A Dense Convolutional Bi-Mamba Framework for EEG-Based Emotion Recognition

In recent times, emotion recognition based on electroencephalograms (EEGs) has found extensive applications. Although numerous approaches leveraging CNN and Transformer have been put forward for automatic emotion recognition and have achieved commendable performance, several challenges remain: (1) Transformer-based models are proficient at capturing long-term dependencies within EEG signals. However, their quadratic computational complexity poses a significant hurdle. (2) Models that combine Transformers with convolutional neural networks (CNNs) often fail to effectively capture the coarse-to-fine temporal dynamics of EEG signals. State Space Models (SSMs), exemplified by Mamba, have emerged as a promising solution. They not only showcase outstanding capabilities in modeling long-range interactions but also maintain a linear computational complexity, which is highly advantageous. To address these challenges head-on, we introduce Emotion-Mamba, an innovative framework designed specifically for EEG-based emotion recognition. The proposed framework initiates the process by employing the CNN Encoder to extract information from both the temporal and spatial dimensions of EEG signals. Subsequently, the extracted feature information is relayed to the Hierarchical Coarse-to-Fine Bi-Mamba (HBM) block, which is adept at efficiently processing these features. Furthermore, a Dense Temporal Fusion (DTF) module has been incorporated. This module capitalizes on the multi-level, purified temporal information sourced from CNN Encoder and HBM blocks, with the aim of bolstering decoding accuracy. We conduct comprehensive evaluations of Emotion-Mamba using the SEED and SEED-V datasets. The experimental findings unequivocally demonstrate that our proposed approach surpasses the existing state-of-the-art methods.