基于Swin Transformer的四维脑电情绪识别

Swin Transformer-based 4-D EEG Emotion Recognition

近年来,基于脑电图(Electroencephalogram, EEG)的情绪识别研究主要使用卷积神经网络、循环神经网络和深度信念网络模型。这些方法能利用全局差异来区分不同情绪状态,但忽视了局部脑电的变化对情绪状态的影响。针对上述问题,使用了一种基于Swin Transformer的EEG四维脑电情绪识别模型,能够更好地捕捉到细小的局部空间特征和复杂的时间序列特征。相较于其它情绪识别方法,该模型通过基于滑动窗口的自注意力机制提高了不同块之间的特征连接,使得模型的建模能力更强,也降低了计算的复杂度。此外,利用情绪脑电公开数据集SEED来评估该模型的可行性与有效性,在单被试情绪三分类的准确率为94.73%±1.72%,跨被试情绪三分类的准确度为89.63%±3.42%,并且测试速度能达到实时处理的水平。实验结果表明,基于Swin Transformer的EEG四维脑电情绪识别模型通过局部特征的学习在小样本训练上也能达到较高的情绪分类准确率和较快的测试速度。

In recent years,electroencephalogram(EEG)-based emotion recognition has focused on the use of convolutional neural networks,recurrent neural networks and deep belief network models.These methods can use global differences to distinguish between different emotional states,but ignore the effect of local EEG changes on emotional states.To address these issues,we use a 4-dimensional EEG emotion recognition model based on the Swin Transformer.The model can better capture both small local spatial features and complex time-series features.Compared with other emotion recognition methods,the model proposed improves the feature connectivity between different blocks through a self-attention mechanism based on shifted windows,which makes the model more modelable and also reduces the computational complexity.In addition,we use the public emotion EEG dataset SEED to evaluate the feasibility and effectiveness of this model,with an accuracy of 94.73%±1.72%for single-subject emotion triple classification and 89.63%±3.42%for cross-subject emotion triple classification,and the testing speed can reach the level of real-time processing.The experimental results show that 4-D EEG emotion recognition based on the Swin Transformer model can achieve high emotion classification accuracy and fast testing speed even with small sample training through local feature learning.