5. Experiments and Results
The performance of the word recognizer was evaluated
used the ready segmented Arabic Database. The wide known
IFN/ENIT [21] Tunisia database is used and the result is
concluded in table(1).A high precision can be obtained for
words classification from printed documents see[23]. The
software used is LIBSVM [24] with additional utilities
developed in mat lab. The set of samples obtained from
78 word/shapes is divided into two to be
considered as training and testing set, respectively.
In the final testing stage, ten-fold cross validation is
used, followed by actual classification using these values.
The recognition rates obtained are in the range of 94-99%
correct classification, depending on the kind and number of
classes (words) used. For the printed database, which is 7
example for every class of 56 class segmented from the
PATS-A01[23]and achieved a recognition rate between 96 %
to 1.The difference of the number of feature vector affect the
requerd time but the output precision is the sameas in table 1.
Feature vector by SIFT/feature vector SIFT with maxima =
1400/200 = 7 times faster