An Epilepsy Detection Method Using Multiview Clustering Algorithm and Deep Features

Table 2

Network structure of deep feature extraction from various views.

Parameter/feature	Deep time domain feature	Deep spectrum feature	Deep time-frequency feature
Layer 1 (input layer)	Original matrix: 23256; Convolution kernel: 1128; Step size: 1	Original matrix: 2327; Convolution kernel: 44; Step size: 1	Original matrix: 2562314; Convolution kernel: 12911; Step size: 1
Layer 2 (convolutional layer)	Feature map: 1@23129; Convolution kernel: 165; Step size: 30	Feature map: 20@2024; Convolution kernel: 88; Step size: 1	Feature map: 1@1282324; Convolution kernel: 6544; Step size: 1
Layer 3 (convolutional layer)	Feature map: 30@236; Convolution kernel: 433; Step size: 1	Feature map: 10@1317; Convolution kernel: 88; Step size: 1	Feature map: 30@642411; Convolution kernel: 3044; Step size: 1
Layer 4 (convolutional layer)	Feature map: 20@2033; Convolution kernel: 818; Step size: 1	—	Feature map: 20@32178; Convolution kernel: 1781; Step size: 1
Layer 5 (convolutional layer)	Feature map: 10@1316	—	Feature map: 10@16108
Layer 6 (fully connected layer)	The 10 feature maps of size 1316 output from the convolutional layer of the fifth layer are converted into a vector of size 12080, which is used as the input of the fully connected layer.; Feature map: 1@1102	The 10 feature maps of size 1317 output by the convolutional layer of the third layer are converted into a vector of size 12210, which is used as the input of the fully connected layer.; Feature map: 1@1512	The 10 feature maps of size 16108 output by the convolutional layer of the fifth layer are converted into a vector of size 112800, which is used as the input of the fully connected layer.; Feature map: 1@12048
Layer 7 (fully connected layer)	Feature map: 1@1100	Feature map: 1@1100	Feature map: 1@11024
Layer 8 (fully connected layer)	Output: 11024 vector	Output: 1512 vector	Feature map: 1@1100
Layer 9 (fully connected layer)	—	—	Output: 12048 vector