Table 4.

Various CNN models for comparisons (convN means the filter of N×. For example, conv3 represents the filter of 3 × 3).

Input Image	Original Image as Input to CNN (Case 1 of Table 2)			Difference Image as Input to CNN (Case 2 of Table 2)
Net configuration	VGG Face (no fine-tuning/fine-tuning)	VGG Net-16 (no fine-tuning/fine-tuning)	VGG Net-19 (no fine-tuning/fine-tuning)	Revised Alexnet-1 (whole training)	Revised Alexnet-2 (whole training)	VGG Net-16 (fine-tuning) (Proposed method)
Method name	A/A−1	B/B−1	C/C−1	D	E	F
# of layers	16	16	19	8	8	16
Filter size (# of filters)	conv3 (64) conv3 (64)	conv3 (64) conv3 (64)	conv3 (64) conv3 (64)	Conv11 (96)	conv3 (64)	conv3 (64) conv3 (64)
Pooling type	MAX	MAX	MAX	MAX	MAX	MAX
Filter size (# of filters)	conv3 (128) conv3 (128)	conv3 (128) conv3 (128)	conv3 (128) conv3 (128)	Conv5 (128)	conv3 (128)	conv3 (128) conv3 (128)
Pooling type	MAX	MAX	MAX	MAX	MAX	MAX
Filter size (# of filters)	conv3 (256) conv3 (256) conv3 (256)	conv3 (256) conv3 (256) conv3 (256)	conv3 (256) conv3 (256) conv3 (256) conv3 (256)	conv3 (256)	conv3 (256)	conv3 (256) conv3 (256) conv3 (256)
Pooling type	MAX	MAX	MAX		MAX	MAX
Filter size (# of filters)	conv3 (512) conv3 (512) conv3 (512)	conv3 (512) conv3 (512) conv3 (512)	conv3 (512) conv3 (512) conv3 (512) conv3 (512)	conv3 (256)	conv3 (256)	conv3 (512) conv3 (512) conv3 (512)
Pooling type	MAX	MAX	MAX		MAX	MAX
Filter size (# of filters)	conv3 (512) conv3 (512) conv3 (512)	conv3 (512) conv3 (512) conv3 (512)	conv3 (512) conv3 (512) conv3 (512) conv3 (512)	conv3 (128)	conv3 (128)	conv3 (512) conv3 (512) conv3 (512)
Pooling type	MAX	MAX	MAX	MAX	MAX	MAX
Fc6 (1st FCL)	4096	4096	4096	4096	2048	4096
Fc7 (2nd FCL)	4096	4096	4096	1024	2048	4096
Fc8 (3rd FCL)	2622/# of class	1000/# of class	1000/# of class	2	2	2