CLA-Net: A Deep Spatio-Temporal Attention Network Based on ConvLSTM for EEG Emotion Recognition

With the rapid development of artificial intelligence and brain-computer interface technologies in recent years, emotion recognition based on Electro-encephalogram (EEG) data has become widely used. In order to simultaneously extract the temporal and spatial features of EEG signals, we propose a deep spatio-temporal attention network based on convolutional long short-term memory network (ConvLSTM) for EEG emotion recognition named CLA-Net. Firstly, CLA-Net maps one-dimensional EEG signals into two-dimensional vectors according to the spatial position of EEG electrodes. The two-dimensional vectors are arranged on the time axis to obtain spatio-temporal sequences with spatial features. Then, ConvLSTM blocks with residual structure are used to extract local emotional features of different depths and scales. The residual learning strategy is convenient to integrate spatio-temporal cues in this local feature. Finally, we send the extracted new spatio-temporal feature sequence into the self-attention block to extract the global feature weight, and use softmax to classify emotion. Extensive tests in SEED and SEED-IV were carried out in this work, and good classification results have been achieved.