情绪识别中EEG信号的特征表示研究

发布时间：2017-12-28 21:38

本文关键词：情绪识别中EEG信号的特征表示研究　出处：《中央民族大学》2015年硕士论文　论文类型：学位论文

【摘要】：情绪识别是人工智能、人机交互等领域的关键技术。信息时代的来临,要求机器能够更友好的理解和表达人类的情绪。在现实生活中,情绪识别已经应用到医疗、教育、商业等领域.但是由于情绪是一个非常复杂的认知过程,情绪识别若想取得较好效果,有赖于不断深入的研究。脑电信号是一种电生理信号,具有客观性和精确性,能直接反映大脑的活动,因而被广泛应用到情绪识别中。本文基于脑电信号,进行情绪识别研究,重点在于情绪相关的脑电特征提取、特征表示问题研究。在本文中,主要开展了三个方面的工作,分别是：1)情绪诱发：情绪诱发是情绪识别研究的关键前提,影响着脑电数据的准确性和可用性。本文采用界内普遍认可的CAPS (Chinese Affective Picture System,中国情绪图片库)和IAPS (International Affective Picture System,国际情绪图片库)作为刺激材料,设计情绪诱导文件；以同类图片连续刺激的方式诱发被试者的三种情绪,分别是积极、中性和消极情绪。连续刺激的方式使被试者的情绪体验更加强烈,因而能获取更好的情绪识别结果。2)脑电时频特征提取：本文利用小波变换,在时频域提取了三类脑电特征、分别是子频带能量、能量比以及小波系数的根均方,他们很好的反应了情绪相关的脑电活动。SVM平均分类正确率能够达到82.87%,表明这三种特征在情绪识别中是非常有效的；与IAPS相比,CAPS刺激采集的EEG数据具有更高的情绪识别率,表明情绪存在着文化背景的差异。另外,考虑到脑电信号非线性时变特性以及情绪的过程性,我们利用了有重叠的方式截取脑电信号样本。SVM识别的结果表明了这种截取方式的优势。3)情绪脑区划分：越来越多的研究关注特定脑区的情绪特征在识别中的重要性。但对脑区的划分多是简单的基于距离和对称性原则来完成的,忽略了电极间的情绪相关性和差异性。本文提出了一种基于电极相似度聚类的情绪脑区划分方法,以脑区中心表示区域内所有电极的特征。该方法以电极时频特征计算互相关度,以相关度最大且大于阈值的方式判断电极是否连通,进而将连通电极聚为一类,即划分为一个脑区。用聚类中心表示类中所有电极的特征,消除了数据冗余,也达到了降维的效果。与一般的基于距离进行脑区划分的方法相比,该方法获得了更高的情绪识别率。
[Abstract]:Emotion recognition is the key technology in artificial intelligence, human-computer interaction and other fields. The advent of the information age requires machines to be able to understand and express human emotions in a more friendly way. In real life, emotion recognition has been applied to medical, educational, commercial and other fields. However, because emotion is a very complex cognitive process, emotion recognition, if we want to achieve better results, depends on continuous in-depth research. Electroencephalogram (EEG) is a kind of electrophysiological signal, which is objective and accurate, and can directly reflect the activity of the brain, so it is widely used in emotion recognition. In this paper, based on EEG, the study of emotion recognition is focused on the study of EEG feature extraction and feature representation. In this paper, there are three main works: 1) emotion induction: emotion induction is the key prerequisite for emotion recognition research, which affects the accuracy and availability of EEG data. The industry generally accepted CAPS (Chinese Affective Picture System, China emotional picture library) and IAPS (International Affective Picture System, the International Affective Picture System) as stimuli, design emotion induced by documents; three kinds of emotion induced by similar images of continuous stimulation mode was that subjects are positive, neutral and negative emotions. The method of continuous stimulation makes the subjects' emotional experience more intense, and thus can obtain better results of emotion recognition. 2) EEG time frequency feature extraction: in this paper, three kinds of EEG characteristics, namely the subband energy, energy ratio and wavelet coefficients of root mean square are extracted from wavelet transform, and they are very good responses to emotional related brain activity. The average classification accuracy of SVM can reach 82.87%, indicating that these three characteristics are very effective in emotion recognition. Compared with IAPS, the EEG data collected by CAPS has higher emotional recognition rate, indicating that there are differences in cultural background between emotions. In addition, considering the nonlinear time-varying characteristics of EEG signals and the process of emotion, we use the overlapping method to intercept the samples of EEG signals. The results of SVM recognition demonstrate the advantage of this interception. 3) emotional brain region division: more and more attention is paid to the importance of emotional characteristics in specific brain areas in recognition. However, the division of the brain area is mostly done based on the principle of distance and symmetry, ignoring the emotional correlation and difference between the electrodes. In this paper, an emotional brain region division method based on electrode similarity clustering is proposed, which represents the characteristics of all electrodes in the region of the brain region. The method calculates the correlation degree based on the time frequency characteristics of electrodes, and determines whether the electrodes are connected by the maximum correlation and the maximum threshold, and then connects the connected electrodes into one class, that is, a brain area. The clustering center is used to denote the characteristics of all the electrodes in the class, and the data redundancy is eliminated, and the effect of reducing the dimension is also achieved. Compared with the general method of dividing the brain based on distance, this method obtains higher emotion recognition rate.
【学位授予单位】：中央民族大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TN911.7

【参考文献】