A compact multi-branch 1D convolutional neural network for EEG-based motor imagery classification

•We design a simple and effective model to solve the recognition based on motor imagery tasks. The model is able to fully learn the event-related desynchronization/synchronization (ERD/ERS) phenomenon in the raw MI EEG signal to accurately classify the motor intention.•The model structure is simple and only requires a simple knowledge of deep learning to reproduce the model.•Data augmentation methods using data interpolation and clipping.•Analyze the network using a variety of visual methods.•The average classification accuracy of 83.92% and 87.19% was achieved on the two public datasets, respectively. Motor imagery (MI) EEG signals are considered a promising paradigm for BCI systems that enable humans to communicate with the outside world through the brain and have a wide range of applications to improve patients' quality of life with muscle or nerve damage. Due to the low signal-to-noise ratio of the acquired EEG signals, it is challenging to decode the intent accurately and even more challenging to decode the raw EEG signals. Currently, there is no deep learning method to achieve high classification performance in decoding raw EEG signals. We propose a new end-to-end network for decoding MI EEG signals, Compact Multi-Branch One-dimensional Convolutional Neural Network (CMO-CNN), without some pre-processing such as filtering, using the original EEG signals. The 1D convolution is used as the feature extractor to extract diverse and multi-level features for fusion using different filter scales and depths of different branches. 1D Squeeze-and-Excitation blocks (SE-blocks) and shortcut connections are added to further improve the generalization and robustness of the network. 83.92% and 87.19% classification accuracies were achieved in the BCI Competition IV-2a and the BCI Competition IV-2b datasets. An 8% improvement to 63.34% was achieved in the cross-subject test, demonstrating that our proposed CMO-CNN outperforms the current state-of-the-art methods. Visual analysis of the network shows that the proposed model can accurately learn the event-related desynchronization/synchronization (ERD/ERS) phenomenon in the signal, and 1D convolution is actively used for feature extraction suitable for feature extraction of the original EEG signal.