EEG emotion recognition approach using multi-scale convolution and feature fusion

Electroencephalogram (EEG) signal has been widely applied in emotion recognition due to its objectivity and reflection of an individual's actual emotional state. However, current EEG-based emotion recognition methods still suffer from limitations such as single-feature extraction, missing local features, and low feature extraction rates, all of which affect emotion recognition accuracy. To address these challenges, we propose an EEG emotion recognition approach based on multi-scale convolution and feature fusion, aimed at integrating low-dimensional and high-dimensional EEG features to enhance classification performance. The proposed approach first extracts multi-scale features of EEG using a multi-scale convolutional neural network. Subsequently, it employs the convolutional block attention module to further refine the obtained multi-scale features, highlighting emotion-related high-quality features while suppressing unimportant features. Finally, emotion-related time series are extracted by long short-term memory (LSTM) networks to achieve time-space-frequency feature fusion of EEG. We conducted fivefold cross-validation on the DEAP and SEED datasets to demonstrate the effectiveness of the proposed approach. For the DEAP dataset, the proposed approach achieved average accuracies of 97.48% and 97.75% for valence and arousal dimensions, respectively, and 96.98% for the SEED dataset, outperforming other models in terms of recognition accuracy.