Exécution de notre outil de synthèse sur un nouvel article
Vérifiez l’article ci-dessous, puis lancez sa synthèse.
Multi-Scale Attention-Enhanced BiLSTM Network for EEG Emotion Recognition
Yongbin Zhang (Beijing Institute of Graphic Communication), Runyi Li (Beijing Institute of Graphic Communication), Xiuli Fu (Beijing Institute of Petrochemical Technology), Meijuan Cao (Beijing Institute of Graphic Communication)
Résumé : Over the last decade, significant progress has been made in detecting human emotions via electroencephalographic (EEG) recordings. Nevertheless, contemporary EEG-driven affect recognition techniques continue to encounter multiple hurdles, including insufficient feature extraction capability, inadequate modeling of local patterns, low feature extraction efficiency, high model complexity, and difficulty in practical deployment. To address these issues, we introduce an EEG-specific lightweight architecture for affective computing titled multi-scale Bidirectional Long Short-Term Memory with Attention (MS-BiLSTM-Attention). This model uses DE features, which are embedded into the EEG electrode topology to preserve spatial information. By combining a multi-scale convolutional neural network (CNN), a Bidirectional Long Short-Term Memory (BiLSTM), and a channel–spatial attention mechanism, the proposed framework jointly extracts and integrates spectral, topology-indexed, and channel-wise information from EEG signals. The main architectural contributions of this work are as follows: (i) dividing the EEG signal into 0.5-second time windows, extracting DE features across multiple frequency bands, and mapping each segment onto the EEG electrode topology to construct segment-level topology representations for subsequent multi-scale feature encoding; (ii) developing an EEG-specific Parallel Multi-scale Spectral–Topological Bidirectional Encoder for multi-scale spectral and topology-indexed feature extraction; (iii) proposing a lightweight Convolutional Additive Attention Module (CAAM) with a Convolutional Additive Token Mixer (CATM) to adaptively enhance channel-wise and spatial feature representations with low computational overhead. Independent experiments were conducted on the DEAP and MAHNOB-HCI datasets under both subject-dependent and subject-independent protocols. Across DEAP and MAHNOB-HCI, the model achieved subject-dependent arousal and valence accuracies above 92%, while using 25.52 K parameters and 0.6 MFLOPs per sample, establishing a favorable accuracy–complexity profile for EEG affective computing.