基于多尺度卷积和混合注意力机制的情绪脑电识别研究

Emotion EEG recognition based on Multi-scale convolution and hybrid attention mechanism

建立了基于深度学习框架的情绪识别模型——M-Attention-EmotionNet。采用多尺度卷积提取不同尺度的特征。为避免特征冗余,引入了混合域注意力机制,从通道和空间两个维度对特征进行赋权。在DEAP数据集上,首先提取预处理后脑电信号不同频带的功率谱密度特征作为模型的输入,然后被分类。实验结果表明,该模型在唤醒和效价两个维度上分类准确率分别为95.64%和96.49%,同时,在四分类和八分类的细粒度情感分类上,平均准确率分别为90.89%和89.22%。

An emotion recognition model, M-Attention-EmotionNet, is built based on deep learning framework. Multi-scale convolution is applied to extract features of different scales. A hybrid domain attention mechanism is introduced to avoid feature redundancy, and each dimension weights the features to characterize the features. On the DEAP dataset, power spectral density features of different frequency bands are extracted from the pre-processed EEG signals as the input of the model to be classified. The experimental results show that the classification accuracy rates of the model in Arousal and Valence dimensions are 95.64% and 96.49% respectively. In fine-grained emotion classification of four categories and eight categories, the average accuracy rates are 90.89% and 89.22%.