An electroencephalogram-based emotion recognition method using multi-branch convolutional neural networks and Transformer

29/06/2026

AI Summary

Proposed MCT uses multi-scale CNN and parallel CNN-Transformer to capture short-term variations and long-term dependencies in EEG temporal features.
Dual-branch convolution learns global and local spatial channel features, with CBAM fusing spatio-temporal features for improved discriminative representation.
MCT outperforms baselines, achieving 83.83% on SEED, 90.00% arousal and 92.62% valence on MEEG, and about 61% on DEAP.

TL;DR (AI Generated)

To address the insufficient modeling of long-term dependencies in EEG signals, this paper proposes an EEG emotion recognition method based on multi-branch convolutional neural networks (CNN) and Transformer (MCT).

The proposed method employs a multi-scale CNN to extract local temporal features from EEG signals and constructs a parallel CNN-Transformer architecture to capture both short-term variations and long-term dependencies, thereby enabling temporal feature modeling at different time scales.

These findings indicate that MCT can effectively learn discriminative features associated with emotional states, providing a new perspective for EEG-based emotion recognition research.

Basic summary

Sheng Wu Yi Xue Gong Cheng Xue Za Zhi. 2026 Jun 25;43(3):513-520. doi: 10.7507/1001-5515.202602031.

ABSTRACT

Electroencephalogram (EEG)-based emotion recognition is an important research area in affective computing and mental health assessment. To address the insufficient modeling of long-term dependencies in EEG signals, this paper proposes an EEG emotion recognition method based on multi-branch convolutional neural networks (CNN) and Transformer (MCT). The proposed method employs a multi-scale CNN to extract local temporal features from EEG signals and constructs a parallel CNN-Transformer architecture to capture both short-term variations and long-term dependencies, thereby enabling temporal feature modeling at different time scales. Furthermore, a dual-branch convolutional structure is utilized to learn both global and local spatial channel features of EEG signals. A convolutional block attention module (CBAM) is then introduced to fuse the spatio-temporal features of EEG signals. Experimental results show that the proposed MCT model achieves a classification accuracy of 83.83% on the Shanghai Jiao Tong University emotion EEG dataset (SEED). On the music emotion EEG dataset (MEEG), it attains accuracies of 90.00% and 92.62% for the arousal and valence dimensions, respectively. On the database for emotion analysis using physiological signals (DEAP), it attains accuracies of 61.30% and 61.04% for the arousal and valence dimensions. The accuracy results on all three datasets outperform those of the best-performing baseline models. These findings indicate that MCT can effectively learn discriminative features associated with emotional states, providing a new perspective for EEG-based emotion recognition research.

PMID:42366434 | DOI:10.7507/1001-5515.202602031

Document this CPD