An Efficient and Robust Iris Segmentation Algorithm Using Deep Learning

Table 2

Experimental results of finding the best architecture of CNN. Note that the “conv_1” and “conv_2” layers shown in the table correspond to the sequences Conv-ReLU-CN and Conv-BN-ReLU, respectively. Further, all the convolution layers run with stride one.

	Test time	Input image size	Layers	RoI pooling grid size	Fully connected layer
A	0.191 s	480 × 640	(3 × 3 × 64) × 3 conv_2	120 × 120	128 ⟶ 2
			(2 × 2) max pool
			(3 × 3 × 64) × 3 conv_2
B	0.197 s	480 × 640	(3 × 3 × 64) × 2 conv_2	120 × 120	256 ⟶ 2
			(2 × 2) max pool
			(3 × 3 × 64) × 3 conv_2
C	0.286 s	480 × 640	(7 × 7 × 128) × 1 conv_1	120 × 120	128 ⟶ 2
			(2 × 2) max pool
			(3 × 3 × 128) × 3 conv_2
D	0.295 s	480 × 640	(7 × 7 × 128) × 1 conv_1	120 × 120	128 ⟶ 2
			(2 × 2) average pool
			(3 × 3 × 128) × 3 conv_2
E	0.286 s	240 × 320	(7 × 7 × 128) × 1 conv_1	64 × 64	128 ⟶ 2
			(2 × 2) max pool
			(3 × 3 × 128) × 3 conv_2
F	0.042 s	120 × 160	(3 × 3 × 128) × 1 conv_1	32 × 32	128 ⟶ 2
			(2 × 2) max pool
			(3 × 3 × 128) × 3 conv_2
G	0.036 s	120 × 160	(5 × 5 × 128) × 1 conv_1	32 × 32	128 ⟶ 2
			(2 × 2) max pool
			(3 × 3 × 128) × 3 conv_2
H	0.037 s	120 × 160	(5 × 5 × 64) × 1 conv_1	32 × 32	128 ⟶ 2
			(2 × 2) max pool
			(3 × 3 × 64) × 3 conv_2
I	0.037 s	120 × 160	(5 × 5 × 64) × 1 conv_1	32 × 32	128 ⟶ 2
			(2 × 2) max pool
			(3 × 3 × 64) × 3 conv_2
J	0.033 s	60 × 80	(5 × 5 × 64) × 1 conv_1	16 × 16	128 ⟶ 2
			(2 × 2) max pool
			(3 × 3 × 64) × 3 conv_2