基于双时间尺度卷积神经网络的微表情识别

发布时间：2018-03-24 22:23

本文选题：自发微表情　切入点：双时间尺度　出处：《西南大学》2017年硕士论文

【摘要】：人类面部表情在人们的日常生活、交流中扮演着十分重要的角色。通常,我们所指的人类面部表情被称之为“宏表情”,其持续时间一般在0.5s~4s之间,容易被人察觉和辨别。然而,有心理学研究表明,“宏表情”在表达人类真实情感上具有一定的掩饰性,即面部“宏表情”能够掩饰真实情感的流露,而与“宏表情”相对的“微表情”,由于其能够表达人类试图压抑的情感,近年来受到了人们的广泛关注。微表情是一种不受人控制的、简短的面部表情,它能够反映人试图掩饰的情感以及人未意识到的情感体验,因此通过“微表情”来识别人类的情感显得更加真实、可靠。遗憾的是,由于“微表情”具有持续时间短(1/25s~1/5s),活动幅度、区域小等特点,不仅人难以识别,并且在利用模式识别等方法对微表情视频片段进行分类识别时,很难有效的表征不同微表情所具有的特征信息;除此之外,由于自发微表情数据库难以采集,数据量缺乏等要因素,使得训练一个有效的微表情识别算法也变得十分艰难。针对以上问题,本文提出了一种利用双时间尺度卷积神经网络(DTSCNN)对微表情进行识别的方法。该方法首先对微表情数据集(CASMEI、CASMEII)进行扩充处理,以此降低网络训练过程中过拟合的风险,然后利用双通道卷积神经网络分别对微表情视频序列在64fps和128fps两个时间尺度进行特征提取,最后对所提取的特征采用SVM进行决策级融合分类。DTSCNN不仅解决了由于微表情数据库样本少、难以训练的问题,而且在CASMEI、CASMEII数据库上验证的结果显示其识别率(66.67%)比最新的、传统的微表情识别算法(MDMO:55.45%、FDM:56.97%、STCLQP:56.36%)的识别率提高了10%以上。
[Abstract]:Human facial expressions play a very important role in people's daily life and communication. Usually, we refer to human facial expressions as "macro expressions". The duration of facial expressions is generally between 0.5s~4s, which is easy to be detected and distinguished. However, Psychological studies have shown that "macro expression" has a certain concealment in expressing human true emotion, that is, facial "macro expression" can conceal the expression of real emotion. "microexpressions", as opposed to "macro expressions", have attracted widespread attention in recent years for their ability to express feelings that humans are trying to suppress. Microexpressions are an uncontrolled, brief facial expression. It reflects the emotions that people try to hide and the emotional experiences they don't realize, so it's more real and reliable to identify human emotions through "microexpressions." unfortunately, Because the "microfacial expression" has the characteristics of short duration of 1 / 25 / 1 / 5 / 5 s-1, range of activity, small area, etc., it is not only difficult for people to recognize, but also in the process of classifying and recognizing microfacial video fragments by using pattern recognition and other methods. It is difficult to effectively represent the characteristic information of different microexpressions. In addition, because the spontaneous microfacial expression database is difficult to collect, the amount of data is scarce and so on. It makes it very difficult to train an effective micro-expression recognition algorithm. In this paper, a method of recognition of microfacial expression by using dual time scale convolution neural network (DTSCNN) is presented. The method firstly expands the data set of microfacial expression (CASMEI / CASMEII) to reduce the risk of over-fitting in the course of network training. Secondly, two-channel convolution neural network is used to extract the features of microfacial video sequences at 64fps and 128fps time scales, respectively. Finally, SVM is used to classify the extracted features in decision level fusion classification. DTSCNN not only solves the problem that it is difficult to be trained because of the small number of samples in the microfacial expression database, but also shows that the recognition rate is 66.67% higher than that of the latest one, which is verified on CASMEI / CASMEII database. The recognition rate of the traditional microfacial expression recognition algorithm, MDMO: 55.45 / FDM: 56.97 / STCLQP: 56.36, has increased by more than 10%.
【学位授予单位】：西南大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【参考文献】