Li S.Z., Jain A.K. (eds.) Encyclopedia of Biometrics
Подождите немного. Документ загружается.
position, because the angle (or gradient) is one of the
most important pieces of information that help differ-
entiate off-line signature images. In the case of digi-
tized images, rotating an image may add some noise to
deteriorate the transformed patterns.
Feature Extraction
There are numerous proposals for the method of fea-
ture extraction. In most of them, two types of princi-
ples, i.e., pattern matching and multivariate analysis,
are combined to construct an effective procedure for
similarity measurement and decision-making.
In pattern matching, we place a questionable image
on a template for measuring the difference in patterns.
The simple matching of gray levels on meshed pixels
does not yield robust estimates of similarity due to the
shift in the location of images. Therefore, the blurring
or smoothing of gray levels on any pixel with respect to
adjacent pixels, using a suitable weighted average, is
adopted as a method of feature extraction. A set of
shadow masking may be considered to apply this prin-
ciple. The measurement of the fringe of the image is
also a technique for feature extraction in pattern
matching. The accumulation of black pixels in a binary
image along a certain direction such as the x- or y-axis
yields a one-dimensional pattern in the form of a time
series with the coordinate considered as time.
For a multivariate analysis, multiple variables must
be properly defined. The frequency of properly selected
primitive patterns can be treated as variables for anal-
ysis. The number and location of pikes in the image
strokes along a certain di rection may serve as variables
required to lend features. In certain situations, the gray
levels of pixels may act as variables, although the num-
ber of variables could go up to 262,144 when the pixels
are defined in a mesh of 256 1024. The reduction of
dimensionality is an important issue to be resolved.
The size of the envelope or gradient of strokes also may
be used as a feature.
The measurement of frequencies of various local
arc patterns on an image yields multiple variables;
this is possible by the application of a hybrid principle
that combines pattern matching and multivariate
analysis.
Measurement of Similarity
For pattern matching, a weighted sum of local simila-
rities may be an index for evaluating the total similarity.
When the x- or y-axis is regarded as the time axis, the
▶ dynamic programming comparison method and the
framework of a
▶ hidden Markov model may be useful
as the tool to construct a good similarity index [15].
In a multivariate analysis, the Euclidian distance,
▶ Mahalanobis distance, Kullback-Leibler divergence,
or the deviance based on some probabilistic models
can be used, depending on the situation. Linear dis-
criminant functions, quadratic discriminant functions,
▶ support vector machine, or various nonparamet ric
statistics for discrimination are also used for discrimi-
nation, although they do not explicitly measure simi-
larity [16, 17, 18].
Judgment on Authenticity
Since there are several criteria of variability among
writers, simple judging procedures cannot, in general,
extract good performance from the devised system.
A multi-step procedure is more practical for achieving
better results. Therefore, the technique of fuzzy logic
may be effective [19]. The incorporation of the hidden
Markov model or neural network technology in the
judgment system may also be useful for ensuring rea-
sonably correct judgment.
The critical value or boundary line for judgment
should be determined based on the balance of the
possibilities of false acceptance and false rejection;
here, each possibility is, in general, evaluated by the
proportion of the incidence of false acceptance or re-
jection in the test sample used for designing the system.
An increase in false rejections may have grave repercus-
sions on the use of cheques, while an increase in false
acceptances may cause serious damage to businesses.
Signature Recognition. Figuer 2 An example of
background on a traveler’s cheque.
1200
S
Signature Recognition
Review of Judgment
The varia bility in w ritten images is less than that in the
motion of writing, be cause the writer can adjust his
motion by visually tracking the pen movement. This
visuality helps seasoned imposters to carry out skilled
forgery after extensive training. Even when the impos-
ter is not very skillful, the level of forgery is not very
inferior, as seen in Fig. 3. As a result, off-line signature
recognition is not a guaranteed way of identifying or
verifying identity of an individual. This factor necessi-
tates that judgment be reviewed using other means
such as identification cards or passwords, depending
on the gravity of the consequences of misjudgment.
Template Construction
A template is provided for measuring the similarity of
the questionable image with authentic signature
images registered in a database or saved in a system.
In the availability of only one authentic image, it
should be processed in exactly the same manner as
questionable images, in order to construct a template.
In the case wherein multiple images for the same
signature are stored in a reference database, there are
various means of constructing a template. One method
is to amalgamate a set of images in to an image using a
suitable tool. Another method is to select a representa-
tive image that is considered as the preferred image.
Influence of Writing System and
Nationality
There are many letter systems worldw ide, such as Latin,
Chinese, Arabic, Cyrillic, Japanese, and Hangul. The
distinctiveness of signatures is highly dependent on
the letter system. With respect to the man ner of writing,
letters are written separately in Chinese, Japanese, and
Hangul, whereas in English, French, and German, each
word is considered as one unit. In European letter
systems, a name is considered one word, whereas in
some Asian letter systems, it is regarded as a set of
disparate characters. The distinctiveness of signatures
is not as high in the latter group of countries as it is
in the former group due to the social habit of writing
[20, 21, 22]. In the Arabi c letter system, words are
written from right to left.
With respect to the use of signatures, a study of the
Japanese case reveals singular circumstances. In 1883,
the Japanese government legislatively forced all Japa-
nese citizens to use red stamps called hanko or inkan
for official authentication. Further, the Japanese are
taught to write their names in the print form in their
childhood. As a result, Japanese signatures generally
lack uniqueness.
In countries such as Indonesia, the government
allows the people to periodically change their signatures
for registration in order to maintain security. Given such
a situation, the period in which the concerned signature
was written is important to verify the authenticity.
Some people have mu ltiple signatures depending
on their use, examples of which are shown in Fig. 4.
Signature Recognition. Figure 3 Examples of reference images (left), genuine probes (center), and imitated forgeries
(right) for a Japanese signature.
Signature Recognition. Figure 4 Examples of multiple
signatures.
Signature Recognition
S
1201
S
In the left side of this figure, the middle name is
inserted in one signature, while different letter systems
produce different signatures for the same name, as seen
in the right side of the figure. People in countries that
have their own letter system can devise their signatures
based on their native alphabet, although they may
often use signatures written in Latin letters due to the
use of Eng lish as the lingua franca internationally.
On-line Signatures
Introduction
On-line signature verification uses data obtained while
a signature is being written. The data obtained during
the process of writing a signature is called an on-line
signature. Figure 5 depicts a sample of on-line signa-
ture data. On-line signature verification is based on the
hypothesis that the writing style of a signature differs
from person to person and cannot be easily forged.
On-line signature verification verifies whether an
input signature is a genuine signature or a forgery.
Ideally, it is a two-class partitioning problem; however,
it is not an easy problem to solve, because of the
following reasons:
People do not reproduce their signature exactly
each time. Characteristics of the w riting manner
of genuine writers can change over time. There is,
necessarily, intra-class (intra-person) variability. In
contrast, forgers attempt to make their forged sig-
natures as similar as possible to genuine signatures,
and thus inter-class (inter-person) variability
decreases. Therefore, it is difficult to distinguish
between genuine signatures and forgeries.
Both the number and type of signatures available for
training are often severely limited. As on-line signa-
ture verification is a two-class partitioning problem,
general pattern recognition techniques can be ap-
plied if enough data is available from both the clas-
ses. In practice, however, only a few genuine
signatures are available from the genuine class.
Moreover, there are several types of forgeries in the
forgery class, but only a few types of forgery can be
collected for the following reasons: The forgeries
that are most similar to genuine signatures and the
most difficult to distinguish from genuine signatures
will be signatures that were produced by imitating
genuine signatures well. Because genuine signatures
differ from writer to writer, well-imitated forged
signatures should be collected for every writer; how-
ever, this is extremely difficult. Scarcity of genuine
Signature Recognition. Figure 5 On-line signature data.
1202
S
Signature Recognition
training data exist in all biometric methods; however,
scarcity of forgeries exist only in methods that must
prepare for imitation attacks.
A variety of algorithms have been proposed for on-
line signature verification. Early work carried out up to
1999 is summarized in [2, 3, 23]; subsequently, many
studies continue to be reported [6, 7, 24, 25, 26]. A first
on-line signature verification competition was held in
2004 [5].
Basic Operation
Data Acquisition
Several data acquisition devices are used for on-line
signature verification, for example, digital tablets with
pens, pen displays on Tablet PC’s, touch screens on
PDA’s, data acquisition pens, and cameras [6, 7]. These
acquisition devices provide
▶ pen coordinate informa-
tion as time-series data. In addition to pen coordinate
information, some devices can acquire information
about
▶ pen pressure and ▶ pen tilt (▶ altitude, ▶ azi-
muth) as time-series data.
Preprocessing
After data acquisition, preprocessing such as resam-
pling [6, 7, 27], noise filtering [26], rotation normali-
zation [7 ], and size normalization [25] is conducted to
suppress insignificant information that is expression of
random variation, but not of individual signature dy-
namics, and to even out differences between data cap-
tured with different capture devices.
Feature Extraction
This process extracts discriminative features from the
acquired data. The features should allow to distinguish
a genuine signature from forgeries and be suitable for
automatic comparison.
Comparison (Similarity/Dissimilarity Measure)
In the next step, similarity/dissimilarity scores between
the extracted features for the input signature and a
reference associated with the claimed identity are com-
puted. On-line signature verification algorithms can be
classified into parameter-based and function-based
approaches, depending on the features compared [3].
In the parameter-based approach, N features, such
as average writing speed, total signature duration, and
the number of pen ups/pen downs, are extracted as
parameters that represent the signature data [26, 28 ].
In this approach, signature data Sig
para
after feature
extraction can be represented as an N-dimensional
parameter vector
Sig
para
¼ðp
1
; p
2
; :::; p
N
Þ; ð1Þ
where p
n
is the n-th extracted feature, and each feature
has a scalar value. In the parameter-based approach,
Euclidean distance, weighted Euclidean distance, or cor-
relation are computed as similarity/dissimilarity scores.
In the function-based approach, several features,
such as pen coordinates,
▶ velocity, and acceleration,
are extracted as a sequence of data. In this approach,
signature data Sig
func
after feature extraction can be
represented as
Sig
func
¼fsig
func
ðtÞg
T
t¼1
¼fðf
1
ðtÞ; f
2
ðtÞ; :::; f
N
ðtÞÞg
T
t¼1
; ð2Þ
where f
n
(t) is the n-th feature and T is the number of
sample points. Each feature is a function of t. In the
case where the features are time-series data, t stands
for a time stamp and T is the total duration. In the
function-based approach, there are different compari-
son methods:
If the reference is also a feature set, two sets of
functions are compared, namely, the reference and
the features extracted from the input signature.
These functions have different durations and are
nonlinearly distorted with respect to each other.
Then, an elastic comparison algorithm such as
dynamic time warping is applied to compute dis-
similarit y scores [7, 24, 25, 27].
If the reference is an enrollee-specific function,
such as a statistical model, similarity scores are
computed using the statistical model and the fea-
tures extracted from the input signature. For exam-
ple, the Hidden Markov Model can be used as a
statistical model, and probabilities are computed as
similarit y scores such that the input features are
reproduced from the model [26].
After the comparison process, some studies apply
score normalization [6, 26]. In the case where multiple
scores are computed, score-level fusion strategies [29 ]
are adopted in some studies [6, 7, 25, 26, 27]. In the
case where statistical models are used in score-level
Signature Recognition
S
1203
S
fusion [25], model parameters should be estimated in
the enrollment process.
Decision Making
In the decision making process, a final decision is made
as to whether or not the input signature is considered a
genuine signature or a forgery. Some researchers use
a classifier for decision making [24], though generally a
decision is made by comparing the computed scores
with a decision threshold. For example, when similarity
scores (a larger score indicate s a better match) are
used, a decision is made based on
Signature ¼
Genuine signature if score > threshold
Forgery if score threshold
ð3Þ
where threshold is a decision threshold. A decision-
level fusion strategy [29] can also be used in combina-
tion with this decision rule [ 28].
The decision result can be changed by adjusting the
decision threshold value; thus, the decision threshold
should be set according to the expected security level.
Two types of decision threshold are possible: a writer-
dependent threshold (user-dependent threshold) and a
common threshold (global threshold). It is reported
that a writer-dependent threshold yields better results
[6]; however, setting a good user-dependent threshold
for each user is a difficult problem. In both types, some
parameters estimated in the enrollment process are
used to set the decision threshold.
Enrollment
Signatures input during enrollment are called ▶ train-
ing signatures. Feature sets are extracted from the
training signatures, and these extracted feature sets
are used in the enrollment process.
For the parameter-based comparison approach, the
extracted feature sets are directly enrolled as reference
data together with an identifier [28].
For the function-based comparison approach, two
types of enrollment are possible depending on the com-
parison approach adopted, namely, reference-based and
model-based [26]. In reference-based enrollment, the
feature sets extracted from training signatures are
directly enrolled as reference data together with
the identifier [6, 7, 25, 27]. In model-based enrollment,
the extracted features sets are used to estimate statisti-
cal models, and these models are enrolled as reference
models f(·;Y) together with the identifier [26]. Here, f
(·;Y) is a statistical model and Y is a parameter set of
the statistical model.
Summary
Signature recognition is a historically established
method for the authentication of an individual or a
document. It is recognized as a common technique
worldwide and therefore, is willingly accepted as a
means of person authentication by ordinary people.
Moreover, signature verification has the notable fea-
ture that the signature can be modified in the event
that it is compromised. However, since a signature
image can be easily copied by other people and there
are many people whose signatures are quite variable,
the capability of off-line signature recognition as an
exclusive means of person authentication is limited.
Many pen input devices, such as tablet PC’s and
PDA’s, are now available that allow to capture the
signature dynamics, which is more difficult to copy
than signature images. On-line signature verification
is thus a promising candidate for person authentica-
tion methods. However, the performance of on-line
signature verification needs to be improved before it
can be used for high-security applications.
Related Entries
▶ Biometrics, Overview
▶ Performance Evaluation
References
1. ISO TC 68/SC 2: Financial services – Biometrics – Security
framework. International Standard ISO 19092 (2008)
2. Plamondon, R. ed.: Progress in automatic signature verification.
World Scientific, Singapore (1994)
3. Plamondon, R., Lorette, G.: Automatic signature verification
and writer identification – the state of the art. Pattern Recognit.
22, 107–131 (1989)
4. Deng, P., Jaw, L., Wang, J., Tung, C.: Trace copy forgery detection
for handwritten signature verification. In: 37th Annual 2003
International Carnahan Conference on Security Technology,
pp. 450–455 (2003)
5. Yeung, D., Chang, H., Xiong, Y., George, S., Kashi, R., Matsumoto,
T., Rigoll, G.: SVC2004: First international signature verification
competition. Number 3072 in Lecture Notes in Computer Sci-
ence, Springer, Hong Kong, China 16–22 (2004)
1204
S
Signature Recognition
6. Jain, A., Griess, F., Connell, S.: On-line signature verification.
Pattern Recognit. 35, 2963–2972 (2002)
7. Munich, M., Perona, P.: Visual identification by signature
tracking. IEEE Trans. Pattern Anal. Mach. Intell. 25, 200–217
(2003)
8. Ortega-Garcia, J., Fie
´
rrez-Aguilar, J., Simon, D., Gonzalez, J.,
Faundez-Zanuy, M., Espinosa, V., Satue, A., Hernaez, I., Igarza,
J.J., Vivaracho, C., Escudero, D., Moro, Q.I.: MCYT baseline
corpus: a bimodal biometric database. IEEE Proc. Vis. Image
Signal Process. 150, 395–401 (2003)
9. Garcia-Salicetti, S., Beumier, C., Chollet, G., Dorizzi, B.,
les Jardins, J., Lunter, J., Ni, Y., Petrovska-Delacre
´
taz, D.: Biomet:
A multimodal person authentication database including face,
voice, fingerprint, hand and signature modalities. Number
2688 in Lecture Notes in Computer Science, Springer, Guildford,
UK 845–853 (2003)
10. Dumas, B., Pugin, C., Hennebert, J., Petrovska-Delacre
´
taz, D.,
Humm, A., Eve
´
quoz, F., Ingold, R., Rotz, D.: Myidea – multi-
modal biometrics database, description of acquisition protocols.
59–62 (2005)
11. Sabourin, R.: Off-line signature verification: Recent advances
and perspectives. [30] 84–98
12. Leclerc, F., Plamondon, R.: Automatic signature verification and
writer verification: The state of the art 1889–1993. Int. J. Pattern
Recognit. Artif. Intell. 8, 643–660 (1993)
13. Impedovo, S., Simon, J., eds.: 1From Pixels to Features III: Fron-
tiers in Handwriting Recognition. North Holland, Amsterdam
(1992)
14. Yoshimura, I., Yoshimura, M.: Off-line verification of japanese
signatures after elimination of background patterns. Int. J. Pat-
tern Recognit. Artif. Intell. 8 , 53–68 (1994)
15. Yoshimura, I., Yoshimura, M.: An application of the sequential
dynamic programming matching method to off-line signature
verification. [30] 299–310
16. Justino, E., Bortolozzi, F., Sabourin, R.: A comparison of SVM
and HMM classifiers in the off-line signature verification. Pat-
tern Recognit. Lett. 26, 1377–1385 (2005)
17. Justino, E., Bortolozzi, F., Sabourin, R.: Off-line signature verifi-
cation using HMM for random, simple and skilled forgeries.
1031–1034 (2001)
18. Kalera, M., Srihari, S., Xu, A.: Off-line signature verification and
idenitification using distance statistics. Int. J. Pattern Recognit.
Artif. Intell. 18, 1339–1360 (2004)
19. Hairong, L., Wang, W., Wang, C., Zhuo, Q.: Off-line signature
verification and forgery detection using fuzzy modeling. Pattern
Recognit. 38, 341–356 (2005)
20. Hanmandlu, M., Yusof, M., Madasu, V.: Off-line chinese signa-
ture verification based on support vector machines. Pattern
Recognit. Lett. 26, 2390–2399 (2005)
21. Ismail, M., Gad, S.: Off-line arabic signature recognition and
verification. Pattern Recognit. 33, 1727–1704 (2000)
22. Yoshimura, I., Yoshimura, M.: Evaluation of signature quality
as a function of nationality via an off-line signature verification
system. Intell. Automat. Soft Comput. 7, 195–203 (2001)
23. Plamondon, R., Srihari, S.: Online and off-line handwriting
recognition: a comprehensive survey. IEEE Trans. Pattern Anal.
Mach. Intell. 22, 63–84 (2000)
24. Kholmatov, S., Yanikoglu, B.: Identity authentication using im-
proved online signature verification method. Pattern Recognit.
Lett. 26, 2400–2408 (2005)
25. Muramatsu, D., Kondo, M., Sasaki, M., Tachibana, S.,
Matsumoto, T.: A marvkov chain monte carlo algorithm for
bayesian dynamic signature verification. IEEE Trans. Inform.
Forensic. Secur. 1, 22–34 (2006)
26. Fie
´
rrez-Aguilar, J., Ortega-Garcia, J.: On-line signature verifica-
tion. In Jain, A., Flynn, P., Ross, A. (eds.): Handbook of
Biometrics, pp. 189–209. Springer, New York (2008)
27. Nalwa, V.: Automatic on-line signature verification. Proc. IEEE
85, 215–239 (1997)
28. Lee, L., Berger, T., Aviczer, E.: Reliable on-line human signature
verification systems. IEEE Trans. Pattern Anal. Mach. Intell. 18,
643–647 (1996)
29. Ross, A., Nandakumar, K., Jain, A. (eds.): Handbook of Multi-
biometrics. Springer, New York (2006)
30. Nabeel, N., Bortolozzi, F. (eds.): Advances in Document Image
Analysis. First Brazilian Symposium, BSDIA Number 1339 in
Lecture Notes in Computer Science, Springer, Curitiba, Brazil
(1997)
Signature Sample Synthesis
LIANG WAN
1
,ZHOUCHE N LIN
2
1
City University of Hong Kong, Hong Kong, China
2
Microsoft Research Asia, Beijing, China
Synonyms
Signature synthesis; Handwriting synthesis; Handwrit-
ing sample synthesis
Definition
Signature sample synthesis is the generation of synthetic
signature from a user’s signature samples. It is a special
case of handwriting sample synthesis which generates
novel handwriting in a particular person’s handwriting
style. A handwriting or signature synthesis system has
two basic modules: the modeling module and the
synthesis module. In the modeling module, the system
collects handwriting/signature samples of a specific
writer (online or offline), and identifies and stores
the basic characteristics of samples (for example,
shape and spatial layout). In the synthesis module,
Signature Sample Synthesis
S
1205
S
synthetic ▶ glyphs are generated from the stored tem-
plates, and they may be further aligned and connected
to form synthetic handwriting data.
Introduction
Like many biometric characteristics, such as face, fin-
gerprint, and iris, signature/handw riting has been
widely accepted by people as an effective way to iden-
tify a specific writer. Historically, signatures and forged
handwritings have always been of interest to forensic
experts. Many signature verification technologies have
been reported in the literature to detect handwriting
forgeries [1, 2]. Signature sample synthesis [3], as an
inverse biometrics problem, is the process of generat-
ing synthetic signatures that mimic real signature sam-
ples. The literature on signature sample sy nthesis is
quite rare. However, it is a special case of handwriting
sample synthesis [4, 5], which generates novel artificial
handwriting in a person’s handwriting style. The article
begins with handwriting sample synthesis techniques
and then narrows down to signature sample synthesis.
Handwriting sample synthesis has become active in
recent years, because the flourish of pen-based devices,
such as Tablet PCs, touch-screen mobile phones, per-
sonal digital assistants (PDAs), and electronic white-
boards, has brought users more natural communication
ways in human–computer interaction. In many situa-
tions, writing with a pen on the screen is more conve-
nient than typing on the keyboard. Yet, many users
find that keyboards are more efficient than handwrit-
ing because typing is faster than writing, and his/her
handwriting may become illegible after long-time
writing. Handwriting sample synthesis addresses this
dilemma by converting ASCII text to handwriting that
is close to the user’s personal handwriting . For those
people who prefer handwriting personal letters, greet-
ings, and compliments, handwriting sample synthesis
adds a personal touch to communications. Like wall-
papers and favorite software settings, synthesized
handwriting also contributes to the personalization of
one’s computing devices. Moreover, it can always gen-
erate legible handwriting and free the user from
lengthy and stressful writing, for example, while pre-
paring many hand-written documents such as greeting
cards with different content [6].
Handwriting sample synthesis is helpful to build a
signature/handwriting recognizer which heavily depends
on the size and quality of the training set [7]. It can
generate a large database of handwriting/signature
samples that look natural. This not only greatly
reduces the manual intervention in the preparation
of hand-written sample, but also provides the ability
to perform operational testing in a laboratory environ-
ment. The automated synthesis is also useful to evalu-
ate existing signature/handwriting verification methods
[8], including the accuracy and reliability against
fraudulent signature/handwriting. In addition, this tech-
nique can help forensic examiners [9] to understand
the key factors that affect a person’s handwriting or
signature.
Characteristics of Handwriting/Signature
As a behaviorial biometrics characteristic, handwrit-
ing/signature is affected by various factors which the
synthesis process should consider. For instance, the
signatures are quite different when the specific writer
writes in different languages (the article mainly con-
siders the Eng lish language, which contains a small set
of single characters, and the combination of individual
characters in a linear fashion forms various words).
Writing with different digital pen devices can also
cause direct changes in the appearance of handwriting.
Likewise, the person’s mood, his/her hand health and
the surrounding environment are also possible factors
to affect the acquisition process of handwriting sam-
ples. Despite those factors, signatures from a single
individual tend to be different even using the same
digitizing device. For different people, the character
shapes can vary greatly and the amount of shape varia-
tion may also differ from person to person. In addi-
tion, people may tend to write handwriting/signature
in a
▶ cursive style or in a partially cursive and partially
▶ handprint st yle, which makes the problem of synthe-
sis more difficult.
As suggested by handwriting analysis techniques in
forensic inspection [10], the specific features that are
easily noticeable to ordinary people to distinguish dif-
ferent handwriting styles include: (1) the glyph and the
size of single characters; (2) the pressure distribution
and the slant of handwriting; (3) the relative sizes of
the middle, the upper, and the lower zones of letters;
(4) the existence and the shape of head, connecting,
and tail parts; (5) the letter, the word, and the line
spacings; (6) the embellishment in strokes or character
1206
S
Signature Sample Synthesis
glyphs; and (7) the simplified or neglected strokes.
These features can be roughly classified into three
types: features of character gly phs, spatial layout of
characters, and connection between characters. Differ-
ent from the English language, the oriental languages,
such as Korean and Indian, often contain a large num-
ber of characters that share the same small set of
strokes, and the characters are usually unconnected.
Therefore, the handwriting characteristics of the ori-
ental languages are depicted by two types of features:
features of strokes and spatial layout of strokes to form
characters [11, 12]. For simplicity, the following dis-
cussion focuses on the synthesis of English scripts.
Modeling Process
To generate signatures that look natural, it is important
to model the characteristics of signatures/handwriting
samples. Figure 1 shows the general outline of the
modeling process. It collects handwriting data of a
specific user and learns his/her handwriting model. In
the modeling process, the handwriting samples are first
obtained by acquisition devices. For example, digitiz-
ing pen-based devices capture the handwriting sample
by a sequence of discrete 2D points. Kinematic infor-
mation such as pressure and duration of writing can be
recorded during the acquisition. Handwriting samples
may also be acquired via scanning the off-line sample
images, but all kinematic charac teristics will be lost.
The general handwriting model consists of three
parts: a character model which captures the shapes and
variations in single characters; an alignment model
which controls the spatial layout of the individual
characters that form words; and a connection model
which simulates how two characters are connected
together. The simplest form to represent the signa-
ture/handwriting is a planar curve. Hence, the char-
acter model often extracts a set of control points
to represent the hand-written character glyph [1, 3,
13–17]. Bezier curves or polynomials are then used
for curve approximation. The shape variation such as
scale, position, and slant can be learnt from multiple
samples of one character. Besides the geometric infor-
mation, physically plausible models have been pro-
posed to model the speed and acceleration in the
writing [4, 5]. These kinematic-based models are cap-
able of representing, compressing and reconstructing
input handwriting data, but they do not target at
synthesizing new handwriting. In contrast, geometry-
based models can generate handwriting with natural
shape variations and sup port different handwriting
styles, i.e., from handprint style to fully cursive style.
The alignment model records the horizontal letter
spacings and the relative vertical positions of charac-
ters with respect to a horizontal baseline [17]. The
connection model may record which character pairs
are likely to connect to each other. It may also extract
the distributio n of concatenation strokes which are
formed by the tail and the head parts of adjacent
characters [14].
In the modeling process, the system usually
requires the users to provide adequate handwriting
samples, so a practical concern is to keep user involve-
ment at a reasonable level. In fact, the burden of user
involvement in the sample collection process depends
upon the handwriting models that the systems use. For
example, users are asked to write more than one thou-
sand letter groups in [13]. Besides writing 80–200
words, the work in [14, 15] may need user interaction
Signature Sample Synthesis. Figure 1 Block diagram of the model training process.
Signature Sample Synthesis
S
1207
S
in order to get good segmentation results. In contrast,
the user is only required to input each single character
three times, several special pairs of letters and several
multiletter words (Fig. 2)in[17].
Synthesis Process
The synthesis process should have synthetic data visu-
ally similar to the samples and incorporate sufficient
variability in synthetic data. This objective, however, is
not easily achieved, especially when considering cur-
sive writing styles. Figure 3 shows the basic flowchart
of handwriting synthesis systems. For an input ASCII
text, each individual character glyph is first generated
from the character model. Then the g lyphs are
arranged and adjacent characters are connected when
needed to form a cursive word. The words are further
aligned into lines and paragr aphs. In the following
section, each step of synthesis is detailed.
Character selection generates glyphs based on the
stored character models. In [4], the handwriting
Signature Sample Synthesis. Figure 2 The user interface in [17] to collect user handwriting samples: samples of
lowercase letters, capital letters, digits, punctuations, special letter pairs, and multi-letter words.
Signature Sample Synthesis. Figure 3 Block diagram of the handwriting sample synthesis process.
1208
S
Signature Sample Synthesis
sample is modeled by the writing velocity equations.
Then a similar synthetic data are generated by setting
initial conditions to the equations. In [14, 15], the
character model is the statistical distribution of control
points that are learnt from multiple gly ph samples for
each character. By randomly sampling the distribution ,
new handwriting trajectory can be generated. In [17],
three instances of characters are stored as the templates
which are supposed to appear at the beginning, the
middle, and the end of words, respectively. Then accord-
ing to the character position in the word, one of the
three templates is chosen as the initial glyph. After
selection, affine transformation (scale, rotation and
slant) is applied to the character glyphs. To mimic the
variability in the natural handwriting/signature, some
randomness is often added in the transformation.
Character alignment then places the glyphs with
respect to the baseline, both vertically and horizontally.
Vertical alignment is necessary for the generation of
smooth handwriting. For example, middle-zone letters
(‘‘a,’’ ‘‘c,’’ ‘‘e,’’ etc.), ascendent letters (‘‘b,’’ ‘‘d,’’ etc.),
and digits are expected to have their glyph bottoms
meet the baseline. Horizontal alignment, on the other
hand, separate s the bodies of adjacent letters at a dis-
tance along the horizontal baseline. However, the adja-
cent character gly phs may have severe overlapping
when their head or the tail parts are too long. As a
result, the synthesized handwriting may look weird or
it may be hard to produce smooth connection part.
Hence, redundant portions of heads/tails may be
trimmed to alleviate the overlapping problem [17].
Character connection is developed to simulate cur-
sive handwriting, in which adjacent letters are connected
by smooth ligature parts (the head and tail parts of
adjacent characters). In [14], a statistical ligature gen-
eration model is learnt from handwriting samples.
when synthesis, a ligature stroke is generated for
every pair of adjacent letters. The final ligature part is
determined by jointly deforming the letter strokes and
the ligature stroke. This method needs sufficient sam-
ples for training. In [15], a delta log-normal model is
employed to represent the head/tail parts as pieces of
arcs. By changing the arc parameters, the trajectories of
letters are deformed to create a smooth ligature part.
This model may be interfered by too long overlaps
between head/tail parts. The work [17] adopts a high-
order polynomial to fit the ligature part. The problem
becomes to determine the control points given the
head/tail parts of adjacent glyphs. To solve the fitting
problem, three constraints are imposed on the liga ture:
similarity to the original ligature, deformation energy
from the original ligature, and smoothness of the liga-
ture. The control points should minimize the sum of
these three energy terms.
After the processes mentioned above, an ASCII
word is converted to a hand-written word. Then it is
natural to synthe size handwriting text by rendering
multiple words one by one. Figure 4 shows an example
of the communication via emails with handwriting
synthesis. The sender types in a text email. The hand-
writing synthesizer automatically converts it into a
handwriting email and sends it to the receiver. Then
Signature Sample Synthesis. Figure 4 Integration of handwriting sample synthesis system with Microsoft
1
Office
Outlook
1
.
Signature Sample Synthesis
S
1209
S