FB-CGANet: filter bank channel group attention network for multi-class motor imagery classification

Abstract

Objective. Motor imagery-based brain computer interface (MI-BCI) is one of the most important BCI paradigms and can identify the target limb of subjects from the feature of MI-based Electroencephalography (EEG) signals. Deep learning methods, especially lightweight neural networks, provide an efficient technique for MI decoding, but the performance of lightweight neural networks is still limited and need further improving. This paper aimed to design a novel lightweight neural network for improving the performance of multi-class MI decoding. Approach. A hybrid filter bank structure that can extract information in both time and frequency domain was proposed and combined with a novel channel attention method Channel Group Attention (CGA) to build a lightweight neural network Filter Bank Channel Group Attention Network (FB-CGANet). Accompanied with FB-CGANet, the Band Exchange data augmentation method was proposed to generate training data for networks with filter bank structure. Main results. The proposed method can achieve higher 4-class average accuracy (79.4%) than compared methods on the BCI Competition IV IIa dataset in the experiment on the unseen evaluation data. Also, higher average accuracy (93.5%) than compared methods can be obtained in the cross-validation experiment. Significance. This work implies the effectiveness of channel attention and filter bank structure in lightweight neural networks and provides a novel option for multi-class motor imagery classification.

IFNet: An Interactive Frequency Convolutional Neural Network for Enhancing Motor Imagery Decoding From EEG

“…Although it has shown success in image recognition, it does not transfer well in MI decoding. Further investigation on attention mechanism might help facilitate EEG feature representation [24]. In conclusion, the summation operator deals with cross-frequency interactions effectively and efficiently.…”

“…EEGNet learns frequency filters entirely by backpropagation, which omits hand-crafted spectral bands that might be beneficial for mining mutual spectral information. More recently, channel group attention was introduced in [24] to deal with filter bank inputs. Different from the perspective of band interaction, their inter-channel attention targets scaling the feature channels to improve the expression ability of representative features in both bands.…”

Filter bank sinc-convolutional network with channel self-attention for high performance motor imagery decoding

“…Our implementation of FB-Sinc-ShallowNet was based on that of Sinc-ShallowNet, and the structure was modified according to the original work, with similar performance obtained. • FB-CGANet [15]: In FB-CGANet, a CGA method was proposed for efficient feature integration, and a dual-branch lightweight network was designed with a hybrid filter bank structure in both the time and frequency domain. Since this work described the details of the network structure, we implemented it in the PyTorch framework, and comparable performance was found on the BCIC IV IIa dataset.…”

“…The implementation of SE, ECA, and SA was based on the PyTorch version of SE 7 , the project created by Wang et al 8 , and codes released by Zhang et al 9 , respectively. CGA was implemented with the PyTorch framework according to the report of Chen et al [15]. In SA and CGA, the number of channel groups was set as 4 to group features by corresponding frequency bands for allocating attention via information of four motor rhythms.…”

“…Li et al proposed the Temporal-Spectral-based Squeeze-and-Excitation Feature Fusion Network that introduced SE blocks for solving the redundant information problem in feature fusion and amplifying the discriminative features [14]. Chen et al designed the Filter Bank Channel Group Attention Network (FB-CGANet) that has a hybrid filter bank structure in both time and spatial domain, and they proposed the channel group attention (CGA) method for feature selection based on local information of features corresponding to one frequency band [15]. Currently, channel attention methods in MI decoding can compute attention values based on either global or local information.…”