毕业生知识库

近年来，人脸识别技术发展迅速，在公共安全、门禁考勤等众多方面得到了广泛的应用。然而，现有的人脸识别系统缺乏检测仿冒攻击（也称为呈现攻击）的有效手段，其安全性已成为制约人脸识别系统应用的一个重大瓶颈。人脸活体检测技术不论是在学术界还是在工业界，均得到了广泛关注。

本论文总结了国内外人脸活体检测领域已有的成果，并对当前存在的问题进行了深入的分析。除了常见的可见光人脸数据外，本论文还对深度图像和近红外图像对人脸活体检测的贡献进行了探讨。本论文的主要贡献有：

（1）提出一种基于人脸深度图的活体检测方法。由于真实人脸和人脸照片、视频等具有不同的立体结构，使用深度信息能够有效区分真人和二维假体，本文分别采用基于支持向量机和卷积神经网络的方法对人脸深度图进行活体检测，取得了较好的效果。

（2）采集了一个可见光近红外双模态人脸面具防伪数据集，并提出三种多模态融合算法作为基准算法。现有的大多数人脸活体检测数据集和方法集中在人脸照片和视频等二维攻击，少有三维面具人脸方伪数据集。本文采集的数据集包括可见光和近红外两种模态，包含67个真人和48个3D面具，同时提供了基准测试协议和评测结果，初步探讨了双模态融合对人脸防伪的作用，有助于推动多模态的3D人脸面具防伪技术的发展。

（3）提出一种3D辅助的虚拟假体样本生成方法，扩充人脸活体检测训练集，提高分类器检测精度和推广性能。人脸活体检测假体样本数据采集费时费力，本方法引入弯折和旋转，可以逼真生成任意数量负样本，缓解负样本数据少给人脸活体检测算法带来的困难。

综上所述，本论文主要研究基于不同模态的人脸活体检测技术，并提出一种虚拟假体样本生成方法以防止数据量少带来的过拟合问题。

Recently, face recognition techniques develop rapidly and have been deployed in many real applications. However, various presentation attacks threaten existing face recognition systems. The security has become a bottleneck restricting the applications of face recognition systems. Nowadays, the threat from the spoofing attacks, such as face images, videos or 3D masks of legitimate users, has been realized and the face presentation attack detection (PAD) has attracted a lot of attention.

This thesis summarizes previous methods of face presentation attack detection and provides a systematic and deep analysis for the issue of face presentation attack detection. To make face presentation attack detection systems more effective, the author makes use of near-infrared (NIR) images and depth images to detect spoofing faces. The main contributions of this work are listed as follows:

(1) This thesis proposes a face presentation attack detection method based on depth images. Since genuine face and printed photo have different 3D structures, the use of 3D information is straightforward and significantly benefits face spoofing detection systems. The SVM-based method and the convolutional neural network-based method are proposed to utilize depth images for face spoofing detection and achieve good performance.

(2) This thesis collects a multi-modality 3D mask face anti-spoofing database and presents three multi-modal fusion methods as baseline algorithms to effectively merge the involved 2 modalities. Most of the existing databases focus on the 2D attacks, including photo and video attacks. The only two public 3D mask face anti-spoofing databases are relatively small. The collected database contains 920 videos of 67 genuine subjects wearing 48 kinds of 3D masks, captured in visual (VIS) and near-infrared (NIR) modalities. Furthermore, the author also builds three protocols and examine the performance of multi-modal fusion methods for face spoofing detection. The author hopes this database would help to promote the development of 3D face mask presentation attack detection techniques.

(3) This thesis proposes a virtual image synthesis method, which is able to generate bent and out-of-plane spoof samples so that large scale spoof data can be generated for training deep neural networks to boost the anti-spoofing performance. Since acquiring spoof data is very expensive because of the live faces should be re-printed and re-captured in many views, the proposed method could synthesize virtual spoof data in 3D space to alleviate this problem.

In summary, this thesis studies face liveness detection techniques based on different modalities. The face spoofing detection methods based on depth image, VIS and NIR modalities are proposed and evaluated. A virtual face synthesis method is also proposed to alleviate the over-fitting problem for deep learning based methods.