Emotion recognition with attention mechanism-guided dual-feature multi-path interaction network

Electroencephalography (EEG)-based emotion recognition has gained widespread attention recently. Although many deep learning methods have been proposed, it is still challenging to simultaneously fuse information in the time–frequency–spatial domain. This paper proposes an attention mechanism-guided dual-feature multi-path interaction network (ADMIN) for emotion recognition. Firstly, the original EEG signal is divided into five frequency bands, and then, differential entropy features and zero crossing rate features are extracted separately. These features are further converted into a two-dimensional (2D) expanded map based on the position of electrodes in the brain area, which is beneficial for the exploration of spatial-frequency information. Next, we transform the generated 2D expanded maps into a 3D cuboid that can simultaneously integrate temporal, frequency, and spatial information as input. Subsequently, the 3D cuboid is processed through convolutional neural network to extract frequency and spatial information from the EEG signals. By embedding the attention mechanism combined with gated recurrent unit, the temporal information can be extracted, which not only solves the problem of long-term dependencies but also strengthens the utilization of key temporal information. We conduct extensive experiments on SEED and DEAP datasets, and the results show that the ADMIN model achieves state-of-the-art performance.