Table 3. Accuracy (in (%)) of our system when using rotated images in different angels
F.E.M.
Matching
Angles of rotation (
◦
)
0
45
90
135
180
225
270
315
SIFT
Min Dist.
100
98.9
97.8
96.7
100
97.8
100
98.9
k-NN_5
100
100
100
96.7
100
98.9
100
100
SVM
100
100
98.9
98.9
100
98.9
100
100
In fourth scenario, we tested our system when the gestures are shifted or occluded horizontally and
vertically in different percentage of its sizes. In other word, we try to investigate whether our proposed
system is robust against occluded images while identifying the character. In this scenario, we used four
training images and the testing images are occluded to identify the UzSL character. The results of this
experiment are shown in Table (4).
Table 4. Accuracy (in (%) of our system in case of image occlusion
F.E.M.
Matching
Percentage of Occlusion
Horizontal
Vertical
20
40
60
20
40
60
SIFT
Nearest Neighbord
98.9
93.3
34.4
98.9
95.6
32.2
k-NN_5
97.8
95.6
38.9
97.8
96.7
53.3
SVM
98.9
95.6
52.2
98.9
96.7
45.6
In order to evaluate the performance of our proposed system, we have considered the percentage
of the total number of Arabic sign characters that were correct as the main factor. The results, summarized
in Table (1, 2, 3 and 4), will be discussed to confirm this factor.
Do'stlaringiz bilan baham: |