5 Recognition results

The Manchester database is described in [ 4 ]. There are 300 training images, 10 for each of 30 people. There are 300 normal test images, again 10 for each of 30 people, and there are 90 difficult test images, 3 for each of 30 people.

5.1 Results obtained by Lanitis et al

Recognition rates
Normal test set		Difficult test set
Fraction	Percentage	Fraction	Percentage
276/300	92	44/90	48

5.2 Combined templates

Recognition rates
Grid dimension, n	Normal test set		Difficult test set
	Fraction	Percentage	Fraction	Percentage
1	152/300	51	29/90	32
3	236/300	79	49/90	54
5	257/300	86	59/90	66
7	262/300	87	59/90	66
9	271/300	90	61/90	68
11	270/300	90	63/90	70
13	274/300	91	55/90	61
15	267/300	89	55/90	61

5.3 Individual templates

The component templates are applied individually to the difficult test set. The template based on the raw gradient magnitude is also tested. For this comparison, the training images are obtained using combined templates.

Recognition rates - Grid dimension: n =11
Template employed	Difficult test set
	Fraction	Percentage
Ranked grey-levels	53/90	59
Ranked gradient magnitude	47/90	52
X-component of the unit gradient vector	55/90	61
Y-component of the unit gradient vector	42/90	47
Gradient magnitude	21/90	23

5.4 Using a subset of windows for matching

Combined templates are used.

Recognition rates - Grid dimension: n =11
Percentage of	Normal test set		Difficult test set
windows summed	Fraction	Percentage	Fraction	Percentage
100	270/300	90	63/90	70
87.5	283/300	94	65/90	72
75	289/300	96	66/90	73
50	295/300	98	68/90	76
25	289/300	96	69/90	77
12.5	276/300	92	67/90	74

Recognition rates - Grid dimension: n =15
Percentage of	Normal test set		Difficult test set
windows summed	Fraction	Percentage	Fraction	Percentage
100	267/300	89	55/90	61
87.5	274/300	91	56/90	62
75	286/300	95	59/90	66
50	288/300	96	65/90	72
25	288/300	96	72/90	80
12.5	286/300	95	70/90	78