Academia.eduAcademia.edu

Figure 1 – uploaded by Khaled Younis

See full PDFdownloadDownload figure
Fig. 1: A Set of 48 Samples of the Isolated Ain Form We have extracted from these page samples about 440 collections of individual words, sub-words, and letter forms. Each collection comprises 48 samples from 48 different persons. Fig. 1 shows the collection of 48 samples of the Isolated Ain form.

Figure 1 A Set of 48 Samples of the Isolated Ain Form We have extracted from these page samples about 440 collections of individual words, sub-words, and letter forms. Each collection comprises 48 samples from 48 different persons. Fig. 1 shows the collection of 48 samples of the Isolated Ain form.

Related Figures (6)

Fig. 2: Arabic OCR Application Tool and its Routine Selection Dialog Box
Fig. 2: Arabic OCR Application Tool and its Routine Selection Dialog Box
Each classifier was tested using the 10-fold cross- validation to find the classification accuracy [19], [20]. Fig. 3 shows the classification accuracy of these five classifiers. The QDA classifier has the best classification accuracy when using up to 20 features. This is expected because the nonlinear decision surface obtained with the QDA classifier separates classes better than the linear surface. The QDA classifier’s accuracy falls with features more than 20 because the number of free parameters to be estimated approaches the number of available training data points. For g > 40, it is not possible to estimate a sample covariance matrix because the number of training samples is not sufficient. Fig. 3: Classification Accuracy using Five Classifiers
Each classifier was tested using the 10-fold cross- validation to find the classification accuracy [19], [20]. Fig. 3 shows the classification accuracy of these five classifiers. The QDA classifier has the best classification accuracy when using up to 20 features. This is expected because the nonlinear decision surface obtained with the QDA classifier separates classes better than the linear surface. The QDA classifier’s accuracy falls with features more than 20 because the number of free parameters to be estimated approaches the number of available training data points. For g > 40, it is not possible to estimate a sample covariance matrix because the number of training samples is not sufficient. Fig. 3: Classification Accuracy using Five Classifiers
Fig. 4: QDA Classification Accuracy for the Four Forms Taken Separately Fig. 4 shows the QDA classi letter forms taken separately when the number of features fall down. The Final form 86% and the Isolated form 76%. The curves of the Fina fication accuracy for the four . The four curves climb fast is increased, then the curves has the highest accuracy of has the lowest accuracy of and Initial forms rise faster than the curves of Medial and Isolated forms.
Fig. 4: QDA Classification Accuracy for the Four Forms Taken Separately Fig. 4 shows the QDA classi letter forms taken separately when the number of features fall down. The Final form 86% and the Isolated form 76%. The curves of the Fina fication accuracy for the four . The four curves climb fast is increased, then the curves has the highest accuracy of has the lowest accuracy of and Initial forms rise faster than the curves of Medial and Isolated forms.
Table 2: Worst Ten letters Recognized Nine of the worst 10 letters are of the Isolated or Medial forms, consistent with the results shown in Fig. 5. Letters with dots above the main body tend to have low classification accuracy because the variations in drawing the dots give inaccuracies in extracting the important secondary type feature. Note that some Medial letters that have loops such as Ghain (s) and Feh (a) have subtle difference in the way the loop is drawn and, consequently, they have low classification accuracy.
Table 2: Worst Ten letters Recognized Nine of the worst 10 letters are of the Isolated or Medial forms, consistent with the results shown in Fig. 5. Letters with dots above the main body tend to have low classification accuracy because the variations in drawing the dots give inaccuracies in extracting the important secondary type feature. Note that some Medial letters that have loops such as Ghain (s) and Feh (a) have subtle difference in the way the loop is drawn and, consequently, they have low classification accuracy.
etter forms taken separa features. The Final form Fig. 5 shows the LDA classification accuracy for the four ely. The four curves also climb fast when the number of features is increased from 5 to 20, then the curves almost cease to rise after using 30 has the highest accuracy of 91% and the Isolated form has the lowest accuracy of 84%. The curves of the Final and Initial forms rise faster than he curves of Medial and Isolated forms, and the former wo curves reach higher accuracies. Fig. 5: LDA Classification Accuracy for the Four Forms Taken Separately
etter forms taken separa features. The Final form Fig. 5 shows the LDA classification accuracy for the four ely. The four curves also climb fast when the number of features is increased from 5 to 20, then the curves almost cease to rise after using 30 has the highest accuracy of 91% and the Isolated form has the lowest accuracy of 84%. The curves of the Final and Initial forms rise faster than he curves of Medial and Isolated forms, and the former wo curves reach higher accuracies. Fig. 5: LDA Classification Accuracy for the Four Forms Taken Separately
Fig. 6: LDA Classification Accuracy for the Average of the Four Forms Taken Separately, All Forms with the Form Type Feature, and All Forms without the Form Type. Fig. 6 shows the curve of the average of these four curves which has a maximum of 87%. The lower curve in this figure shows the LDA classification accuracy when the classifier is used to recognize all four forms. With a maximum of 76%, the lower curve indicates that we get lower accuracy when the classifier is required to recognize all forms.
Fig. 6: LDA Classification Accuracy for the Average of the Four Forms Taken Separately, All Forms with the Form Type Feature, and All Forms without the Form Type. Fig. 6 shows the curve of the average of these four curves which has a maximum of 87%. The lower curve in this figure shows the LDA classification accuracy when the classifier is used to recognize all four forms. With a maximum of 76%, the lower curve indicates that we get lower accuracy when the classifier is required to recognize all forms.
Related topics:
Optical Character RecognitionInformation Processing and Pattern RecognitionMultiple Classifiers
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved