结合互信息通道选择与混合深度神经网络的脑电情感识别方法

EEG Emotion Recognition Based on Normalized Mutual Information Channel Selection and Hybrid Deep Neural Network

针对单模态深度神经网络(Deep Neural Network, DNN)难以充分提取情感分类任务中脑电信号的多域特征,且脑电信号中存在通道冗余的问题,提出一种结合互信息通道选择与混合深度神经网络的脑电情感识别方法,首先提取各通道信号中γ节律的微分熵(Differential Entropy, DE)特征,通过DE计算通道间的归一化互信息(Normalized mutual information, NMI),将所得NMI矩阵按列求和后的向量作为表征各通道任务相关性的权值,根据权值选出最优通道集,之后采用卷积神经网络(Convolutional Neural Networks, CNN)和长短期记忆网络(Long-short term memory neural network, LSTM)相结合的混合DNN网络进行样本特征提取和分类。该方法分别在DEAP数据集的效价(Valence)和唤醒度(Arousal)上取得了87.60%和88.58%的平均分类准确率,表明了所提出方法的可行性和有效性。

Aiming at solving the problem that single mode deep neural network( DNN) fails to fully extract the multidomain features of EEG signals and redundant channel selecting in EEG emotion recognition tasks. A Normalized Mutual Information( NMI) based channel selection for hybrid DNN is proposed in this paper. Firstly the differential entropy( DE) are extracted from the γ EEG rhythm. The NMI are calculated between each channel according to the DE. The vector summed by the column of the NMI matrix is used as the weight to characterize the task correlation of each channel. Then the optimal channel set is selected by channel’s weight. Secondly the hybrid DNN with convolutional neural network( CNN) and long-term and short-term memory network( LSTM) is used for feature extraction and classification after channel selection. Extensive binary emotion classification experiments are carried out on DEAP dataset,and proposed method achieved 87.60% and 88.58% average accuracy in Valence and Arousal respectively,which shows the feasible and effective of proposed method.