【作者】 李嘉豪；

【导师】 陈虎；

【作者基本信息】 四川大学 ， 工程硕士（专业学位）， 2021， 硕士

【摘要】 三维人脸识别技术是一种应用广泛的身份认证技术,在安防、网络支付等领域都有着成熟的应用,其中,基于深度学习的三维人脸识别技术表现突出。一种好的三维人脸识别方法应该对人脸表情的变化有足够的鲁棒性,这就需要在训练过程中有大量的多种表情下的三维人脸数据。然而,由于难以获得三维人脸数据以及一些三维人脸数据缺乏准确性,三维人脸识别技术的发展遇到了比较大的阻碍。通过对已有的三维人脸进行变形来对三维人脸数据进行扩充,成为了解决三维人脸数据缺少的问题的一种途径。本文介绍了当前主流的三维人脸变形方法,基于前人的研究,提出了一种基于解耦表示学习的三维人脸变形方法,并基于该方法设计了一个三维人脸采集系统,为解决三维人脸数据不足的问题提供了有效的途径。本文的主要工作如下:第一,本文针对中性表情人脸样本不足的问题,提出了一种基于区域重组的三维人脸样本扩充方法和基于插值法的三维人脸样本扩充方法,该方法结合了两种数据扩充方法的优点,既能使生成的人脸具有足够的身份特征多样性,在人脸的细节处的表现也更好。这种方法能够生成多数量、高质量的三维人脸样本,扩充中性表情三维人脸,为研究三维人脸变形提供了大量的中性三维人脸样本。第二,本文对基于VAE网络的三维人脸解耦表示算法进行了研究和改进。为了得到更好的三维人脸解耦表示效果,本文对网络的输入方式和卷积方式上做了改进,并将算法运用到三维人脸变形当中。该方法能够将一个三维人脸解耦成为身份和表情两大特征,对表情特征进行改变后,再和身份特征进行融合,可以得到变形后的三维人脸。这种方法能够在尽可能保持身份特征不变的情况下对三维人脸进行变形,生成逼真的变形后的三维人脸,为三维人脸识别研究提供大量且有效的数据基础。本文在公开数据库上,设计了与另外三个变形模型的对比实验,在三维人脸变形的几个评估指标的表现上都更优,证实了本文提出方法的可行性。第三,本文在三维人脸变形算法的基础之上,设计了一个基于三维人脸变形方法的三维人脸采集系统。该系统使用三维相机进行拍摄并建模,然后对三维人脸进行变形,达到采集一次人脸,存储多个不同表情的三维人脸模型的效果,能够从有限的采集到的三维人脸数据集中大幅扩充三维人脸数据。本文还在系统实现过程中开发了一种基于事件响应机制的高并发服务器框架,该框架在多个指标上都优于当前的主流高并发服务器框架,在性能上表现优异。更多还原

【Abstract】 3D face recognition technology is a widely used identity authentication technology,which has mature applications in security,online payment and other fields,and deep learning has a good performance in 3D face recognition.A good 3D face recognition method should have good robustness to changes in facial expressions,which requires a large amount of 3D face data under various expressions during the training process.However,due to the difficulty in obtaining 3D face data and the lack of accuracy of some 3D face data,the development of 3D face recognition technology is facing huge difficulties.Performing 3D face deformation from the existing 3D face data set to expand other 3D faces is a effective way to solve the problem of lack of 3D face data.This paper introduces the current main 3D face deformation methods,proposes a 3D face deformation method based on disentangled representation learning,and designs a 3D face acquisition system based on this method which provides a effective solution to solve the problem of lacking 3D face data.The main work of this paper is as follows:Firstly,in order to solve the problem of insufficient face samples for neutral expressions,this paper proposes a 3D face sample expansion method based on region reorganization and face interpolation which combines the advantages of two data expansion methods.This method can generate a large number of high quality 3D face samples with natural expression to expand the neutral expression 3D face data.It provides a large number of neutral 3D face samples for the study of 3D face deformation.Secondly,this paper studies and improves the 3D face disentangled representation algorithm based on VAE network.In order to get a better disentangled representation of 3D face,this paper improves the input mode and convolution mode of the network.In addition,this paper applies the algorithm to 3D face deformation.It can decouple a 3D face into identity and expression characteristics,and could obtain deformed 3D face from changing the expression characteristics and fusing with the identity characteristics.The method provides a way to deform the 3D face while keeping the identity features as unchanged as possible,and generate a realistic deformed 3D face,which provides a large and effective data basis for the research of 3D face recognition.This paper designs comparative experiments with three other deformation models on the public database.From the experimental results,the performance of several evaluation indicators of 3D face deformation is better,which confirms the feasibility of the method proposed in this paper.At last,this paper designs a 3D face acquisition system based on the 3D face deformation method which uses the 3D face deformation framework.The system uses a 3D camera to shoot human faces and build 3D models,and then deform and save 3D faces.It can store multiple 3D face models with different expressions from only once face collecting.The system could achieve the goal to greatly expand the 3D face data from the limited collection of 3D face data.This paper also developed a high-concurrency server framework based on the event response mechanism to provide sufficient and stable network service to the 3D face collection system.This framework is superior to the current mainstream high-concurrency server framework in many aspects,and has excellent performance.更多还原