Page 85 - 2024-Vol20-Issue2
P. 85
81 | AlKarem, Khalid, & Ali
Fig. 3. Progress of model fitting using loss error metric.
Fig. 4. Progress of model fitting based on accuracy metric. Fig. 5. The outcomes of signature recognition using images
from Arabic dataset.
ing progress based on loss error and accuracy are depicted in
Fig. 3, and Fig. 4. The obtained results from model training recognition procedure. The method can perform correct pred-
and fitting are summarized in Table II. The model is saved ication and recognize the genuine signatures from forgery
using “.h5” file and deployed on test data (non-trained im- signature based on Arabic signature data as depicted in Fig. 5,
ages) to make prediction, the obtained accuracy and loss error and Fig. 6. The Arabic signature dataset has prepared by this
are 100%, and 0.0488 consecutively as illustrated in Table II, work to be employed within this prediction task.
which demonstrate that the model has succeed in recognizing
all images of test data. Furthermore, the deployed model is Comparison evaluation with recent studies has been made
used to make prediction on Arabic signature data images as to recognize the performance of the proposed method from
non-trained data. Arabic signature dataset was prepared by previous studies performance. Table III. conducts a compari-
this study, which consists of 80 images of forgery signature son with previous methods based on accuracy metric, which
and 80 images of real signatures. Consequently, the gained reveals that the proposed method has good performance versus
outcomes presented a good performance for the proposed other stated studies in Table III.
method including the deployed model based on the proposed
architecture of CNN network for both features extraction and IV. CONCLUSION
This paper highlighted the challenges of WI offline signature
verification systems and the necessity to improve the perfor-
