A novel 3D feature fusion network for EEG emotion recognition

• Published in: Biomedical signal processing and control Vol. 102; p. 107347
• Main Authors: Chen, Kun, Chai, Shulong, Cai, Mincheng, Liu, Quan, Ai, Qingsong, Zhou, Chiming, Ma, Li
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.04.2025
• Subjects: 3D feature fusion; Attention mechanism; EEG; Emotion recognition; Selective kernel convolution; Emotion recognition; Attention mechanism; Selective kernel convolution; 3D feature fusion; EEG
• ISSN: 1746-8094
• DOI: 10.1016/j.bspc.2024.107347

Emotion recognition through EEG signals occupies a pivotal role in the domain of brain-computer interfaces. However, the effective integration of features remains a challenge. In this paper, we propose a novel three-dimensional feature fusion network. This network is designed to delve deeper into the intricacies of 3D features, extracting latent features, and subsequently performing a comprehensive fusion of these potential features. First, we extract differential entropy (DE) and power spectral density (PSD) from raw EEG signals. Then, we independently construct 3D features from these sets. By integrating a convolutional neural network (CNN) with an attention mechanism, we enhance the understanding of 3D features, deriving latent features. Finally, we employ a feature fusion module using Selective Kernel Convolution’s fuse and select modules. On the DEAP dataset, our model achieved promising results in both the 10-fold cross-validation task (84.93% arousal, 83.54% valence) and subject-dependent experiments (87.02% arousal, 86.12% valence). On the SEED dataset, our model achieved an impressive average accuracy (96.64%). These experimental outcomes underscore the capability of our model to adeptly fuse disparate 3D features, yielding more effective fusion features and consequently enhancing emotion recognition performance.