Feature Pyramid Networks and Long Short-Term Memory for EEG Feature Map-Based Emotion Recognition.

Area of Science:

• Neuroscience
• Computer Science
• Machine Learning

Background:

• Electroencephalography (EEG) data is often processed as 1D sequences, neglecting crucial spatial topology.
• Traditional Convolutional Neural Networks (CNNs) struggle with feature extraction during scale transformations and small target detection compared to Feature Pyramid Networks (FPN).

Purpose of the Study:

• To develop an advanced method for emotion recognition using EEG feature maps.
• To improve the extraction and utilization of spatial and temporal features from EEG data for more accurate emotion detection.

Main Methods:

• Proposed a novel Feature Pyramid Network (FPN) and Long Short-Term Memory (FPN-LSTM) model for EEG-based emotion recognition.
• Utilized Azimuth Equidistant Projection (AEP) to create 2D EEG maps, preserving electrode spatial topology.
• Extracted features (average power, variance, standard deviation) from alpha, beta, and gamma bands, generating RGB EEG feature maps.
• Implemented channel weight proportion distribution, emphasizing electrodes with higher emotion correlation.
• Applied BiCubic interpolation to handle missing pixel data.

Main Results:

• The FPN-LSTM model achieved high recognition rates for emotional dimensions.
• Achieved a Value recognition rate of 90.05%.
• Achieved an Arousal recognition rate of 90.84%.

Conclusions:

• The proposed FPN-LSTM method effectively utilizes spatial topology information from EEG data for emotion recognition.
• The integration of FPN and LSTM, along with strategic feature extraction and channel weighting, significantly enhances recognition accuracy.
• This method offers a promising advancement in EEG-based affective computing.