Attention-Based DSC-ConvLSTM for Multiclass Motor Imagery Classification
Table 5
Evaluated deep learning structure.
Design structure
Our choice
Aim
Average accuracy
bciv iv 2a (%)
HGD (%)
Convolution in first layer
Splitted convolution
The first layer of convolution is divided into time-domain convolution and spatial filtering, which can better process the input of EEG signal and improve the classification accuracy.
60.2
82.3
ConvNet
Separable convolution
One is to reduce the number of network parameters and improve the training speed; the other is to show the relationships within and across decoupled feature maps
60.8
82.9
LSTM
BiConvLSTM
Improves LSTM’s disadvantage of extracting only temporal features of EEG signals and enables it to extract spatial features of EEG signals
65.3
84.6
Data processing
Sliding window
It not only increases the number of training samples but also fully extracts the differential features and global features of all EEG data
