毕业生知识库

随着人脸识别技术的成熟及其商业化应用的普及，人脸识别系统面临着各种各样的挑战，特别是随着高清电子设备、3D打印等仿造手段的迅速发展与应用，人脸更容易以高清晰度照片、高清视频片段、人脸面具模型等形式被伪造、复制与传播，这类安全性威胁使得研究高效、准确的活体检测技术成为目前亟待解决的重要课题。为了进一步提高活体检测系统的有效性和鲁棒性，本文将在传统的单目可见光谱摄像头外，借助更加先进的硬件设备，从双目成像、主动红外光照、RGB-D图像、多光谱成像等角度研究人脸活体检测技术。本文主要的工作和贡献有：

一、基于双目图像的活体检测：大多数常见的攻击媒介与真实人脸的深度信息有很大的差异，使用双目摄像头与三维重建技术，本文提出了一种基于双目图像的活体检测算法。该算法一方面根据双摄像头位置不同而导致的双目图像差异提取二维纹理特征，另一方面根据人脸特征点构成的稀疏点云提取三维深度特征，最后将两个特征融合起来以克服各自的缺陷，并达到提高活体检测算法准确率的效果。

二、基于主动红外光照的活体检测：每个人脸伪造方案均依赖某种攻击媒介作为展示载体，本文主要采用前景背景一致性分析策略，提出了一种旨在分析攻击媒介存在性的活体检测算法。该算法使用额外的主动红外光源，通过分别开启和关闭该光源构造相应的红外差分图像，从而对前景人脸区域与背景区域之间光照变化的一致性进行分析；同时，在人脸区域中，算法依据不同材质的反射特性提取光照纹理特征，进一步增强算法对于经过精心剪裁的攻击媒介的抵抗能力。

三、基于RGB-D图像的活体检测：使用三维深度信息会较明显地提高活体检测算法的准确性，然而较高的硬件或时间成本限制了这类方案的使用。借助较为廉价且常见的RGB-D摄像头，本文提出了一种基于RGB-D图像的活体检测算法，该算法一方面分析彩色图像与深度图像之间的关联特性，另一方面计算人脸深度图像中各区域间的一致程度，最后将两个活体检测特征进行融合，提升算法的通用性，并增强对照片、屏幕和面具等不同攻击媒介的检测效果。

四、基于多光谱关联分析的活体检测：相对单一特定光谱而言，伪造人脸同时在多个光谱种模仿真实人脸成像要困难得多，本文提出了一种基于分析多光谱间成像关联特性的活体检测算法，该算法分别在可见光谱和近红外光谱中采集被测用户的图像，通过分析这两张人脸图像特征间的潜在关联信息来进行活体检测。为了增强对不同光照等外界环境的鲁棒性，算法使用图像分区策略，将可见光和红外光图像平均分为若干子区域，并在训练阶段计算置信图以衡量各个区域对整体判别能力的贡献程度，以增强算法的准确率和跨数据集判别能力。

With the rapid development and wide commercial applications of face recognition techniques, there are more and more security challenges made by spoofing faces. Due to advanced electronic devices and three-d printing technology, spoofing faces could be easily made, copied or exchanged in terms of high-definition photos, high-definition videos or face masks, which contributes to the difficulty and importance of effectual and efficient face anti-spoofing algorithms. To make face spoofing systems robuster and more effective, we adopt more advanced sensors apart from common cameras, and make use of dual cameras, the active near-infrared light, RGB-D images and multispectral images to detect spoofing faces. The main contributions of this work are listed as follows:

1. Face spoofing detection based on dual cameras: most spoofing faces behave rather differently in terms of depth information from genuine ones. With the help of dual cameras and stereo reconstruction techniques, we propose a spoofing detection method. On one hand, the texture differences caused by dual camera positions are explored; on the other hand, depth features are extracted based on the reconstructed three-d sparse point cloud. Two spoofing cues are fused to overcome their own deficiencies and to enhance the final classification results.

2. Face spoofing detection based on active near-infrared light: every face spoofing attack needs  certain kinds of spoofing media as displaying carriers, and we propose a spoofing detection method by analyzing the existence of spoofing media. By turning on and off the additional near-infrared light, we obtain the near-infrared differential image caused by the active light, and the context consistency between face and non-face areas is analyzed. At the same time, we further extract lighting texture cues based on the reflectance model, to boost the performance of cropped spoofing faces. 

3. Face spoofing detection based on RGB-D images: while using three-d information might significantly benefit face anti-spoofing systems, it is restrained by several issues such as the expensive hardware, high time cost or poor accessibility. Thus, we utilize RGB-D images captured by relatively low cost sensors, and propose a spoofing detection method. We analyze the correlation between color and depth images, and calculate the consistency among different parts of depth images. Finally two spoofing cues are fused in decision level to make proposed method robuster towards different spoofing media and enhance its generalization.

4. Face spoofing detection based on multispectral correlation analyses: motivated by the thought that it is easier for spoofing faces to cheat on one spectrum than defraud two different spectra, we propose a face spoofing detection system analyzing potential multispectral information between visual and near-infrared images. To handle different lighting conditions, we choose part-based schema and divide both images into smaller regions. A confidence map is also calculated during the  training process to measure the contribution of each patch, which increases the final classification accuracy and the generalization ability.