Improving Motor Imagery Intention Recognition via Local Relation Networks.

Brain-computer interface (BCI) is a new communication and control technology established between human or animal brains and computer or other electronic equipment that does not rely on conventional brain information output pathways. The non-invasive BCI technology collects EEG signals from the cerebral cortex through signal acquisition equipment and processes them into signals recognized by the computer. The signals are preprocessed to extract signal features used for pattern recognition and finally are transformed into specific instructions for controlling external types of equipment. Therefore, the robustness of EEG signal representation is essential for intention recognition. Herein, we convert EEG signals into the image sequence and utilize the Local Relation Networks model to extract high-resolution feature information and demonstrate its advantages in the motor imagery (MI) classification task. The proposed method, MIIRvLR-Net, can effectively eliminate signal noise and improve the signal-to-noise ratio to reduce information loss. Extensive experiments using publicly available EEG datasets have proved that the proposed method achieved better performance than the state-of-the-art methods.