CASIA OpenIR  > 毕业生  > 硕士学位论文
基于脸部图像的个性分析
秦日臻
学位类型工学硕士
导师胡占义
2016-05
学位授予单位中国科学院大学
学位授予地点北京
学位专业模式识别与智能系统
关键词现代相面术 人格特质 智力水平 人脸特征 机器学习方法
其他摘要      个性是指一个人在应对周围环境时所表现出来的系统和独特的反应方式,一个人的个性在人际交往过程中发挥着非常重要的作用。日常生活中,我们总是会下意识的根据一个人的脸部特征对其个性进行推断。而且,大量的研究结果表明:一方面,我们根据人脸推断出的性格在一定程度上具有可靠性;另一方面,我们在根据人脸对个性做出推断的时候,所依据的脸部特征具有高度的一致性。因此,本文通过实验来探究是否能够根据人类脸部图像来推断出一个人的个性。为了对这一问题进行研究,首先,我们提取了具有判别力的脸部图像特征;其次,基于这些人脸特征,我们从分类和回归两个方面对个性预测问题进行了分析。本文的主要工作包括以下几个方面:
      1. 设计了两种特征提取方法,分别提取了脸部图像的结构特征和纹理特征。提取脸部图像的结构特征时,使用人脸关键点检测算法检测了人脸图像的21个关键点的位置,然后基于这些关键点之间的空间几何关系构造了一个1134维的特征向量。提取脸部图像的纹理特征时,我们采用了五种广泛使用的特征描述方法(HOG,LBP,Gabor,GIST 和 SIFT),首先分别提取了人脸图像的五种纹理特征,然后通过降维操作并将这些纹理特征串接起来之后,构成了一个新的全局纹理特征。在构造人脸图像的结构特征时,为了提高人脸关键点检测精度,提出了一种新的人脸形状初始化方法,来增强非约束环境下关键点检测的鲁棒性。   
      2. 为了通过实验探究能否根据人类脸部图像准确的推断一个人的个性,我们首先构建了一个包含有186个样本的实验数据集。针对每个样本,我们采集了被试者的脸部图像,并测试了其人格特质(用来描述个性)及智力水平。然后,基于这一实验数据集,我们分别使用分类和回归方法对个性预测问题进行探究。分类实验中,我们使用五种分类方法来预测人格特质和智力水平的类别标签;回归实验中,我们使用六种回归方法预测对应的得分。分类正确率、回归误差以及皮尔逊相关系数等实验结果表明,有些人格特质和人脸图像的相关性较大(如“责任性”和“怀疑性”),可以基于人脸进行比较准确的预测,而大多数人格特质和人脸图像的相关性较小,无法进行可靠的预测。;       Personality is a kind of systematical and distinctive react expressed by the people when they deal with the surrounding environment. The personality always plays a crucial role in human relations. In our daily life, we often evaluate other people’s personality based on their face unconsciously, and many experimental results demonstrate that on the one hand, there is some truth behind the facial traits evaluation, and on the other hand, the personality prediction based on the face is in a highly similar manner. Therefore, in this thesis, we experimentally explore whether the personality traits can be predicted reliably from a facial image. In order to study this problem, we extract the distinctive facial feature firstly. Then, the personality prediction is cast as a classification problem and regression one respectively based on these facial features. The main contributions include:
      1. We design two kinds of methods to extract structural facial feature and appearance facial feature respectively. When extracting the structural feature, we use a face alignment method to detect the positions of the 21 facial salient points, and construct a feature of 1134 dimensions from the geometric relations among them. In order to construct the appearance facial feature, we first use five widely used texture descriptors (HOG, LBP, Gabor, GIST and SIFT) for subfeature extraction, then the five subfeature vectors are dimensionally reduced and concatenated. When constructing the structural facial feature, we propose a new shape initialization method to enhance the robustness of the face alignment method in the unconstrained environment.
      2. In order to investigate whether the personality traits can be predicted reliably from a facial image, we create a dataset containing 186 samples. For each sample, we take his or her facial image, and measure the personality traits and the intelligence of the participants using some questionnaires. Then we use the classification method and regression method to study the problem based on this dataset. In the classification experiments, we use five kinds of classification methods to predict the category labels of the personality traits and the intelligence. In the regression experiments, we use six kinds of methods for the personality trait prediction. The classification accuracy, the Root Mean Square Error and the Pearson Correlation Coefficient show that some personality traits are more related to facial morphological characteristics and can be predicted more accurately, while some other personality traits may have little correlation with facial features and can not be predicted reliably.
学科领域工学
语种中文
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/11492
专题毕业生_硕士学位论文
作者单位中国科学院自动化研究所
推荐引用方式
GB/T 7714
秦日臻. 基于脸部图像的个性分析[D]. 北京. 中国科学院大学,2016.
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
基于脸部图像的个性分析.pdf(19193KB)学位论文 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
[秦日臻]的文章
百度学术
百度学术中相似的文章
[秦日臻]的文章
必应学术
必应学术中相似的文章
[秦日臻]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。