Table 3.

Descriptions of the three databases used in our research (*: index, middle, and ring fingers, **: from authentic matching, ***: from imposter matching).

			Good-Quality Database	Mid-Quality Database	Low-Quality Database
Original images		# of images	1200	1980	3816
		# of people	20	33	106
		# of hands	2	2	2
		# of fingers *	3	3	3
		# of classes (# of images per class)	120 (10)	198 (10)	636 (6)
Data augmentation for training	CNN using original image as input (Case 1)	# of images	72,600 (60 classes × 10 images × 121 times)	119,790 (99 classes × 10 images × 121 times)	230,868 (318 classes × 6 images × 121 times)
	CNN using difference image as input (Case 2)	# of images	15,480	25,542	48,972
			7740 ** ((10 images × 13 times – 1) × 60 classes)	12,771 ** ((10 images × 13 times – 1) × 99 classes)	24,486 ** ((6 images × 13 times – 1) × 318 classes)
			7740 ***	12,771 ***	24,486 ***