基于多视角唇动时空动态特征的身份识别研究

Biometric technologies have rapid development in national defense, military security and civil identity. Speaker identification based on lip motion also gain lots of attention, because of its less restrictive factors, relatively easy, and meeting the demand of real-time identification. Nowadays Speaker identification based on lip motion has two big problems. It is sensitive to the head’s rotation. If speaker rotate his head, the recognition rate will be reduced sharply. The other is stringent request of the video equipment. Low resolution also makes the recognition result be worse. To solve these problems, the key research points of our project are as follow: 1) AS the rotation problem of the speaker, how extract the lip motion features based on. 2) Lip super resolution features are extracted using super resolution algorithm from low resolution image features. 3) The methods of multi-view super resolution feature fusion.. This topic has a very specific application prospect: speaker recognition. It has the big challenge in theory, and the research content also gets people's attention more and more. Over these years, the applicant of this project has a good cooperation with Professor Alexander I.Rudnicky who works in the school of computer science in Carnegie Mellon University. We focus on the research of audiovisual information processing and pattern recognition and so on. In all, the applicant of this project has good research foundation for this project. The results of this project will make the identification system be more robust and more humanity, and accelerate the products application of the system..

生物识别技术在国防和民用身份认证中都有较快的发展,特别是基于动态唇形的身份识别, 因其受制约因素少、实现相对容易、可满足实时识别的需求等特点而受到广泛重视。目前唇动身份识别系统存对使用者头部转动敏感和视频设备的要求苛刻等问题，随着使用者头部轻微转动，或图像分辨率降低使得识别率急剧下降。针对这些问题，本项目重点研究：1）针对视频中说话人头部转动问题，提取基于多视角的唇动身份特征。2)利用超分算法处理低分辨率图像，从而得到有效的超分特征。3）基于时空的多视角超分特征融合方法。. 本课题有非常明确的应用前景：身份认证。理论上也具有很大的挑战性。项目申请者多年来与卡耐基梅隆大学计算机学院的Alexander I.Rudnicky教授有很好的研究合作关系，一直合作致力于音视频信息处理和模式识别的研究，有良好的研究基础。项目成果使身份识别系统更鲁棒，更人性化，加速其产品化。

生物识别技术在国防和民用身份认证中都有较快的发展,特别是基于动态唇形的身份识别,因其受制约因素少、实现相对容易、可满足实时识别的需求等特点而受到广泛重视。本项目主要研究内容有：1）针对视频中说话人头部转动问题，提取基于多视角的唇动身份特征。2)利用超分算法处理低分辨率图像，从而得到有效的超分特征。3）基于时空的多视角超分特征融合方法。该项目按照计划执行后，主要有以下研究结果：1）建立和扩展了基于人脸和唇部的数据库，并对其数据库进行了大量处理，大大地节约了同行使用该数据库的工作量。2）提出一种基于权重值的视角特征融合方法来进行身份识别，而实验表明比均分的视角特征融合的识别率提高了5%。3）利用对图像特征来进行超分特征提取，就可以不用考虑图像本身的分辨率了。这也给了对处理不同图像数据库的研究有很好地借鉴。4）动态特征能很好地提高唇读的识别效果，实验表明利用该动态特征提高了唇动说话人身份识别率。这种方法对基于视频的人脸识别系统也是有很大地帮助。 总体该项目的研究可以提高目前流行的人脸认证系统的安全性和人脸识别率，以及对在高噪声背景下的说话人识别系统有很好地辅助作用。在目前的身份认证方面有很大地应用前景。