![](bg4.png)
for 10 times. Previous studies (Mezghani et al., 2003;
Mezghani et al., 2005) tested on this database, used
288 and 144 samples of each class for training and
testing, respectively. We have used the same 2 out
of 3 ratio between the number of training and testing
samples to make our results more comparable to those
previously reported. We aimed for more confidence in
the reported recognition rate by following the experi-
mental set up as explained over ten runs reported the
statistics on recognition results with 95% confidence
interval. This total recognition is 95.2± 0.12.
Table 1: Performance comparison of different recognition
systems.
Performance 1-NearestNeighbor Kohonen memory RF-Tangent
Recognition Rate Ref -1.19 + 4.22
Training Time - 2 hrs 7.5 min
Recognition Time 526 s 26 s 23 s
Table 2: Confusion matrix for a sample run using the com-
bination of relational and directional features.
class Recog
label 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 rate%
1 143 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 99.31
2 0 137 0 0 0 1 1 0 0 1 0 0 0 4 0 0 0 95.14
3 0 0 144 0 0 0 0 0 0 0 0 0 0 0 0 0 0 100.00
4 0 0 0 130 0 0 0 0 0 1 0 12 0 1 0 0 0 90.28
5 0 2 0 1 140 0 0 0 0 0 0 1 0 0 0 0 0 97.22
6 0 1 0 1 0 137 0 0 0 2 0 0 1 0 0 0 2 95.14
7 0 1 0 0 0 1 142 0 0 0 0 0 0 0 0 0 0 98.61
8 0 0 0 0 0 0 0 144 0 0 0 0 0 0 0 0 0 100.00
9 0 0 0 0 0 0 0 0 141 0 0 0 1 0 1 0 1 97.92
10 0 6 0 0 1 2 1 0 0 133 0 0 0 0 0 0 1 92.36
11 0 0 0 0 0 0 0 0 0 0 144 0 0 0 0 0 0 100.00
12 0 0 0 7 0 0 0 0 0 1 0 133 0 3 0 0 0 92.36
13 0 0 0 0 0 1 0 0 1 0 0 0 139 0 0 3 0 96.54
14 0 0 0 3 0 0 0 0 0 0 0 2 0 132 3 0 4 91.67
15 0 0 0 0 0 1 1 0 0 0 1 0 1 1 139 0 0 96.53
16 0 0 0 0 0 0 0 0 0 3 0 0 0 0 0 141 0 97.92
17 0 1 0 0 0 0 0 0 0 1 0 0 0 2 5 0 135 93.75
total
rate 96.16
Table 1 summarizes the results. The first column
presents our experiments with the 1-Nearest Neighbor
classifier as this classifier is often used as a bench-
mark. Its asymptotic error rate is less than twice the
optimal Bayes rate. However, its asymptotic through-
put is zero. Second column shows the results reported
in (Mezghani et al., 2005), and the last column shows
the results of our experiments with feed forward MLP
neural network classifier equipped with the RF fea-
ture and the tangent feature, described in Section 3,
conducted on Pentium(R)D with 340-MHz CPU. In
the first row, the recognition rate of all methods pre-
sented. The recognition rate for 1-Nearest Neigh-
bor classifier is considered as the reference point for
this comparison. Our method shows 4.22% improve-
ment, on average, compare to the one achieved by
1-Nearest Neighbor classifier. This is while the re-
sults in (Mezghani et al., 2005) in the best case can
only reach 1.19% less than 1-Nearest Neighbor clas-
sifier. The training and recognition times presented
in the second and third rows of Table 1. Although
the results reported in (Mezghani et al., 2005) are
based on slightly faster machine than ours, recogni-
tion is faster by our method and the training time is
significantly shorter than the ones in (Mezghani et al.,
2005). The confusion matrix of a sample run is pre-
sented in Table 2. The recognition rate for different
classes vary. On average, our recognition rate for dif-
ferent letters was more robust than the best ones re-
ported in (Mezghani et al., 2005). This confirms the
high discrimination ability of the feature vectors in-
troduced in our recognition system.
5 CONCLUSIONS
We introduced relational feature for online handwrit-
ing recognition and showed the usefulness of this fea-
ture in Arabic character recognition. Our experiments
suggest that this feature representation improves the
state-of-the-art recognition performance. This repre-
sentation provides a rich enough representational fea-
ture for the global shape of these characters. A com-
bination of this global feature, and the local tangent
feature which captures the temporal information of
the online data, improves the recognition rate com-
pared for the same database. In future, we intend to
investigate the use of these features in other super-
vised learning techniques and also for online word
recognition.
REFERENCES
Bahlmann, C., Haasdonk, B., and Burkhardt, H. (2002).
On-line handwriting recognition with support vector
machines – a kernel approach.
Belongie, S., Malik, J., and Puzicha, J. (2002). Shape
matching and object recognition using shape contexts.
IEEE Trans. Pattern Anal. Mach. Intell., 24(4):509–
522.
Mezghani, N., Cheriet, M., and Mitiche, A. (2003). Com-
bination of pruned kohonen maps for on-line aracic
characters recognition.
Mezghani, N., Mitiche, A., and Cheriet, M. (2005). A new
representation of shape and its use for high perfor-
mance in online arabic character recognition by an as-
sociative memory. IJDAR, 7(4):201–210.
Oh, J. and Geiger, D. (2000). An on-line handwriting recog-
nition system using fisher segmental matching and hy-
potheses propagation network. IEEE Conference on
Computer Vision and Pattern Recognition, 2:343 –
348.
Verma, B., Lu, J., Ghosh, M., and Ghosh, R. (2004). A
feature extraction technique for online handwriting
recognition. IEEE International Joint Conference on
Neural Networks, 2:1337 – 1341.
VISAPP 2008 - International Conference on Computer Vision Theory and Applications
562