基于多元模式分析的情绪脑电识别

发布时间：2018-04-08 22:03

本文选题：脑电　切入点：功率谱　出处：《电子科技大学》2017年硕士论文

【摘要】：受诸多因素的影响,情绪目前在心理学上还很难给出一个确切的定义,但是情感计算是实现高级人机交互(Human Computer Interaction,HCI)的关键技术之一。一些较早关于情绪识别的研究所用的信号都是非生理信号,一些计算机视觉技术有能力识别到非生理情感特征,如面部表情、手势和语音信息等,但这些外部特征容易被伪装而且不稳定,这很容易导致不可靠的结果。相反,从生理反应获得的各种生理指标,如脑电(EEG)、皮肤电反应、血液循环、呼吸活动等不能被伪装且稳定。目前情绪识别技术广泛应用于商业、测谎、安检等领域,如何有效识别情绪一直是一个十分值得研究的问题。本论文的主要工作如下:1.网络的方法被广泛应用于疾病识别、测谎、磁共振等模式识别领域,在已有的功率谱研究的基础上我们首先提出网络分析的情绪识别方法。多模态特征的方法在模式识别中也得到广泛地应用,其次我们提出将功率谱特征和网络特征相结合的方法以提高分类的准确率。我们用五个频段(theta,slow alpha,alpha,beta,gamma)下的功率谱作为特征进行三种情绪(正性,中性,负性)分类,结果发现在beta和gamma频段有较高的分类准确率,分别为62.8%和64.2%。然后提取五个频段下的网络属性作为特征进行情绪分类,结果也发现在beta和gamma频段有较高的分类准确率分别为56%和67%,这说明我们可以用网络的分析方法来研究情绪识别。最后我们结合功率谱特征和网络特征进行分类,在beta和gamma频段的准确率分别为63.3%和68.2%,比功率谱和网络特征的分类准确率都要高,这说明结合不同类型的特征可以提高分类准确率。2.随着人工智能的快速发展,机器学习成为了热门的研究,作为机器学习的一个重要分支,深度学习在其中扮演了十分重要的角色,深度学习能够构建学习网络并表现出来优越的分类性能。卷积神经网络(Convolutional Neural Network,CNN)是一个比较成熟的深度学习模型,目前CNN被广泛应用于图像识别,语音识别等领域,已有相关研究将CNN用在脑电上。本文通过CNN模型去研究脑电信号中的情绪识别,可以得到75.2%的准确率,大部分被试的分类准确率高于功率谱和网络属性以及融合特征的分类准确率,平均分类准确率较功率谱高11%,比网络特征高8.2%,比融合特征高7%,跨被试也能得到比较好的分类效果,结果为68.4%,这说明CNN能够用于情绪识别并可以得到较好的分类效果。
[Abstract]:Under the influence of many factors, it is difficult to give a precise definition of emotion in psychology at present, but emotional computing is one of the key technologies to realize advanced human-computer interaction (HMI).Some of the earlier studies on emotional recognition used signals that were non-physiological, and some computer vision techniques were capable of recognizing non-physiological emotional features, such as facial expressions, gestures and voice messages.But these external features are easily disguised and unstable, which can easily lead to unreliable results.On the contrary, the physiological parameters obtained from physiological responses, such as EEGG, skin electrical response, blood circulation, respiratory activity and so on, cannot be disguised and stable.At present, emotion recognition technology is widely used in business, lie detection, security inspection and other fields, how to effectively identify emotions has been a problem worth studying.The main work of this thesis is as follows: 1.The method of network is widely used in the fields of disease identification, polygraph detection, magnetic resonance and so on. Based on the research of power spectrum, we first put forward a method of emotion recognition based on network analysis.The method of multi-modal features is also widely used in pattern recognition. Secondly, we propose a method which combines power spectrum features with network features to improve the classification accuracy.Three emotions (positive, neutral and negative) were classified by using the power spectrum under the five bands of beta and gamma, which were 62.8% and 64.2%, respectively.Then the network attributes of five bands are extracted as features for emotion classification. The results also show that the classification accuracy in beta and gamma bands is 56% and 67% respectively, which indicates that we can use the network analysis method to study emotion recognition.Finally, combining the power spectrum features and the network features, the accuracy of beta and gamma is 63.3% and 68.2% respectively, which is higher than the classification accuracy of power spectrum and network features.This shows that combining different types of features can improve the classification accuracy. 2. 2.With the rapid development of artificial intelligence, machine learning has become a hot research, as an important branch of machine learning, in-depth learning plays a very important role in it.Deep learning can build learning networks and demonstrate superior classification performance.In this paper, the CNN model is used to study the emotion recognition in EEG signals, and the accuracy is 75.2%. The classification accuracy of most of the subjects is higher than that of power spectrum, network attributes and fusion features.The average classification accuracy is 11% higher than the power spectrum, 8.2 higher than the network feature, 7% higher than the fusion feature. The results show that CNN can be used in emotion recognition and can get better classification effect.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：B842.6

【参考文献】