MDRNet: Multi-Branch with Different Feature Representations Network for Motor Imagery Classification

A brain-computer interface (BCI) offers an innovative solution for facilitating communication and control in individuals with paralysis. BCI reflects brain activity by decoding electroencephalogram (EEG) signals. Despite numerous techniques for classifying motor imagery (MI) EEG signals, challenges such as low signal-to-noise ratio (SNR), nonlinearity, non-stationarity, and feature information loss continue to persist. This paper proposes an innovative Multi-Branch with Different Feature Representations Network (MDRNet) to address these issues. Firstly, we introduce the Temporal Gramian Angular Summation Field (T-GASF) method, which enhances signal features with temporal information, thereby improving SNR. Secondly, we present a hybrid attention mechanism based on empirical mode decomposition (EMD), combining channel attention and intrinsic mode function (IMF) attention to effectively address signal nonlinearity and non-stationarity. Lastly, we develop a U-net branch with T-GASF input, a Channel-IMF Hybrid Attention (CIHA) branch based on EMD-decomposed signals, and a Transformer Encoder-CNN branch. These branches ensure comprehensive extraction of critical EEG features. Experimental findings from the BCI Competition IV 2a and 2b datasets indicate that MDRNet attained classification accuracies of 95.71% and 90.28%, respectively, markedly outperforming current methods.