CFBC: A Network for EEG Emotion Recognition by Selecting the Information of Crucial Frequency Bands

• Published in: IEEE sensors journal Vol. 24; no. 19; pp. 30451 - 30461
• Main Authors: Zhu, Mu, Bai, Zhongli, Wu, Qingzhou, Wang, Junchi, Xu, Wenhui, Song, Yu, Gao, Qiang
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.10.2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Band theory; Brain modeling; Convolution; Convolutional neural networks; Crucial frequency band selector; Electrodes; Electroencephalography; electroencephalography (EEG); Emotion recognition; Emotions; Feature extraction; Feature maps; Feature recognition; Frequencies; Frequency conversion; Frequency domain analysis; Frequency ranges; Position (location); Power spectral density; Spatial data; spatial feature extractor; spatial-frequency feature
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2024.3440340

In this work, inspired by the frequency band division theory of electroencephalogram (EEG) signals, we propose the crucial frequency band convolution (CFBC) network method to explore the crucial frequency range of frequency domain features. CFBC includes two parts: spatial feature extractor and crucial frequency band selector. First, we extracted differential entropy (DE) and power spectral density (PSD) features from the frequency domain of each EEG channel by different frequency bands. To avoid the loss of effective spatial information in processing, we interpolate the EEG feature maps by the relative location of electrodes. The spatial feature extractor captures spatial information using channel-by-channel convolution with the 2-D EEG feature maps containing electrode position information. The crucial frequency band selector performs causal dilated convolution for the selected frequency band sequence so that the features contained in different frequency bands are stacked into a feature vector. Finally, CFBC realizes the purpose of combining multiple selected frequency bands to extract the cross-frequency band feature vector. To evaluate the proposed method, we conducted a subject-dependent EEG emotion recognition experiment in the SEED dataset. The experimental result shows that the selection of a frequency band has an impact on the emotion classification effect of frequency domain features.