A multi-scale deep CNN based on attention mechanism for EEG emotion recognition.

BACKGROUND: Recognizing emotion is a crucial project within the domain of brain-computer interface technology. Recently, researchers have found that deep learning have been proven to be superior to machine learning, but how to obtain more discriminative features still faces great challenges. NEW METHOD: We propose a multi-scale convolutional neural network (MSCNN) based on channel attention and spatial attention (CSA-MSCNN) for EEG emotion recognition. The channel attention enhances the feature extraction ability of critical channels by generating channel weights, while suppressing noise or interference from redundant channels. The spatial attention helps the model to more precisely locate key areas related to emotion by generating a spatial weight matrix. To extract more comprehensive features, CSA-MSCNN uses MSCNN for feature extraction, with smaller convolutional kernels capturing the local details of the signals, and larger convolutional kernels with a broader receptive field to obtain deeper signal information. RESULTS: CSA-MSCNN achieves average accuracies of 95.75 % and 95.39 % for three-class classification of valence and arousal on DEAP, respectively, while achieving an average three-class classification accuracy of 90.48 % on SEED. COMPARISON WITH EXISTING METHODS: The classification accuracy of CSA-MSCNN is not only significantly better than traditional machine learning models, but also shows strong competitiveness compared with mainstream deep learning models such as graph convolutional neural network (GCNN). CONCLUSIONS: CSA-MSCNN addresses the issues of multiple EEG signal channels and complex regional information.