Time-Frequency Representation and Convolutional Neural Network-Based Emotion Recognition.

Area of Science:

• Neuroscience
• Computer Science
• Artificial Intelligence

Background:

• Emotions involve complex physiological, cognitive, and behavioral changes.
• Electroencephalogram (EEG) signals offer a noninvasive method for emotion identification.
• Accurate emotion classification is crucial for advancing human-computer interfaces.

Purpose of the Study:

• To propose an automatic feature extraction and classification method for emotions using convolutional neural networks (CNNs).
• To evaluate the performance of different CNN architectures for EEG-based emotion recognition.

Main Methods:

• Filtered EEG signals were transformed into images using the Smoothed pseudo-Wigner-Ville distribution.
• Pretrained models (AlexNet, ResNet50, VGG16) and a configurable CNN were utilized.
• Performance was assessed using accuracy, precision, F1-score, and other metrics.

Main Results:

• The configurable CNN demonstrated superior performance with fewer learning parameters.
• Accuracy scores reached 90.98% (AlexNet), 91.91% (ResNet50), 92.71% (VGG16), and 93.01% (configurable CNN).
• The proposed method using configurable CNN outperformed existing approaches.

Conclusions:

• The developed method effectively automates feature extraction and emotion classification from EEG data.
• Configurable CNNs present a promising approach for accurate and efficient emotion recognition.
• This research contributes to the development of more sophisticated human-computer interaction systems.