In recent years, emotion recognition methods based on deep learning and electroencephalogram (EEG) have achieved good results.However, existing methods still have issues such as incomplete extraction of emotional features from EEG and significant impact from artificially mislabeled emotional labels.A parallel separable convolution and label smoothing regularization (PSC-LSR) network model was proposed.Firstly, through the attention mechanism, EEG important time points and important channels were given greater weight to obtain shallow emotional features of EEG.Secondly, a parallel separable convolution module was used to comprehensively extract EEG emotional information and obtain deep emotional features.Finally, the emotion label smoothing regularization method was used to optimize the model parameters, which increased model’s fault tolerance probability for incorrect labels, enhanced the generalization and robustness of the network model, and improved accuracy of EEG emotion recognition.The proposed method has been validated in two datasets, in which the average accuracy rates of arousal and valence dimensions in the DEAP dataset reaches 99.23% and 99.13%, respectively.In the Dreamer dataset, the average accuracy rates for both arousal and valence dimensions reaches 97.33% and 97.25%.
