毕业生知识库

人脸属性识别在人机交互、安防系统等众多领域有着广阔的应用前景，受到了研究者们的广泛关注，

已成为计算机视觉和模式识别领域的热点研究课题。目前，由于样本多样性、

数据收集困难等问题都对人脸属性的识别算法提出了重重挑战，从而限制了其在实际场合中的应用与推广。

本文对人脸属性中的活体检测、年龄估计和表情识别课题的关键问题进行了探索和研究。

主要工作和贡献有：

 

1、基于判别高斯过程隐变量模型的年龄估计。受到基因、生活习惯等多种内外因素的影响，

不同个体的年龄发展模式截然不同。本文提出了一种基于流形学习的年龄估计方法，

旨在从不同用户的大量数据中挖掘出可能潜在的年龄模式。

主要贡献为：在判别高斯过程隐变量模型的基础框架上加入隐变量的基于广义判别分析的

先验信息，

从而将其从解决线性可分问题推广到线性不可分问题——年龄估计。

 

2、基于多形状对齐的表情识别。针对表情识别中的几何多样性问题，

即不同个体对同一种表情的表达程度不同，甚至同一个体在不同时间不同场景下也不同，

本文提出了一种基于多形状对齐的表情识别方法，

旨在减少几何多样性的同时尽可能地保持不同表情类别下个体的差异性。

该方法首先通过不同表情类别的样本获取相应的平均形状，进而基于多个平均形状进行加权投票

获取最后的表情类别预测。和传统基于单一人脸形状模板相比，基于多形状模板的方法充分利用了不同表情的

特点，有效提高了表情识别的准确率。

 

3、基于视线估计的活体检测。首先从单一人脸线索出发，基于合法用户的视线具有不可预测

的特点，提出了基于视线估计的活体检测方法。该方法预先建立屏幕中多个视点的视线估计统计模型，

进而通过估计用户的视线行为进行活体判断。

提出的方法包括两种框架，分别用于处理照片与视频播放的识别和抵御。

首先提出了基于信息熵的方法用于抵御照片攻击，该方法采用信息熵指标来衡量用户在短时间内视线

行为的不确定程度。其次，进一步提出基于挑战-响应机制的方法用于抵御视频播放。

受随机验证码的启发，系统发出的随机点序列是挑战，用户的视线行为是对应的响应。

综合使用以上两种基于视线估计的活体检测方法可以有效识别和抵御人脸活体检测中的照片和视频播放攻击，

为活体检测提供了一种新的思路。

 

4、基于镜面反射和关键点区域变化的活体检测。针对单一人脸线索不能同时应对多种攻击的问题，

本文提出了一种基于混合线索的活体检测方法，该方法在得分层面上对两种线索进行融合。

首先，实现了一种基于镜面反射的人脸描述子，并基于支持向量机获取模型得分。

其次，提出一种描述人脸非刚体运动的特征描述子，用于表征人脸图像序列的运动状态。

并且基于贝叶斯理论定义了和该描述子有关的模型得分。

最后，提出了基于得分对齐的融合策略。该融合方法有效增加了能够同时应对的攻击人脸形式，

提高了活体检测的识别率。

 

5、基于指令选择的活体检测。鉴于仅依靠分析人脸线索并不能完全区分合法用户与攻击者，

本文关注了基于系统指令和用户配合的活体检测方法。针对该类方法中无法确定系统指令

个数这一问题，提出了基于指令选择的活体检测方法，旨在综合考虑系统识别准确率和用户体验。

该方法通过在损失函数的基础上加入基于群组的指令稀疏项构造指令选择学习模型实现指令的选择。

提出的方法首次在人脸活体检测方法中加入用户体验的因素，在实际应用中具有一定的指导意义。

 

With the development of human-computer interactions, face attribute recognition has received a great deal of attentions and has become a hot topic in pattern recognition and computer vision.

However, existing face attribute recognition techniques

still have many challenges, such as the diversity of samples and the difficulties of sample collection.

The performance of existing face attribute recognition algorithms is far behind

satisfactory and hence the application in real world is limited.

 

In this thesis, we focus on the issue of face liveness detection,

age estimation and facial expression recognition in face attribute recognition field.

The main contributions of this thesis include the following issues:

 

1. Age estimation based on discriminant Gaussian process latent variable model. Affected by various factors (for example, genes and living habits), different people present distinct aging patterns. To discover the underlying trend of aging patterns, in this paper we propose a novel age estimation method based on improved DGPLVM (Discriminative Gaussian Process Latent Variable Model). DGPLVM is an effective probabilistic model for manifold learning, in which

the corresponding low-dimensional representations of high-dimensional data are

given linearly separable discriminative priors and achieved by maximizing the posterior

probability. Considering that age estimation is a complex and nonlinear problem,

we propose improved DGPLVM to get the low-dimensional representations

by placing nonlinearly separable priors instead of linear one.

 

2. Facial expression recognition based on multiple base shapes.

Geometric variation is one of the important components deteriorating the facial

expression recognition performance. Aligning the face image to a base shape is a commonly used preprocess step to alleviate the variation. However, the assumption of single base shape can not necessarily guarantee the best performance. In this paper, we propose a facial expression recognition framework based

on multiple base shapes, which aims to minimize the geometric variation

between face images with the same facial expression and retain the

geometric shape difference between face images with different facial expressions.

 

3. Face spoofing detection based on gaze estimation. From the perspective of single facial clue

of valid user, considering the unpredictability of gaze movement, we propose two face spoofing

detection methods based on gaze estimation for photo and replay attack, respectively. In proposed work, statistical models of multiple points on the screen are built in advance and nonlinear regression is used to predict user's gaze for liveness detection. Specifically, we firstly proposed the photo spoofing detection method based on information entropy which is used as the measurement for evaluating the uncertain degree of gaze movement.

Secondly, we further propose an effective replay attack detection method under challenge-response mechanism. The random points on the computer

screen create the challenge, and the gaze positions of the user as they

look at the computer screen form the response. The matching degree between predicted gaze positions

and system point locations is used to evaluate the liveness of user. A combination of two methods mentioned above

can effectively deal with photo and replay attack and offers a new research way by using gaze estimation.

 

4. Face spoofing detection based on specular component and landmark patch movement. Considering that the method based one single facial clue can not deal with multiple forms of attacks at the same time, we present an effective face spoofing detection method based on hybrid models on the score level. Firstly, an specular component based descriptor is employed to capture the texture information and the model score is obtained by support vector machine. Secondly, a patch movement based descriptor is presented to capture the face region motion and model score is defined based on Bayesian theory. Finally, score fusion is performed according to score calibration strategy. This work can detect effectively photo and replay attack at the same time thus improve the recognition rate.

 

5. Face liveness detection based on instruction selection learning. Besides the facial clues,

user interaction based on system instruction also plays an important role for face spoofing detection.

In this work we focus on this kind of anti-spoofing system in which the user

is asked to follow a series of instructions sent randomly by the system

and estimated whether he is a genuine access or not according to the response. While how to choose the number of instructions is an key problem. Too many instructions can guarantee

the high recognition accuracy and distinguish effectively attacks from genuine accesses,

however, it lasts too much time thus brings burden on users. Therefore, instruction selection learning based spoofing detection framework is proposed for compromising between system recognition performance and user experience. This work provides the first investigation in research literature on the user experience for user interaction.