Page 82 - 2024-Vol20-Issue2

78 |
styles [9, 11]. In WD style, a classifier is trained separately
on samples of signatures for each person, and when a new
person’s signature joins the system, the classifier is retrained. CEDAR that involves 24 images of genuine and 24 images
In contrast, in the WI approach the system is developed so that
a new person’s signature is checked without the necessity to
retrain the classification model [12, 13]. This study proposes
to use WI style. Signature verification is an essential part of
many business processes.
There are many of introduced studies for online and offline
signature verification in the literature. Ismail et al. [14] In-
troduced offline Arabic signature recognition and verification
system of two separate phases. This technique based on a
multistage classifier and a set of global and local features.
The algorithms for signature verification are relies on fuzzy
concepts. Iranmanesh et al. [15] was employed a systematic
method for online signature verification using multilayer per-
ceptron (MLP) on a subset of principal component analysis
(PCA) features to analyze the signature time series signals.
This method explain a feature selection technique utilizing
information extracted from PCA on handwritten signature
which can be significant by obtaining reduced error rates, this
technique obtain an 93.1% accuracy on 200 users and 8,000
signatures consisting of genuine and forger signatures. In
Hafemann et al. [16] the method employed the representations
from signature images, using CNN to process the difficulty
of obtaining good features, and improve system performance.
They suggested a new formulation that includes knowledge of
the professional forgeries from a subset of users in the feature
learning process, which aims to capture visual sign that distin-
guish real signatures and forgeries regardless of the user. In
Gideon et al. [17] handwritten signature forgery detection is
explored on English signature dataset. The forgery signature
system used the static features which involves image pro-
cessing techniques to analyses the accuracy of the signatures
based on a CNN. Poddar et al. [18] used a method special-
ize in the signature as biometric feature to discern forgery
in signature. It used a Convolution Neural Network (CNN)
for signature forgery detection and relies on Crest-Trough
method, speeded up robust features (SURF) algorithm and
Harris corner detection algorithm; this system got an accuracy
of 85-89% for forgery detection and 90-94% for signature
recognition. Ghanim et al. [19] introduced SVM and CNN
classifiers independently for offline Arabic handwritten recog-
nition method, by implementing multistage cascading system.
This approach began with applying the Hierarchical Agglom-
erative Clustering (HAC) technique to divide the database
into partially interrelated clusters. The inter-relations support overfitting problem. At last of model architecture, Softmax
representing the database as a big search tree model and help
to reduce the recognition complexity in matching for each
test image with a cluster, and higher recognition accuracy of
90%. Upadhyay et al. [20] introduced offline signature veri-
AlKarem, Khalid, & Ali
fication model based on multi-dilation convolutional neural
network, the proposed model was validated using dataset of
of forgery for every 55 signer. Longjam et al. [21] proposed
multi-scripted writer independent offline signature verification
method by suggesting hybrid of CNN and Bi-directional Long
Short Term memory (BiLstm) techniques, to recognize skilled
forgery and genuine signatures. Their study used different
datasets from the literature such as GPDS-300, GPDS-Bengali,
GPDS-Devanagari, CEDAR, BHSig260-Bengali, BHSig260-
Hindi, and Meitei Mayek signature. The system evaluated on
various multi-scripted signatures based offline that belongs
to multiple lingual Indian community. However, lots of stud-
ies have been proposed for WI offline signature verification.
Yet, it is necessary to improve the overall accuracy measure-
ments [7, 22–25]. Working this paper proposed a handwritten
signature recognition method using CNN model architecture.
The developed model is tested on Arabic handwriting signa-
tures data to recognize whether the provided signature is real
or forged. The contribution of this study is summarized as
below:
1- Developing deep learning model for writer independent
(WI) offline signature verification system based on convolu-
tional neural network model architecture.
2- Developing an Arabic signature dataset that consisting of
genuine signature samples and forged signature samples.
3- Evaluating and testing the deployed model to make classifi-
cation and verification on the Arabic signature dataset.
The paper is organized as follows: Section II. presents the
proposed method, Section III. presents the results and discus-
sions, and Section IV. presents the conclusion.
II. PROPOSED METHOD
This paper introduced a method for offline handwritten sig-
nature recognition, by using a deep learning concept. The
proposed method implemented a model architecture based on
CNN [1]. Features of CNN network is proposed to extract
depth and high descriptive cues from signature images, by
using multiple CNN layers. Max pooling layer is proposed
to interpose between CNN layers to select good and most
representative features each time. The extracted features are
flattened into one feature vector. Then, the fully connect
layer (Dens network layer) with 128 neurons is proposed for
features mapping and to make classifier learns from these
features. Dropout layer of 20% is proposed here to overcome
classifier is proposed to recognize real signatures from forged
signatures. The framework of the proposed method is depicted
in Fig. 1.