Emotion recognition based on 3D matrices and two-way densely connected network

In order to fully consider the time, space, frequency features of EEG signals, a new model CA-ResNet-DenseNet for feature extraction and recognition of EEG signals is proposed. Firstly, the differential entropy features of different frequency bands of EEG signals in each channel after time-frequency segmentation were extracted, which were transformed into three-dimensional feature matrices. Secondly, the frequency and spatial information of the EEG signal was mainly extracted by CA-ResNet, and then the weights of the relevant frequency band were reassigned. Finally, the improved densely connected network was used for secondary feature utilization of frequency and time information, and makes full use of multi-source information. The SEED and SEED-IV dataset were used to validated the effectiveness of the algorithm. On SEED dataset, the model achieved the best performance, with accuracy, kappa value, precision, and F1-score reaching 95.23%, 0.94, 0.96, and 0.96, respectively. The results of 85.15% accuracy and kappa value of 0.80 were obtained on SEED-IV dataset, which achieve the highest accuracy compared with existing baseline models. The model can fully decode the EEG signals, extract key emotional features, and improve the accuracy of emotion classification. The algorithm has important significance for the development of more efficient interactive emotion recognition systems.