Simplified 2D CNN Architecture With Channel Selection for Emotion Recognition Using EEG Spectrogram

• Published in: IEEE access Vol. 11; pp. 46330 - 46343
• Main Authors: Farokhah, Lia, Sarno, Riyanarto, Fatichah, Chastine
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Arousal; Brain modeling; Channels; Classification; Complexity; Computer architecture; Convolutional neural networks; Electroencephalography; Emotion recognition; Emotions; Fourier transforms; inter-subject; Recording; Spectrogram; Training; Two dimensional models
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2023.3275565

Emotion Recognition through electroencephalography (EEG) is one of the prevailing emotion recognition techniques achieving higher accuracy rates. Nevertheless, one of the problems is the emotion recognition for inter-subjects where accuracy measures are lower. This happens because the EEG is a non-stationary signal which is resulting a domain shift across recordings of subjects, even under the same emotions, thus making emotional patterns difficult to identify. Another common observation is the emotion recognition of inter-subject and intra-subject by using all channels originating from the standard EEG recording mechanism. This requires higher computational resources and deep networks which are more complex including DenseNet, ResNet, etc. In this paper, we propose a novelty emotion recognition classification model that offers a simplified structure and employs only the selected channels from the 32 recorded channels. Residing on a standard approach of transforming EEG data- Database for Emotion Analysis of Physiological Signals (DEAP)- into 2D images using Short-Time Fourier Transform (STFT) EEG signals for training and analysis. Follows the channel selection approach and makes use of a simplified 2D CNN model. The channel selection adopts a search and retention approach using the selected samples of the data. The experimental results show that the performance of the proposed architecture improves the accuracy of the inter-subject emotion recognition using 32 channels by 9.73% and 11.7% on valence and arousal, respectively. While the use of the proposed selected channel method, with only 10 channels, performance increased by 3.53% and 7.2% on valence and arousal classes, respectively. Thus, by keeping a lower complexity level, the proposed architecture attains higher performance rates.