最后提出使用迁移学习算法来训练特定人唇语识别模型。实验证明,唇动时序特征能更好地表征出数字串唇语,迁移学习训练的特定人唇语模型能够满足活体检测的需要,所构建的高安全性人脸识别系统具有较好的防攻击效果。Aiming at defending against malicious attacks on face recognition systems using photos and videos, this paper pro- posed a high security face recognition system, which combined face recognition and lip-reading based liveness detection. Firstly, due to the lack of Chinese lip-reading data, it established two large Chinese lip-reading datasets named CNLIP1 and CNLIP2. Secondly, in order to extract temporal information from the data, it used an extensive version of the stacked convolutional independent subspace analysis algorithm to learn features from video data. Finally, it proposed to use transfer learning algorithm to train a speaker-dependent lip-reading model. Experiments show that the proposed feature gives preferable performance, the model can meet the needs of the liveness detection and the constructed high security face recognition system has good anti-attack effect.

关 键 词：高安全性 人脸识别 活体检测 唇语识别 迁移学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]