摘要
随着生物认证技术的飞速发展，近年来人脸识别在许多行业都已经得到商用。然而不容忽视的是，人脸识别系统极易受到非法用户的虚假攻击，目前市面上的人脸识别系统还不具备识别出真实用户和非法用户的功能。因此，人脸防欺骗技术成为计算机视觉中的一项重要的研究课题。现有的人脸防欺骗方法往往有计算复杂度高、需要用户配合检测效果不好等缺点。
基于此，本文提出了两种改进型的人脸活体检测方法，并且优化了传统的人脸检测方法。本文的主要工作包括：
1.在人脸活体检测之前，先进行人脸检测，而传统的SMQT-SNOW人脸检测方法检测人脸不够精准。因此本文加入了3000fps人脸对齐算法，通过定位人脸的68个特征点，对SMQT-SNOW得到的人脸框进行优化，最终获得了精准的人脸检测结果。
2.提出改进的基于微纹理差异的人脸防欺骗方法。通过对人脸灰度图像进行四级Haar小波分解，得到一到四级的子图，提取每级高频子图系数矩阵的均值、方差作为特征向量；同时对人脸灰度图像提取ULBP特征，并统计直方图向量，级联后得到75维特征向量，最后送入SVM分类器进行识别分类。
3.提出改进的并联卷积神经网络（P-CNN）的人脸防欺骗方法。基于深度学习的角度，提出了P-CNN模型，
并首次结合ELM方法。P-CNN分别包括一个改进的八层和一个改进的六层的模型，两个模型的输入图像不一样，以学习到更加全面的特征，并利用PCA降维，采样层交替采用均值采样、随机采样和重叠采样，并分别加入了dropout正则化防止过拟合。
两种方法分别在公开的NUAA、REPLAY-ATTACK数据库上进行实验验证，基于微纹理的方法的平均准确率分别为99.53%和92.95%，基于P-CNN结构的平均准确率分别为99.96%和97.14%。与此同时，ROC曲线图同样也表明了这两种方法的有效性。前一种方法的优势是计算法复杂度小；后一种方法的优势是检测准确率很高，并且ELM获得了较快的分类速度。本文的研究为以后其他的研究者提供了借鉴以及在人脸识别系统上进行商用提供了基础。
关键词：人脸识别，人脸防欺骗，人脸检测，微纹理，卷积神经网络
ABSTRACT
Along with the rapid development of biometric authentication technology, face recognition has been commercially used in many industries in recent years. However, It can not be ignored that the face recognition system is highly vulnerable to attacks by illegal users, the existing face recognition system can not recognize the legal users and illegal users on the market recently. As a result, the technology of face anti-spoofing becomes an important topic research topic in computer vision.
正则化研究背景和意义Therefore，this paper proposes two improved face liveness detection methods and optimizes the traditional face detection approach. The main research contents are as follows:
1. Before face liveness detection, face detection is conducted, however, traditional SMQT-SNOW face detection method can not detect face precisely. This paper adds face alignment algorithm of 3000fps into the SMQT-SNOW face detection method. 3000fps algorithm locates the 68 feature points of the face to optimize the face bounding box, finally gets more accurate face detection.
2. The paper proposes the improved face anti-spoofing approach based on the micro texture features. The algorithm decomposes face images by four-level Haar wavelet decomposition, calculates the means，variances of high frequency sub-bands coefficients matrixes which are obtained from one-level to four-level Haar wavelet decomposition , and extracts ULBP histogram , then cascades them to act as final 75 dimensional feature vectors to train SVM classifier to determine whether the given face is genuine or not.
3. The paper proposes a novel face anti-spoofing approach based on improved parallel convolutional neural network(P-CNN). From the angle of deep learning, P-CNN is designed, and the ELM method is first utilized for face liveness detection. The P-CNN network includes an improved eight-layer model an
d an improved six-layer model respectively. The input image of the two models are not same to learn comprehensive features, then PCA is used to reduce dimensions.The models both have alternating sampling layer using mean sampling, random sampling and overlapping sampling, and utilize the dropout regularization to prevent overfitting.
Experimen verifications about the two approaches are conducted on the two public face databases, e.t. NUAA and REPLAY-ATTACK, the average accuracy rates of the method based on micro-texture are 99.53% and 92.95% respectively, and the average accuracy rates of the method based on P-CNN are 99.96% and 97.14% respectively. Meanwhile pictures about ROC curves show the effectiveness about the proposed methods. The advantage of the first method is that the computational cost is low, the latter’s is the high detection accuracy rate, and ELM acquires the fast classification speed. It provides technological foundation to the future researchers and reference in commercial face recognition system containing face liveness detection technology.
KEY WORDS: Face recognition, Face anti-spoofing, Face detection, Micro texture, Convolutional neural network
目录
摘要............................................................................................................................ I ABSTRACT ............................................................................................................... I II 第1章绪论 (1)
1.1研究背景及其意义 (1)
1.2人脸识别系统中的攻击方式 (2)
1.3国内外研究现状 (3)
1.3.1基于硬件环境的方法 (3)
1.3.2基于软件为基础的非入侵性的方法 (4)
1.3.3基于用户交互式的方法 (6)
1.3.4基于深度学习的方法 (7)
1.4主要研究思路和结构安排 (7)
1.4.1主要研究思路 (7)
1.4.2结构安排 (8)
第2章人脸活体检测理论基础 (9)
2.1人脸活体检测流程 (9)
2.2常用的纹理特征描述子 (10)
2.2.1局部二进制模式 (10)
2.2.2共生邻接局部二值模式 (10)
2.2.3局部相位量化 (11)
2.2.4二值统计图像特征 (11)
2.2.5规模不变描述符 (11)
2.3主流的人脸防欺骗技术 (11)
2.4 SVM分类器 (16)
2.5人脸样本数据库介绍 (18)
2.6本章小结 (20)
第3章基于微纹理差异的人脸防欺骗方法 (21)
3.1人脸检测 (21)
3.2算法基础 (24)
3.2.1 Haar小波分解 (24)
3.2.2 ULBP (25)

版权声明：本站内容均来自互联网，仅供演示用，请勿用于商业和其他非法用途。如果侵犯了您的权益请与我们联系QQ:729038198，我们将在24小时内删除。