Abstract: In order to fully utilize the spatial distribution characteristics and multi-domain deep level features between electrodes, and improve the accuracy of EEG emotion recognition, a method using multi-scale and multi-channel mixed attention mechanism is proposed for EEG emotion recognition. Firstly, baseline processing and spatial reconstruction are performed on EEG to reduce individual differences while enhancing spatial feature information. Then, the time-frequency two-dimensional and spatial three-dimensional data are jointly input into the multi-scale and multi-path hybrid attention network (MS-MPHAN) for training. A dual stream spatiotemporal fusion module is designed to strengthen the connection between temporal and spatial features; it introduces multi-scale convolution kernels for preliminary feature extraction at multiple scales, thereby increasing the field of view dimension of features; it uses a multi-channel structure to extract deep features at different levels, and enhances and fuses features at the channel level, two-dimensional spatial level, and refined coordinate level. Finally, EEG emotion four classification recognition is performed on the DEAP and DREAMER datasets, and the results show that the average accuracy of the DEAP and DREAMER datasets reaches 93.75% and 98.93%, demonstrating the good performance of the proposed method in EEG emotion recognition.

Key words: neural network, electroencephalogram, emotional recognition, attention mechanism

摘要： 为了充分利用电极间的空间分布特性和多域深层次特征，提高脑电信号（electroencephalogram, EEG）情绪识别精度，提出一种采用多尺度多路混合注意力机制的方法进行脑电情绪识别。对EEG进行基线处理与空间重构，减小个体差异的同时增强空间特征信息。将时频二维和空间三维数据共同输入到多尺度多路混合注意网络（multi-scale and multi-path hybrid attention network，MS-MPHAN）进行训练，其中，设计双流时空融合模块加强时域特征和空域特征的联系；引入多尺度卷积核进行多尺度的初步特征提取，进而增加特征的视野维度；利用多路结构二次提取不同层次的深层特征，并在通道层次、二维空间层次和精细化的坐标层次对特征进行增强与融合。在DEAP数据集和DREAMER数据集上进行EEG情绪四分类识别，结果显示DEAP数据集和DREAMER数据集的平均准确率达到93.75%和98.93%，证明了所提方法在脑电情绪识别上的良好性。

关键词: 神经网络, 脑电信号, 情绪识别, 注意力机制