Table 1.

CNN architectures.

CNN1		CNN2		CNN3
Layers	Parameters	Layers	Parameters	Layers	Parameters
Input image	$100\times 100$	Input image	$100\times 100$	Left branch:
Conv 1	$64\times 5\times 5$, pad 0, stride 1	Conv 1	$64\times 5\times 5$, pad 0, stride 1	Input image	$100\times 100$
				Max pool 1	$10\times 10$
Leaky ReLU 1	Alpha = 0.01	Leaky ReLU 1	Alpha = 0.01	Dropout	0.1
Max pool 1	$3\times 3$, stride 3, pad 0	Max pool 1	$3\times 3$, stride 3, pad 0	Right branch: Input image	$100\times 100$
Conv 2	$64\times 2\times 2$, pad 0, stride 1	Conv 2	$64\times 2\times 2$, pad 0, stride 1	Conv 1	$64\times 5\times 5$, pad 0, stride 1
Leaky ReLU 2	Alpha = 0.01	Leaky ReLU 2	Alpha = 0.01	Leaky ReLU 1	Alpha = 0.01
Max pool 2	$3\times 3$, stride 3, pad 0	Max pool 2	$3\times 3$, stride 3, pad 0	Max pool 2a	$3\times 3$, stride 3, pad 0
				Conv 2	$64\times 2\times 2$, pad 0, stride 1
Dropout	0.1	Dropout	0.1
Fully connected	128	Fully connected	128	Leaky ReLU 2	Alpha = 0.01
1 + ReLU		1 + ReLU		Max pool 2b	$3\times 3$, stride 3, pad 0
Fully connected	8	LSTM 1 + ReLU	8	Dropout	0.1
2 + ReLU		L2 regularizer	0.01
L2 regularizer	0.01	Dropout	0.25
Dropout	0.25	Fully connected 2	1 sigmoid	Merge left branch and right branch
Fully connected 3	1 sigmoid			Conv 3 + ReLU	$64\times 2\times 2$, pad 0, stride 1
				Max pool 3	$2\times 2$, stride 2, pad 0
				L2 regularizer	0.01
				Dropout	0.1
				Fully connected 1	1 sigmoid
Total parameters: 841,681		Total parameters: 845,033		Total parameters: 39,553