The original EEG data collected are the 1D sequence, which ignores spatial topology information; Feature Pyramid Networks (FPN) is better at small dimension target detection and insufficient feature extraction in the scale transformation than CNN. We propose a method of FPN and Long Short-Term Memory (FPN-LSTM) for EEG feature map-based emotion recognition. According to the spatial arrangement of brain electrodes, the Azimuth Equidistant Projection (AEP) is employed to generate the 2D EEG map, which preserves the spatial topology information; then, the average power, variance power, and standard deviation power of three frequency bands (α, β, and γ) are extracted as the feature data for the EEG feature map. BiCubic interpolation is employed to interpolate the blank pixel among the electrodes; the three frequency bands EEG feature maps are used as the G, R, and B channels to generate EEG feature maps. Then, we put forward the idea of distributing the weight proportion for channels, assign large weight to strong emotion correlation channels (AF3, F3, F7, FC5, and T7), and assign small weight to the others; the proposed FPN-LSTM is used on EEG feature maps for emotion recognition. The experiment results show that the proposed method can achieve Value and Arousal recognition rates of 90.05% and 90.84%, respectively.

ReliefF Matching Feature Selection (RMFS) is proposed in the paper, which can solve the problem of individual specificity and global threshold mismatch of emotion recognition. Firstly, EEG was decomposed into six emotion-related bands by wavelet packet, then EMD was employed for extracting the 10 categories of features of wavelet coefficient and IMF component of the reconstructed signal; Secondly, the optimization formula of the feature group weight was proposed based on feature sets selected by ReliefF, and it can get the weights of different test features, which were the global optimal matching feature group and the corresponding matching channel, so it can eliminate the redundant information and solve the problem of individual specificity. Finally, SVM was employed to identify the test feature group data to obtain emotional recognition results. The experimental results show that the average correct rates of RMFS for two-category of the valence and the arousal are 93.28% and 93.32%, and the four-categories are higher than 83%. The efficiency of the single subject using RMFS is improved by 42.65%, which is better than the traditional ReliefF algorithm.