Page 83 - 2024-Vol20-Issue2

79 | AlKarem, Khalid, & Ali
Fig. 1. The framework of the proposed method.
A. Data Preparing
This method is utilizing two offline image datasets of handwrit-
ten signatures. The first dataset called handwritten signature,
and it consisted of English signatures. This dataset was used
by [26], and it was on Kaggle platform
(https://www.kaggle.com/datasets/divyanshrai/handwritten-
signatures).
The data consists of 360 genuine signature images and 360
forgery signature images. The total images are 720 images
for both genuine and forgery. This data is prepared for model
training and fitting. The second data is for Arabic signature
and is prepared by this study for model testing and deploy-
ment, the number of signature images within this dataset for
each category is 80 images for genuine and 80 images for
forgery, where 20 different people were asked to sign 4 forged
signatures and 4 real signatures.

B. Data Preprocessing
After data uploading and images are read, a pre-processing
operation is performed to prepare data for features extraction
and recognition process using CNN model of recognition. At
a very beginning the images is resized into 64× 64 dimension,
to facilitate the model process and prevent noise from getting
in, then the images is converted into grayscale color in order
to make the model learning easier. Normalization and Bina-
rization process is then performed. The labels are hot-coded
into binary categorical 0 for genuine class, and 1 for forgery
class. Later, the data are shuffled and splits randomly into
80% train data, and 20% test data. The test data are further
split into 15% validation data and 5% test data.

C. Model Construction
This study constructs a model for offline handwritten signa-
ture recognition utilizing a deep learning concept. The model
consists of features extraction procedure and a classification
or signature verification procedure. The feature’s capturing
procedure is done using the convolutional neural network [14],
since CNN features can extract the deep cues in the image
reaching into an object of interest [27]. This work proposed
building three convolution layers for features extraction in-
cluding three max pooling layers for features selection. The
recognition and signature verification procedure is performed
by proposing two fully connected layers of a deep neural net-
work; the proposed model architecture is shown in Fig. 2.
More details regarding the architecture of the proposed model
can be reached in Table I.

D. Features Extraction
Three conventional layers are proposed for deep features ex-
traction from images of signature. CNN has proved its ability
to capture high hire sophisticated appearance features within
the images. Convolution process is performed with filter size