MCDDT: Mirror Center Loss-Based Dual-Scale Dual-Softmax Transformer for Multisource Subjects Transfer Learning in Motor Imagery Recognition

Accurate recognition of motor imagery (MI)-based electroencephalogram (EEG) signals is crucial for the performance of brain–computer interface (BCI). Given the limited number of EEG signals from a target subject, localizing neural activity in the sensorimotor cortex of the brain and transferring knowledge from source subject data with diverse distributions presented two significant challenges. In this article, we propose a mirror center loss-based dual-scale dual-Softmax transformer (MCDDT) model for multisource subjects transfer learning in MI recognition. Specifically, the mirror center loss is proposed to help the model enhance the localization ability of the neural activity, by minimizing the distance between the features with ipsilateral neural activity and maximizing that with contralateral neural activity. The dual-scale dual-Softmax transformer is introduced to adopt the different distribution of EEG signals from different source subjects, effectively transferring knowledge from these diverse sources. The proposed MCDDT is evaluated on two public datasets and the experimental results demonstrate that MCDDT achieves accuracies of 89. 64% and 90. 96%, exceeding the state-of-the-art models by 2.69% and 2.73%, respectively. Furthermore, the ablation experiments have validated the effectiveness of the dual-scale structure, dual-Softmax mechanism, and mirror center loss, respectively.