I. INTRODUCTION
The handwritten signature has always been one of the most
simple and accepted ways to authenticate an official
document. Research into signature verification has been
vigorously pursued for a number of years and it is being
explored especially in the off-line mode [1, 2]. The
recognition of human signatures is significantly concerned
with the improvement of the interface between human-beings
and computers [3, 4]. A signature verification system and the
associated techniques used to solve the inherent problems of
authentication can be divided into two classes: (a) on-line
methods [
7, 8
] to measure the sequential data such as order of
stroke, and writing speed, pen pressure and other temporal
information by utilizing intelligent algorithms [
9, 10
], and (b)
off-line methods [
11, 12
] that use an optical scanner to obtain
handwriting data written on paper. On-line signature
verification has been shown to achieve much higher
verification rates than off-line verification [
11
] as a
considerable amount of dynamic information is lost in the
off-line mode.
Signatures are not considered as a collection of letters and
words [
14
]. It is often difficult for a human to instantly verify
two signatures of the same person because signature samples
from the same person are similar but not identical and
signatures can change depending on elements such as mood,
fatigue, time etc. Great inconsistency can even be observed in
signatures according to country, habits, psychological or
mental state, physical and practical conditions [
15
].
Significant research has been performed in the field of
signature verification involving English signatures, but to the
best of our knowledge, very little attention has been given
towards non-English signatures such as Chinese, Japanese,
Arabic etc. In order to convey the state-of-the-art of
nonEnglish signature verification, in this paper we present a
survey of non-English and non-Latin signature verification
systems.
III. TYPES OF FORGERIES
There are usually three different types of forgeries to take
into account. According to Coetzer et al. [
16
], the three basic
types of forged signatures are indicated below:
1. Random forgery. The forger has no access to the genuine
signature (not even the author‟s name) and reproduces a
random one.
2. Simple forgery. The forger knows the author‟s name, but
has no access to a sample of the signature.
3. Skilled forgery. The forger has access to one or more
samples of the genuine signature and is able to reproduce it.
But based on the various skilled levels of forgeries, it can
also be divided into six different subsets. The paper [
17
]
shows various skill levels of forgeries and these are shown
below.
1. A forged signature can be another person‟s genuine
signature. Justino et al. [
18
] categorized this type of forgery
as a Random Forgery.
2. A forged signature is produced with the knowledge about
the genuine writer‟s name only. Hanmandlu et al. [
19
]
categorized this type as a Random Forgery whereas Justino et
al. [
18
] categorized this type as a Simple Forgery. Weiping et
al. categorized this type as a Casual Forgery [
20
].
3. A forged signature imitating a genuine signature‟s model
reasonably well is categorized as a Simulated Forgery by
Justino et al. [
18
].
4. Signatures produced by inexperienced forgers without the
knowledge of their spelling after having observed the genuine
specimens closely for some time are categorized as Unskilled
Forgeries by Hanmandlu et al. [
19
].
5. Signatures produced by forgers after unrestricted practice
by non-professional forgers are categorized as Simple
Forgery/Simulated Simple Forgery by Ferrer et al. [
21
], and a
Targeted Forgery by Huang and Yan [
22
].
6. Forgeries which are produced by a professional imposter
or person who has experience in copying Signatures are
categorized as Skilled Forgeries by Hanmandlu et al. [
19
].
IV. NON-ENGLISH SIGNATURE VERIFICATION
TECHNIQUES
We think that the shape of non-English signatures and
writing styles are different to English signatures. Arabic
script is written from left to right. Most of the Japanese
signatures consist of two to six kanji, hiragana and/or
katakana component characters and they are spaced
appropriately from each other. Persian signatures are also
different from other signature types because people usually
do not use text in it and they draw a shape as their signature.
Hence in this work, non-English signature verification
systems are reported and they are described below.
A. Chinese Signature Verification Systems
Chinese signature consists of many strokes and these
strokes can be taken into consideration for signature
authentication. Liu [
23
] discussed this issue, but he discussed
it from the point of view of identifying a signature manually.
Off-line Chinese Signature Verification Systems
Lv et al. [
24
] developed a Chinese off-line signature
verification system. A database of 1100 signatures was
developed for experimentation. Support Vector Machines
(SVM) are used as a classifier. Four different types of
features such as moment feature, direction feature, grey
distribution and stroke width distribution are used here.
Based on each feature, the accuracies are calculated
separately and an average accuracy was also calculated based
on all combined feature sets. An average error rate 5.10% is
found using the combined feature sets. SVM based
techniques are also proposed by Chen et al. [
25
] and Meng et
al. [
26
] for Chinese signature verification. Shen et al. [
27
]
proposed an off-line Chinese signature verification system
based on geometric features. A database of 800 signatures
was used for experimentation and obtained 96.8% accuracy.
Four main features such as: (a) Envelope of the signature (b)
Cross-count feature (c) Centre of gravity feature and
distance between vectors made from the centre of gravity (d)
Embedded white area and position are used to optimise the
verification scheme. Some similar works are also proposed
by Bajaj et al. [
28
] and Huang et al. [
29
].
Lin and Li [
30
] proposed a Chinese signature verification
scheme using normalized Zernike moment invariants
(NZMI). A total of 210 signature samples were collected
from 35 writers. The average accuracies of 8% and 12%
are obtained for FRR and FAR, respectively. Belkasim et
al. [
31
] introduced a new recursive formula to derive
Zernike moments.
In another work of Lin and Li [
32
], they utilized a set of
shape features based on special characteristics of Chinese
signatures along with high pressure feature. Their features
includes: (a) Ratio of a signature's height to its width. (b)
Ratio of a signature's height to its packed width (c) Slant (d)
Stroke width. To define the global high-pressure features
(GHP) they use Ammar et al‟s [
33
] dynamic threshold
selecting method. A database of 100 genuine Chinese
signatures and 50 forged signatures are collected for the
experiment. Reported FRR and FAR rates are 1.0% and
4.0%, respectively.
Chang et al. [
34
] presented a dynamic handwritten Chinese
signature verification system based upon a Bayesian neural
network. Features such as: timing features, average velocity
feature, average length in the eight directions, width/height
ratio, left-part/right-part density ratio, upper-part/lower-part
density ratio etc are utilize in the work. Similar works are
proposed by Brault and Plamondon [
35
] and Lorette [
36
]. A
database of 1200 signature samples is collected. The
experimental results show the type I error is about 2% and
the type II error rates are approximately 0.1% and 2.5% for
“simple” and “skilled” forgeries, respectively.
Ji et al. [
37
] developed an off-line Chinese signature
verification system based on a weighting factor of similarity
computation. Their earlier paper introduces an improved
approach to verify off-line Chinese signatures and it is
described in [
38
]. In their proposed scheme, seven features
such as (a) Relative horizontal centre (b) Relative vertical
centre (c) The number of points having horizontal neighbours
(d) The number of points having vertical neighbours, (e) The
number of points having positive diagonal neighbours (f) The
number of points having negative diagonal neighbours and
(g) Stroke thickness of the segments are used. This technique
for off-line Chinese signature verification based on different
weighting factors is compared with an expert on questioned
documents used to verify a signature sample [
39
]. The
experimental results are generated differently using different
data sets. The average ERR is 3.30% and the average EAR is
16.50% for simple forgeries when the weighting factor is
0.04.
Ji and Chen [
40
] proposed an off-line Chinese signature
verification System. A method to solve the problem for
random forgeries and simple forgeries is presented in their
paper. The pre-processing techniques used here are described
in detail in [
41
]. The features are extracted in seven steps as
discussed in the paper [
33
]. A database of 4800 handwriting
samples from 32 participants is used in this method to
obtain a verification accuracy rate of 91%.
Zuo et al. [
42
] proposed an off-line Chinese signature
verification scheme using Pseudo-Zernike invariant moments
as for static features due to scale and translation invariance.
High-density factors, relative gravity centre and Wavelet
Transform are used as dynamic features. A database of 290
signatures was collected. As a result of their experiments, the
FAR and FRR was 7.84% and 6.89%, respectively.
Cheng et al. [
43
] presented a handwritten Chinese
signature verification scheme. An attributed string matching
approach based on the writing sequences of an input
signature is proposed. In order to obtain an attributed string
that is used in the string matching similarity calculation, the
input signatures are split into several segments. The stroke
attributed feature is used in their proposed technique. A large
database is used to obtain 1.5% and 3.6% for type1 and type2
error rates respectively. A similar matching method is
performed by Chen et al. [
44
].
Ye et al. [
45
] developed an off-line handwritten Chinese
signature verifier with an inflection feature. Different scale
wavelet transforms are used in the curvature signature signals
transformation. The signature curves are divided into several
parts, i.e. the strokes, according to the inflections. The
distance between two corresponding strokes is measured with
a Dynamic Time Warping algorithm. A database of 3120
signatures was collected for the experiments. The rate of FRR
and FAR (skilled forgery) are 1.33 % and 6.72%,
respectively.
On-Line Chinese Signature Verification Systems
Xiao and Dai [
46
] introduced a hierarchical on-line
Chinese signature verification system. First, global features
are applied to obtain a statistical decision through comparing
their weighted distance. Secondly, the input primitive string
is matched with its reference primitive string by attributed
automaton. In their paper an attributed automaton [
47
] which
has four edit operations (insertion, deletion etc.) are applied
to solve the problem of inconsistency of signature
segmentation.
Tseng and Huang [
48
] presented an on-line Chinese
signature verification scheme based on the ART Neural
Network. The verification method based on one bit quantized
pressure patterns, which constitute time domain information.
The timing information contained in the on/off motions of
handwriting is analysed by Zimmermann and Varady [
49
].
Carpenter and Grossberg [
50
] also proposed a method based
on the ART Neural Network. The error rates 4.5% and 5%
are obtained for type1 and type 2, respectively. Techniques
based on neural network expert systems to identify Chinese
signature are proposed by Ng and He [
51
] and He et al. [
52
].
Cheng et al. [
53
] presented an on-line Chinese signature
verification system using a voting scheme. Global feature,
line segment feature, 8-directional chain code feature,
Spectral information, similarity of position sequences,
similarity of velocity sequence, similarity of attribute strings,
segment correlation, Tremor feature are used in these nine
expert steps. A database of 600 genuine signatures and 12000
forge signatures is used. Some similar types of works are
conducted by Suen et al. [
54
] and Jeng et al. [
55
] based on
neural networks and wavelet transforms respectively. Y.
Mizukami [56)] developed a handwritten Chinese character
recognition system using hierarchical displacement extraction
based on directional features. Other techniques involving
online signature verification can be obtained in [
57-64
].
B. Japanese Signature Verification Systems
The Japanese handwritten signature verification is
difficult due to the lack of stability and individuality. Only a
few articles are available on Japanese handwritten
verification and they are discussed as follows. Ueda et al.
[
65
] presented an off-line Japanese signature verification
system using a pattern matching technique. The similarity
between two signatures obtained by pattern matching is
affected by stroke widths. Stroke widths vary with the pen
used for signing, and even if signatures are written with the
same pen, the stroke width may also vary. In their modified
pattern matching method, the strokes of the signatures are
first thinned and then the thinned signatures are blurred by a
fixed point-spread function. A database of 2000 signatures
including 100 genuine signatures from 10 writers and 100
forged signatures from 10 writers are used. An average error
rate 9.10% is obtained. Some techniques for verification of
Japanese handwritten signatures have been proposed in
[6668].
Yoshimura and Yoshimura [
69
] presented off-line
verification of Japanese signatures after elimination of
background patterns. Some preprocessing techniques to
eliminate the background pattern are performed as follows:
position adjustment, filtering, clipping of random noise and
smoothing for noise elimination etc. The verification stage
following the preprocessing stage is based on the Arc Pattern
Method. A small data set is used to obtain an error rate of
approximately 14%. Mizukami et al. [
70
] proposed an
offline Japanese signature verification system using an extracted
displacement function.
C. Persian Signature Verification Systems
Ghandali et al. [
71
] proposed an off-line Persian
signature identification and verification system based on
Discrete Wavelet Transform and image fusion. In this
method, DWT is employed to access high-frequency bands of
signature shape. Then, different samples of a person‟s
signature are fused together based on high frequency bands to
generate the signature patterns. This pattern is saved in the
learning phase. SVMs are used here as classifiers. A database
consists of 6 genuine, 1 simple forgery and 1 skilled forgery
signatures from each of the 90 signers is used. The error
rates, 8.9% and 10% are obtained for FRR and FAR,
respectively. Chalechale and Mertins [
72
], Chalechale et al.
[
73
] proposed a Persian signature recognition system using
line segment distribution. Zoghi et al. [
74
] introduced a
Persian signature verification system using Improved
Dynamic Time Warping-based Segmentation and
Multivariate Autoregressive Modelling. A database including
1250 genuine signatures and 750 forged signatures was used
to obtain an accuracy of 88.8% for the testing of skilled
forgery signatures. The statistical spectral estimate for each
signature segment is obtained via the use of an
AutoRegressive model [
75
]. The verification process is carried out
using an Artificial Neural Network with a multilayer
perceptron architecture described in [
76
].
D. Arabic Signature Verification Systems
Ismail et al. [
77
] proposed an off-line Arabic signature
recognition and verification technique. In the first phase
(Identification phase) some features are extracted and there
features are: area filtering, translation, extraction of the
circularity feature, normalization, image enhancement, partial
histogram (Vertical projection, Horizontal projection),
Centres of gravity, extraction of the global baseline (BSL),
extraction of the upper limit (UL) and lower limit (LL),
thinning, calculation of the global slant etc. In this phase, the
features are classified into two main groups: global features
and local features. In the second phase (Verification phase)
some other features are also extracted such as central line
features, corner line features, central circle features, corner
curve features and critical point features. A set of signature
data consisting of 220 genuine samples and 110 forged
samples is used for experimentation. Their system obtained a
95.0% recognition rate and a 98% verification rate. Other
techniques of Arabic handwritten word recognition systems
are described in [
78-87
].