基于3D矩阵特征与多维卷积网络的脑电信号情感识别

Emotion recognition of EEG signal based on 3D matrix features and multi-dimensional convolutional network

人类大脑皮层能够对不同情感产生动态响应,在这一神经科学研究成果的启发下,提出一种基于3D矩阵特征与多维卷积神经网络的脑电(Electroencephalogram,EEG)信号情感识别方法,用MCNN表示.该3D矩阵特征是指每个时间点上提取的6个频带的PSD特征按大脑电极位置分布,转换为9×9网状矩阵后连接得到的一个9×9×6的三维矩阵,该表征方法能够直接准确反应大脑皮层EEG信号的空间相关性和时频动态,再将该特征输入一个多维卷积神经网络进一步提取相关深度语义特征并进行情感分类.所提方法在DEAP数据集中脑电信号的唤醒度和效价维度上两类情感分类的平均准确率分别达到了85.88%和87.32%,相同实验条件下,比目前较优手工特征的平均分类准确率分别提升了5.41%和5.69%,比最优深度模型的平均分类准确率分别提升了3.52%和4.18%,验证了该方法的先进性和有效性.

Inspired by the neuroscience research results that the human brain can dynamically respond to different emotions,an EEG(Electroencephalogram,EEG)emotion recognition method based on 3D matrix features and multi-dimensional convolutional neural networks(3D-MCNN)is proposed.The 3D matrix feature is a 9×9×6 three-dimensional matrix obtained by concatenating the 6 frequency bands of PSD features separately expressed in 9×9 mesh matrix converted from 1-D chain feature according to the brain electrode position distribution.The 3D matrix features can directly and accurately reflect the spatial correlation and time-frequency dynamics of the cortical EEG channels,which are then input into a multi-dimensional convolutional neural network,to learn deep discriminative semantic features and then make emotion classification.The proposed method has an average accuracy of 85.88%and 87.32%in the arousal and valence dimensions of EEG signals in the DEAP dataset.Under the same experimental conditions,it is more accurate than the current average classification of better manual features.The rate has increased by 5.41%and 5.69%,respectively,which is 3.52%and 4.18%higher than the average classification accuracy of the optimal depth model,which verifies the advanced nature and effectiveness of the method.