Table 1. Parameters of each layer in CNN-GMM.

Layer name	Number of filters	Size of feature map	Size of kernel	Stride
1. Grouped convolution 1	3 groups with 4 filters in each	222 * 222 * 12	[3 3]	[1 1]
ReLU + Batch normalization
2. Grouped convolution 2	12 groups with 4 filters in each	220 * 220 * 48	[3 3]	[1 1]
ReLU + Batch normalization
Max pooling 1	–	108 * 108 * 48	[5 5]	[2 2]
3. Grouped convolution 3	48 groups with 4 filters in each	106 * 106 * 192	[3 3]	[1 1]
Relu + Batch normalization
4. Grouped convolution 4	192 groups with 4 filters in each	104 * 104 * 768	[3 3]	[1 1]
ReLU + Batch normalization
Max pooling 2	–	51 * 51 * 768	[3 3]	[2 2]
5. Grouped convolution 5	786 groups with 2 filters in each	49 * 49 * 1,536	[3 3]	[1 1]
ReLU + Batch normalization
6. Grouped convolution 6	1,536 groups with 2 filters in each	47 * 47 * 3,072	[3 3]	[1 1]
ReLU + Batch normalization
Max pooling 3	–	23 * 23 * 3,072	[3 3]	[2 2]
7. Grouped convolution 7	3,072 groups with 1 filter in each	21 * 21 * 3,072	[3 3]	[1 1]
ReLU + Batch normalization
8. Grouped convolution 8	3,072 groups with 1 filter in each	19 * 19 * 3,072	[3 3]	[1 1]
ReLU + Batch normalization
Max pooling 4	–	9 * 9 * 3,072	[3 3]	[2 2]
9. Fully connected layer	–	1 * 400	–	–
Dropout (0.4) + Batch normalization
10. GMM fully connected layer	–	1 * 400	–	–
Batch normalization
11. GMM fully connected layer	–	1 * 250	–	–
Batch normalization
12. GMM fully connected layer	–	1 * 2	–	–
SoftMax layer
Classification layer