Multi-Scale Convolutional Attention and Riemannian Geometry Network for EEG-Based Motor Imagery Classification

Abstract

The electroencephalogram (EEG) is a non-invasive technique with high temporal resolution that has become the research frontier of brain-computer interface (BCI) systems. It is widely used in medical rehabilitation, gaming, and other industries. However, decoding EEG signals remains a challenging task. A network called MSCARNet, which combines multi-scale convolution and Riemannian geometry, was proposed for classifying motor imagery based on EEG. The network is supplemented by an attention mechanism and sliding window technique. The MSCARNet utilizes sliding windows to expand data dimensions and multiple convolution kernels to obtain spatial and temporal features. These features are then mapped to Riemannian space and undergo bilinear mapping and logarithmic operations for dimensionality reduction. This approach is beneficial in reducing the impact of noise and outliers and provides convenience for classification. Subject-dependent and subject-independent experiments were conducted using the BCI-IV-2a dataset to validate the effectiveness of the MSCARNet. The results show that the accuracy improved by approximately 4% compared to existing state-of-the-art methods. The hybrid network based on Riemannian space can effectively improve the accuracy of EEG motor imagery classification tasks without excessive preprocessing.

Keywords

• Electroencephalogram
• deep learning
• convolution neural network
• motor imagery
• Riemannian geometry