Table 1.

Layer-wise configuration details of the proposed ensemble-SDCNN^a model.^b

Layer name		Output sizec	Filter sized	Iterations	Parameters
DCNNe model
	Input	(224,224,3)	N/Af	—g	—
	Conv1	(112,112,64)	(7,7,64)	1	9600
	Max pooling	(56,56,64)	(3,3)	1	—
	IMh-based RU1i	(56,56,64)	(3,3,64)	2	74,112
	IM-based RU2	(56,56,64)	(3,3,64)	2	74,112
	CMj-based RU3	(28,28,128)	(3,3,128); (1,1,128)	2; 1	230,528
	IM-based RU4	(28,28,128)	(3,3,128)	2	295,680
	CM-based RU5	(14,14,256)	(3,3,256); (1,1,256)	2; 1	919,808
	IM-based RU6	(14,14,256)	(3,3,256)	2	1,181,184
	CM-based RU7	(7,7,512)	(3,3,512); (1,1,512)	2; 1	3,674,624
	IM-based RU8	(7,7,512)	(3,3,512)	2	4,721,664
	Avg pooling	(1,1,512)	(7,7)	1	—
SCNNk model
	Conv1	(112,112,128)	(7,7,128)	1	19,200
	Conv2	(35,35,64)	(5,5,64)	1	204,992
	FC1	(1,1,32)	(5,5,64)	1	2,508,832
	Depth concat	(1,1,544)	—	1	—
	FC2	(1,1,2)	—	1	1090
	SoftMax	(1,1,2)	—	1	—
	Classification	2	—	1	—

^aSDCNN: shallow–deep CNN.

^bTotal learnable parameters: 13,915,426.

^cOutput size (image width, image height, # of channels),

^dKernel size (kernel width, kernel height, # of filters), Max pooling (kernel width, kernel height), Avg pooling (kernel width, kernel height).

^eDCNN: deep CNN.

^fN/A: not applicable.

^g—: not available.

^hIM: identity mapping.

ⁱRU: residual unit.

^jCM: convolutional mapping.

^kSCNN: shallow CNN.