基于纹理特征的2D-3D人脸活体检测关键技术研究
发布时间：2018-12-13 06:21
【摘要】：人脸识别技术是一种精度高、稳定性好、使用方便的生物识别技术,市场应用前景广阔。然而,人脸识别技术频繁受到假冒攻击(或复制攻击),仍存在诸多安全隐患。在抵抗假冒攻击(或复制攻击)方面,活体检测具有显著的效果,它对样本是否具有生命特征进行了辨识。针对人脸识别系统无法识别采集的人脸图像是否来自真人的问题,本文重点研究了基于2D人脸图像和3D人脸深度图的活体检测算法。主要工作包括:1、针对现有3D人脸活体检测数据库较少的问题,本文采集了一个RGBD人脸数据库。该数据库正样本包括使用Kinect和另一双目设备采集的104个真人在0.5-2米处不同姿态的深度人脸数据,共计20973张图片。负样本包括使用Kinect采集在不同环境下0.5-2米处不同角度的ipad、电脑、手机、照片攻击人脸,共计12300张图片。2、针对现有的傅里叶频谱分析方法较为简单且准确率较低的情况,本文提出了一种改进的傅里叶频谱特征方法。该方法在对2D人脸区域图像提取二维离散傅里叶频谱图的基础上,加入分块子空间的方法,将傅里叶频谱图分成若干个子块,并求得每一个子块内图像的平均能量值,归一化后级联成一个全局傅里叶频谱特征向量。实验结果表明,改进后傅里叶频谱特征能有效地提高2D人脸图像的活体检测准确率。3、针对在训练样本增加时,基于傅里叶频谱特征的2D人脸活体检测准确率会进一步下降的情况,本文提出了融合LBP特征的FS-LBP特征人脸活体检测方法。该方法将傅里叶频谱特征和低维的LBP特征级联,并使用SVM来分类判别。实验结果表明,该方法在2D人脸活体检测上更优于时下最主流的MSLBP特征方法。4、针对灰度共生矩阵纬度低,且其3D人脸的活体检测率仍可进一步提升的情况,本文提出了一种多尺度灰度共生矩阵的方法。该方法首先通过对RGB图像进行人脸检测并同步采集深度图的人脸区域图像,其次将人脸区域深度图调整为不同尺度大小的深度图像,并分别提取其灰度共生矩阵特征,并级联成一个多尺度灰度共生矩阵特征,最后使用SVM来分类判别。实验结果表明,该方法在3D人脸深度图上的活体检测准确率高于灰度共生矩阵特征和LBP特征方法。最后对本文工作进行了总结,并对本文后续工作进行了展望。
[Abstract]:Face recognition is a kind of biometric technology with high precision, good stability and convenient use. However, face recognition technology is frequently subjected to fake attacks (or copy attacks), there are still many security risks. In the aspect of resisting counterfeiting attack (or replica attack), in vivo detection has remarkable effect, and it identifies whether the sample has life characteristic or not. In order to solve the problem of whether the human face image can not be recognized by the face recognition system, this paper focuses on the living body detection algorithm based on 2D face image and 3D face depth map. The main work is as follows: 1. Aiming at the lack of 3D human face detection database, this paper collects a RGBD face database. The database includes 104 human face data with different pose depth at 0.5-2 meters collected using Kinect and another binocular device, with a total of 20973 images. The negative samples include ipad, computers, mobile phones and photos that use Kinect to collect 0.5-2 meters of different angles in different environments to attack faces, with a total of 12300 images. In view of the simple and low accuracy of the existing Fourier spectrum analysis methods, an improved Fourier spectrum feature method is proposed in this paper. On the basis of extracting 2D discrete Fourier spectrum from 2D face region image, the method of adding block subspace is used to divide the Fourier spectrum into several sub-blocks, and the average energy value of the image in each sub-block is obtained. After normalization, it is cascaded into a global Fourier spectrum eigenvector. Experimental results show that the improved Fourier spectrum features can effectively improve the accuracy of 2D face image in vivo detection. The accuracy of 2D human face detection based on Fourier spectrum features will be further reduced. In this paper, a FS-LBP feature based face detection method based on LBP features is proposed. In this method, Fourier spectrum features and low-dimensional LBP features are concatenated, and SVM is used to classify and discriminate. The experimental results show that this method is better than the most popular MSLBP feature method in 2D face detection. 4. Aiming at the low latitude of gray level co-occurrence matrix and the fact that its 3D face detection rate can be further improved. In this paper, a method of multi-scale gray level co-occurrence matrix is presented. The method firstly detects the face of RGB image and synchronously collects the facial region image of the depth map. Secondly, the depth map of the face region is adjusted to the depth image of different scales, and its gray level co-occurrence matrix feature is extracted respectively. And cascaded into a multi-scale gray level co-occurrence matrix feature, finally using SVM to classify discrimination. Experimental results show that the accuracy of this method is higher than that of gray level co-occurrence matrix and LBP features on 3D face depth images. Finally, the work of this paper is summarized, and the future work of this paper is prospected.
【学位授予单位】：集美大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
[Abstract]:Face recognition is a kind of biometric technology with high precision, good stability and convenient use. However, face recognition technology is frequently subjected to fake attacks (or copy attacks), there are still many security risks. In the aspect of resisting counterfeiting attack (or replica attack), in vivo detection has remarkable effect, and it identifies whether the sample has life characteristic or not. In order to solve the problem of whether the human face image can not be recognized by the face recognition system, this paper focuses on the living body detection algorithm based on 2D face image and 3D face depth map. The main work is as follows: 1. Aiming at the lack of 3D human face detection database, this paper collects a RGBD face database. The database includes 104 human face data with different pose depth at 0.5-2 meters collected using Kinect and another binocular device, with a total of 20973 images. The negative samples include ipad, computers, mobile phones and photos that use Kinect to collect 0.5-2 meters of different angles in different environments to attack faces, with a total of 12300 images. In view of the simple and low accuracy of the existing Fourier spectrum analysis methods, an improved Fourier spectrum feature method is proposed in this paper. On the basis of extracting 2D discrete Fourier spectrum from 2D face region image, the method of adding block subspace is used to divide the Fourier spectrum into several sub-blocks, and the average energy value of the image in each sub-block is obtained. After normalization, it is cascaded into a global Fourier spectrum eigenvector. Experimental results show that the improved Fourier spectrum features can effectively improve the accuracy of 2D face image in vivo detection. The accuracy of 2D human face detection based on Fourier spectrum features will be further reduced. In this paper, a FS-LBP feature based face detection method based on LBP features is proposed. In this method, Fourier spectrum features and low-dimensional LBP features are concatenated, and SVM is used to classify and discriminate. The experimental results show that this method is better than the most popular MSLBP feature method in 2D face detection. 4. Aiming at the low latitude of gray level co-occurrence matrix and the fact that its 3D face detection rate can be further improved. In this paper, a method of multi-scale gray level co-occurrence matrix is presented. The method firstly detects the face of RGB image and synchronously collects the facial region image of the depth map. Secondly, the depth map of the face region is adjusted to the depth image of different scales, and its gray level co-occurrence matrix feature is extracted respectively. And cascaded into a multi-scale gray level co-occurrence matrix feature, finally using SVM to classify discrimination. Experimental results show that the accuracy of this method is higher than that of gray level co-occurrence matrix and LBP features on 3D face depth images. Finally, the work of this paper is summarized, and the future work of this paper is prospected.
【学位授予单位】：集美大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
【参考文献】
相关期刊论文 前8条
1 谢志华;刘国栋;;基于局部二元模式的快速红外人脸识别系统[J];红外与激光工程;2013年12期
2 尹京;方艳梅;;数码翻拍图像取证算法[J];中山大学学报(自然科学版);2011年06期
3 吴贤星;赵杰煜;沈明p，
本文编号：2376059
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2376059.html
