基于脑电同源样本捆绑法的情绪识别研究^

Study on Emotion Recognition with Integrating EEG Homologous Samples Method

近年来,越来越多的研究者投入到基于脑电的情绪识别研究中,新理论方法的提出使得情绪的识别准确率有了进一步提高。然而,一旦走向实际应用,识别率大幅度下降,建立高精确率的情绪识别模型仍面临巨大的挑战。一个非常重要的原因是同源样本携带的非情绪信息使情绪识别模型更容易识别测试集的样本,使得传统的情绪识别方法出现识别正确率虚高的问题。为此,提出一种新的样本划分方法-同源样本捆绑法,并对10名被试开展情绪诱发实验,利用情绪图片和视频诱发平静、愉悦、难过和恶心等4种情绪状态。提取6个频段的脑电信号（θ∶4~8 Hz,α：8~12 Hz,β1：13~18 Hz,β2：18~30 Hz,γ1：30~36 Hz和γ2：36~44 Hz）的功率谱特征,然后利用支持向量机进行情绪识别,并对比传统随机方法与同源样本捆绑法的情绪识别效果。结果显示,对于图片诱发任务,随机法四分类平均正确率为43.92%,而捆绑法只得到34.15%的平均正确率;在视频诱发任务下,随机法和捆绑法分别得到94.45%和37.88%的平均正确率,随机法显著高于捆绑法,证明传统的样本随机划分方法会带来虚高情绪识别率。最后利用基于递归筛选的支持向量机（SVM-RFE）算法剔除基于同源样本捆绑法下的非情绪特征,提高情绪识别正确率,图片和视频诱发任务分别得到76.22%和72.53%的平均识别正确率。综上,所提出的同源样本捆绑法可以剔除非情绪因素的影响,避免传统的样本划分方法带来的虚高情绪识别率,是情绪识别从理论研究走向实际应用重要且必要的一步。

There are numerous studies measuring brain emotional status by analyzing EEGs under the emotional stimuli that have occurred,and acceptable accuracies were obtained in existing researches. However,emotion classification model would be challenged when it was applied in practical application. Shared non-emotional information in homologous samples may make the classification model easier to recognize the samples in the testing set,resulting in higher accuracies in EEG-based emotion recognition. In the pattern recognition,we proposed a new sample-divided method,named integrating homologous samples method,where the homologous samples were either used to build a classifier,or to be tested. In this paper,affective pictures and videos were used to elicit four emotional states of neutral,happy,sad,and disgust from 10 subjects,and EEG signals were recorded during the pictures or videos display. PSD were extracted from EEGs of 6 frequency bands（ θ： 4- 8Hz,α： 8- 12 Hz,β1： 13- 18 Hz,β2： 18- 30 Hz,γ1： 30- 36 Hz and γ2： 36- 44 Hz）,and then sent to a SVM for classification. The results showed that the classification accuracy was much lower for the integratinghomologous samples method（ IHSM） than for the traditional dividing the samples randomly（ TDSR）. For the image evoked task,43. 92% and 34. 15% were obtained by TDSR and IHSM,respectively. There were94. 45% and 37. 88% for video evoked task. SVM-RFE was employed to select emotional features and improved the classification rates to 76. 22% and 72. 35% for these two tasks. The proposed method avoided the overinflated accuracies brought by the traditional method, and handle this problem is an important and necessary step from the laboratory to the practical application.