蔡冬丽, 钟清华, 朱永升, 张涵. 基于混合神经网络的脑电情感识别[J]. 华南师范大学学报(自然科学版), 2021, 53(1): 109-118. doi: 10.6054/j.jscnun.2021017
引用本文: 蔡冬丽, 钟清华, 朱永升, 张涵. 基于混合神经网络的脑电情感识别[J]. 华南师范大学学报(自然科学版), 2021, 53(1): 109-118. doi: 10.6054/j.jscnun.2021017
CAI Dongli, ZHONG Qinghua, ZHU Yongsheng, ZHANG Han. EEG Emotion Recognition Based on Hybrid Neural Networks[J]. Journal of South China Normal University (Natural Science Edition), 2021, 53(1): 109-118. doi: 10.6054/j.jscnun.2021017
Citation: CAI Dongli, ZHONG Qinghua, ZHU Yongsheng, ZHANG Han. EEG Emotion Recognition Based on Hybrid Neural Networks[J]. Journal of South China Normal University (Natural Science Edition), 2021, 53(1): 109-118. doi: 10.6054/j.jscnun.2021017

基于混合神经网络的脑电情感识别

EEG Emotion Recognition Based on Hybrid Neural Networks

  • 摘要: 为保留脑电(Electroencephalogram,EEG)空间信息的同时充分挖掘EEG时序相关信息,提出了一种三维卷积神经网络(3-Dimensional Convolutional Neural Networks,3D-CNN)结合双向长短期记忆神经网络(Bidirectional Long Short-term Memory Neural Networks,BLSTM)的混合神经网络(3DCNN-BLSTM);为验证该模型的分类性能,在DEAP数据集和SEED数据集上进行情感识别实验. 实验结果表明3DCNN-BLSTM模型能有效学习EEG多通道间的相关性与时间维度信息且提高了情感分类性能:在DEAP数据集的二分类实验中,唤醒度和效价的情感识别平均准确率分别为93.56%和93.21%;在DEAP数据集的四分类实验中,情感识别平均准确率为90.97%;在SEED数据集的三分类实验中,情感识别平均准确率为98.90%.

     

    Abstract: A hybrid neural network (3DCNN-BLSTM) based on a 3-Dimensional Convolutional Neural Network (3D-CNN) combined with a Bi-directional Long Short-term Memory Neural Network (BLSTM) is proposed to preserve the spatial information of the EEG while taking full advantage of its time-related information. Emotion re-cognition experiments on DEAP and SEED datasets are carried out to evaluate the classification performance of the model. The experiment results show that the 3DCNN-BLSTM model can effectively learn the correlation between EEG multi-channels and time dimension information and improve the performance of emotion classification. The ave-rage accuracy of emotion recognition of arousal and valence in the two-classification experiments on DEAP dataset are 93.56% and 93.21% respectively; the average accuracy of emotion recognition in the four-classification experiments on DEAP dataset is 90.97%; and the average accuracy of emotion recognition in the three-classification experiments on SEED dataset is 98.90%.

     

/

返回文章
返回