【摘要】 基于脑电信号的情绪识别是人机交互的重要部分，本文将二维卷积神经网络、三维卷积神经网络、深度可分离卷积进行结合，提出一种基于2D-3D卷积神经网络(2-3DCNN)模型，从时间、空间、频率三个方面进行特征提取。在网络中引入SE-ResNet网络、深度残差收缩网络和Xception网络，挖掘脑电信号中更能显著反映情感变化的空间、时间和频率信息。本文在DEAP公共情感数据集上做性能测试，结果表明，2-3DCNN在唤醒度和效价的两个分类任务上的识别准确率分别达到了97.59%和97.21%,比目前最先进的模型分别高出2.36%和1.34%。更多还原

【Abstract】 Emotion recognition based on EEG signals is an important part of human-computer interaction.This paper combines two-dimensional convolutional neural network, three-dimensional convolutional neural network and depth-wise separable convolution, and proposes a 2D-3D convolutional neural network(2-3DCNN) model to extract features from three aspects: time, space and frequency.SE ResNet network, deep residual shrinkage network, and Xception network are introduced into the network to explore spatial, temporal, and frequency information in EEG signals that can more significantly reflect emotional changes.Experiments are conducted on the DEAP public sentiment dataset using the 2-3DCNN model, and the results show that the recognition accuracy of this method in arousal and valence classification tasks reached 97.59% and 97.21%,2.36% and 1.34% higher than the current state-of-the-art models.更多还原