CASIA OpenIR  > 毕业生  > 硕士学位论文
多视角人脸视频的活体检测技术研究
其他题名Face Liveness Detection from Multiple View Face Images
王涛
学位类型工学硕士
导师李子青
2014-12-01
学位授予单位中国科学院大学
学位授予地点中国科学院自动化研究所
学位专业模式识别与智能系统
关键词人脸活体检测 多视角 3d人脸结构 3d人脸结构 人脸识别 Face Liveness Detection Multiple View 3d Face Structure 3d Face Structure Face Recognition
摘要人脸识别技术因其可靠便捷已经在现代生活中被广泛使用。人脸识别技术及系统在过去几年中获得了很大的进步,但是传统的人脸识别系统在实际应用中却能够被轻易的攻击(比如,用一张打印的照片也可以通过系统)。人脸活体检测正是这样一种技术旨在检测出伪造的人脸,抵御假人脸攻击,增强人脸识别系统的安全等级。在本文中,我们提出了通过挖掘人脸3D结构信息以及人脸的纹理信息等新的活体检测方法。人脸的3D结构对于照片和真人具有显著的差异并且在一定程度上与设备无关,所以本文利用一系列不同姿态的人脸图像恢复出人脸的稀疏3D结构进而判断真假人脸。同时,我们还发现,多视角的真人脸和照片之间在纹理上也存在着明显差异,本文利用LBP特征提取方法实现利用多视角的人脸纹理特征判断真假人脸。在不同的应用场景下,当把基于3D结构和基于纹理的活体检测方法融合到一起的时候,其检测性能又可以进一步的提升。我们还采集一个包含三种不同采集设备更贴近实际的人脸活体检测库,在其上的实验表明,本文提出的方法最高可以达到100%的正确率,表明了该方法在实际应用中的有效性和实用性。本文的主要工作包括: 1) 提出了一种基于人脸3D结构的活体检测方法,该方法通过定位输入视频中人脸的关键点位置,然后根据关键点位置,重构人脸的3D稀疏结构,最后根据重构的3D人脸结构判断真假人脸。由于照片和真人在3D结构上显著的差异,3D结构特征能够很好的辨别真假人脸。 2) 不同于传统的基于单张图片提取纹理的人脸活体检测方法,我们发现基于多视角的纹理活体检测方法,相比于单张图片在能够获得更高的精度并且在跨库条件下也有更好的泛化性能,这对于将纹理活体检测方法应用于实际是一个很大的提升。 3) 将基于3D结构和基于纹理的人脸活体检测技术融合在一起,本文实现了一种更安全更可靠的人脸活体检测方法。其在同一库和跨库的性能表明融合多种方法是解决人脸活体检测问题一个很好的方法之一。 4) 本文采集了一个新的人脸活体检测数据库,该库包含三种不同设备采集的人脸视频,是对现有单一视角人脸活体检测数据库的一个很好补充。
其他摘要Face recognition, which is security-critical, has been widely deployed in our daily life. Face recognition system has achieved great success in recently years. However, traditional face recognition technologies in practice can be spoofed easily (e.g., by using a simple printed photo). Face liveness detection is such a technique that pre-vents the spoofing attacks to enhance the security level of face recognition system. In this paper, we propose a novel face anti-spoofing method by exploiting the structure and texture information from a rotated face image sequences. In particular, we recover the sparse 3D facial structure from a series of face images with different poses. The facial structure is salient and device-independent to some extent to distinguish ge-nuine and spoofing photos. Observed that there are significant differences between the genuine and fake photo in terms of the appearance variation of rotated faces, we apply the simple LBP feature to extract the texture information to detect the spoofing photo attacks. The combination of structure and texture information can further improve face anti-spoofing performance in various environments. Extensive experiments across three databases collected by different devices show that the proposed method can achieve 100% accuracy under some conditions, validating its effectiveness and practicality in real applications. The contributions of this work can be summarized as below: 1) A sparse 3D structure recovery based on detected landmarks from rotated face sequence is proposed. Based on the significant structure difference between the genuine and spoofing samples, the recovered structure is a useful feature to diffe-rentiate genuine and photo replica. 2) Unlike the previous texture based work which is based on a single frontal face image, we find that the texture based method with rotated face sequences can achieve significantly high accuracy and good generalization on cross databases, which is very promising in real application. 3) By combining the structure and texture information, we propose an accurate and reliable face anti-spoofing method based on rotated face sequences. The good performance on both within and cross databases indicates it is a good solution to face anti-spoofing problem. 4) A novel face anti-spoofing database, captured by three different devices are col-lected. The face database consists of rotated face sequences is a good comple-mentary to the existing ones.
其他标识符201128014628057
语种中文
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/7740
专题毕业生_硕士学位论文
推荐引用方式
GB/T 7714
王涛. 多视角人脸视频的活体检测技术研究[D]. 中国科学院自动化研究所. 中国科学院大学,2014.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
CASIA_20112801462805(3323KB) 暂不开放CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[王涛]的文章
百度学术
百度学术中相似的文章
[王涛]的文章
必应学术
必应学术中相似的文章
[王涛]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。