An efficient shallow convolutional decoding network for motor imagery EEG signals

To address the problems of poor temporal-spatial-frequency coupling feature learning and lengthy model training and inference of deep learning-based electroencephalography signal decoding networks (EEGNet ) in existing motor imagery-based brain-computer interfaces (MI-BCIs), an efficient shallow convolutional decoding network called Faster-EEGNet was proposed. In this network, the first layer of two-dimensional serial convolution was optimized to perform parallel convolution on all channels simultaneously. This enabled the completion of temporal fil¬tering and spatial filtering of signals across all channels. Temporal convolutional features were captured from spatial pattern-extracted signals at the intermediate deep convolutional layers. Then, the depthwise separable convolution was used to capture the temporal-spatial coupling fea¬tures of the signals for pattern recognition. Experimental validation was conducted using publicly available datasets. The results demonstrate that the Faster-EEGNet exhibits better performance than the EEGNet in motor imagery recognition accuracy and information transfer rate. This net¬work also achieves good recognition results in small-sample training scenarios. Furthermore, the Faster-EEGNet reduces training lime by 44. 8% and model inference time by more than 43. 6% compared to the EEGNet. These findings demonstrate that the proposed Faster-EEGNet can enhance the recognition accuracy, convenience, and rapid response performance of the motor imagery brain-computer interface system. © 2023 Xi'an Jiaotong University. All rights reserved.