CMLP-Net: A convolution-multilayer perceptron network for EEG-based emotion recognition

Deep learning has demonstrated remarkable performance in emotion recognition tasks based on Electroencephalogram (EEG) signals. The utilization of the spatial and temporal features of EEG signals plays a pivotal role in the task of emotion recognition. Traditional Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Transformer models have encountered challenges in effectively capturing both the temporal and spatial characteristics of EEG signals. Therefore, achieving a comprehensive integration of temporal and spatial features remains a demanding task. In this paper, we propose a Convolution-Multilayer Perceptron Network (CMLP-Net) based on the 2D representation of EEG signals. CMLP-Net comprises a temporal-stream shared convolution, a time-refinement temporal-spatial convolution, and a spatial interaction Multilayer Perceptron (MLP). The temporal-stream shared convolution is employed to uncover shared features across multiple consecutive temporal windows. The time-refinement temporal-spatial convolution is utilized to capture effective temporal-spatial features effectively. Lastly, the spatial interaction MLP facilitates spatial interactions among subregions within the EEG feature maps, thereby enhancing the global spatial dependency of the features. The proposed method was evaluated on the popular DEAP dataset, where 14 channels relevant to EEG emotions were selected. Experimental results validate that the proposed approach exhibits state-of-the-art performance in binary emotion recognition tasks, achieving average accuracies of 98.65%, 98.70%, and 98.63% in valence, arousal, and dominance dimensions, respectively. Furthermore, CMLP-Net also demonstrates excellent capabilities in handling multi-state emotion recognition tasks.