人脸识别系统中活体检测算法研究

论文总字数：34068字

摘 要

贺港龙，指导教师：罗琳

随着各种生物认证技术迅速发展，人脸识别技术在各方面的应用也越来越广泛。与此同时，人脸识别技术的安全性和可靠性也就显得更加的重要。目前来看，人脸识别系统仍然容易受到非法用户的攻击。因此人脸防欺骗技术就成为了非常重要的研究课题。

本文对各种活体检测算法进行研究，详细介绍了各种传统的手工特征提取算法以及近年提出的基于深度学习的活体检测算法。首先详细分析传统方法中的LBP算法、VLBP算法、LBP-TOP算法和 DoG算法的原理，以及这些方法在活体检测中的应用。然后介绍了卷积神经网络CNN以及轻型网络架构FeatherNets。

本文分别利用现有的活体检测数据集NUAA及多模态活体检测数据集CASIA-SURF对传统方法和基于深度学习方法进行实验测试。实验结果表明，对于传统手工纹理特征提取算法，单一的某种算法提取的纹理特征一般会有片面性和主观性，而将不同的纹理提取方法结合可以达到更好的效果。对于卷积神经网络的方法，相较于其他的网络模型，本文使用的带有流模块的精简网络架构FeatherNets在实现多模态的活体检测任务中，实现了识别性能和计算的复杂性之间很好的折中，是一种非常有效且应用性广的活体检测网络架构。

关键词：活体检测，LBP，CNN，FeatherNets

ABSTRACT

Ganglong He, Instructor: Lin Luo

Face recognition technology is more and more important these years with the rapid development of biometric authentication technologies. Simultaneously, the security and reliability of face that is more and more important. However, the face recognition system is still to be attacked easily. To sum up, face anti-spoofing is a important topic.

This paper focuses on the typical living body detection algorithm in face recognition system, and introduces various traditional manual feature extraction algorithms and the deep learning-based living body detection algorithm proposed in recent years. Firstly, the principles of LBP algorithm, VLBP algorithm, LBP-TOP algorithm and DoG algorithm in traditional methods are analyzed in detail, and the application of these methods in living body detection. Then this paper introduces CNN and FeatherNets, a new network architecture.

In this paper, the traditional method and the multi-modal in vivo detection data set CASIA-SURF are used to test the traditional method and the deep learning method based on the existing in vivo detection data set NUAA. The experimental results show that for the traditional manual texture feature extraction algorithm, the texture features extracted by a single algorithm generally have one-sidedness and subjectivity, and different texture extraction methods can achieve better results. For the method of convolutional neural network, FeatherNets, a streamlined network architecture with stream modules, is used in the multi-modal in vivo detection task to achieve the recognition performance and computational complexity compared to other network models. A good compromise is a very effective and widely applicable live detection network architecture.

Keywords: liveness detection, LBP, CNN, FeatherNets

目 录

摘要…………………………………………………………………………………………………Ⅰ

Abstract……………………………………………………………………………………………Ⅱ

1. 绪论…………………………………………………………………………………………1

1.1 引言………………………………………………………………………………………1

1.2 活体检测研究现状………………………………………………………………………1

1.3 论文主要研究内容及意义………………………………………………………………2

1.4 论文组织结构……………………………………………………………………………3

1. 传统活体检测算法…………………………………………………………………………4

2.1 LBP算法…………………………………………………………………………………4

2.1.1 普通LBP………………………………………………………………………4

2.1.2 VLBP……………………………………………………………………………5

2.1.3 旋转不变VLBP…………………………………………………………………8

2.1.4 LBP-TOP…………………………………………………………………………9

2.2 DoG算法………………………………………………………………………………12

2.3 本章小结……………………………………………………………………………13

1. 基于CNN的活体检测算法………………………………………………………………14

3.1 卷积神经网络CNN……………………………………………………………………14

3.1.1 卷积层…………………………………………………………………………14

3.1.2 池化层…………………………………………………………………………15

3.1.3 全连接层………………………………………………………………………15

3.2 FeatherNets………………………………………………………………………………15

3.2.1 FeatherNets设计………………………………………………………………15

3.2.2 网络架构细节…………………………………………………………………17

3.2.3 多模态融合方法………………………………………………………………18

3.3 本章小结 ………………………………………………………………………………19

1. 实验结果与分析…………………………………………………………………………20

4.1 传统方法实验测试……………………………………………………………………20

4.1.1 人脸活体检测数据集NUAA…………………………………………………20

4.1.2 评估指标………………………………………………………………………21

4.1.3 SVM分类器……………………………………………………………………21

4.1.4 实验结果与分析………………………………………………………………22

4.2 深度学习方法实验测试………………………………………………………………23

4.2.1 CASIA-SURF数据集…………………………………………………………23

4.2.2 数据增强………………………………………………………………………24

4.2.3 实验结果与分析………………………………………………………………25

4.3 本章小结 ………………………………………………………………………………25

1. 总结与展望…………………………………………………………………………27

5.1总结………………………………………………………………………………………27

5.2 展望……………………………………………………………………………………27

参考文献……………………………………………………………………………………………28

致谢…………………………………………………………………………………………………30

剩余内容已隐藏，请支付后下载全文，论文总字数：34068字