Multi-branch heterogeneous network of exploiting complementary multi-view features for decoding finger motor imagery EEG.

In contrast to general motor imagery involving large body parts, research on finger motor imagery is very scarce. Due to more refined motor functions, the decoding of finger motor imagery is more arduous and has lower accuracy than that of general motor imagery. In order to improve the decoding accuracy of finger motor imagery, this paper proposes the problem of identifying complementary multi-view decoding features and the problem of electrode channel difference of convolution kernels. A novel multi-branch heterogeneous network (MBHN) consisting of three groups of diversified branches is proposed to effectively extract and exploit complementary features of raw-view, frequency-decomposition-view and wavelet-view. Moreover, the channel adaptive kernel (CAK) module is proposed as a solution to the problem of channel difference of convolution kernels. The experimental results on the public finger motor imagery dataset show that our MBHN model achieves the state-of-the-art decoding accuracy of 58.49%. Additionally, the integration of the auxiliary supervision mechanism and the hybrid loss function is a very effective approach to fully leverage the complementarity of multi-view deep features. Our code is publicly available at https://github.com/ykhdu/MBHN.