基于3D卷积神经网络的活体人脸检测研究

发布时间：2018-07-20 16:26

【摘要】：非法认证者可通过伪装人脸获得进入人脸识别系统的权限,为社会安全带来威胁。因此,活体人脸检测具有现实紧迫性。然而,现有文献多为照片人脸攻击方面的研究,对于视频人脸攻击,识别率不甚理想。3D卷积神经网络(Convolutional Neural Network,CNN)是一个深层结构,能自动学到图像的分布式特征表示;与2D卷积相比,能学到连续视频帧的动作信息。本文结合3D卷积神经网络的特性,提出利用3D卷积神经网络实现视频人脸伪装检测。本文的研究内容主要包括:(1)对现有活体人脸检测算法做了深入研究,分析了2DCNN的弊端,只能对二维图像进行卷积;现有活体人脸检测算法多为手工提取特征,存在的缺点为:(1)手工设计特征只对某种类型的图片有较好的识别率;(2)手工设计特征需要较高的专业知识。本文对视频做卷积,首次提出将3D卷积神经网络应用于活体人脸检测。(2)针对两个公开的人脸伪装数据库,设计合适的3D卷积神经网络结构,包括网络的层数、卷积核大小及数量。此外,根据每个网络在时间维度中下采样不同的先后顺序设计了三种待选网络(Late,Slow,Early)。通过实验测试,比较每个网络识别率高低与网络参数的数量来选择合适的网络。(3)利用前面找到的最优网络,将REPLAY-ATTACK和CASIA-FASD人脸伪装数据库作为实验对象,对不同帧数输入进行实验对比,寻找使分类器性能达到最佳的最优输入帧数。通过提取网络最后全连接层学到的时间空间特征,训练支持向量机(Support Vector Machine,SVM)分类器,实现真实人脸和伪装人脸的分类。(4)将最优网络和最优帧数作为输入,实现两个公开人脸伪装数据库REPLAY-ATTACK和CASIA-FASD的多尺度内部数据库测试和交叉数据库测试。实验分为5个尺度完成测试,在网络输入层,图像分为5帧输入;当利用3D卷积神经网络学习到图像帧的特征后,本文提取网络最后一层全连接层的特征;最后利用提取到的训练集特征训练SVM分类器,从而实现真实人脸和伪装人脸的分类。实验结果相对于纹理特征及2D卷积方法有较大提高,可应用于视频人脸攻击的活体人脸检测。
[Abstract]:Illegal authenticators can gain access to face recognition system by camouflage face, which is a threat to social security. Therefore, in vivo face detection has a realistic urgency. However, most of the existing literatures focus on the face attack of photographs. For video face attacks, the recognition rate is not ideal. The Convolutional Neural Network (CNN) is a deep structure, which can automatically learn the distributed feature representation of images. Compared with 2D convolution, we can learn the action information of continuous video frames. Based on the characteristics of 3D convolution neural network, this paper proposes a 3D convolutional neural network for face camouflage detection. The main research contents of this paper are as follows: (1) the existing human face detection algorithms are deeply studied, and the disadvantages of 2DCNN are analyzed, only 2D images can be convoluted; most of the existing live face detection algorithms extract features by hand. The disadvantages are: (1) manual design features only have a better recognition rate for certain types of images; (2) manual design features require higher professional knowledge. In this paper, we firstly apply 3D convolution neural network to face detection in vivo. (2) for two open facial camouflage databases, we design a suitable 3D convolution neural network structure, including the number of layers of the network. Size and number of convolution nuclei. In addition, according to the order in which each network samples in different order in time dimension, three kinds of waiting network (low order early) are designed. Through the experimental test, the recognition rate of each network and the number of network parameters are compared to select the appropriate network. (3) the REPLAY-ATTACK and CASIA-FASD face camouflage database is used as the experimental object. In order to find the optimal input frame number for the classifier, the different frame number input is compared with each other. By extracting the temporal and spatial features learned from the last full connection layer of the network, the support Vector Machine (SVM) classifier is trained to classify real and camouflaged faces. (4) the optimal network and the optimal number of frames are used as inputs. Two open facial camouflage databases, REPLAY-ATTACK and CASIA-FASD, are implemented for multi-scale internal database testing and cross-database testing. The experiment is divided into five scales to complete the test, in the network input layer, the image is divided into five frames input, when using 3D convolution neural network to learn the features of the image frame, this paper extracts the features of the last layer of the network full connection layer. Finally, SVM classifier is trained with the extracted features of training set to realize the classification of real face and camouflage face. Compared with the texture features and 2D convolution, the experimental results can be applied to live face detection of video face attacks.
【学位授予单位】：五邑大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【参考文献】