See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/275517344

HANDWRITING RECOGNITION METHODS USING

ARTIFICIAL NEURAL NETWORKS

Conference Paper · November 2006

CITATION READS

1 1,590

1 author:

Wojciech Kacalak
Koszalin University of Technology
341 PUBLICATIONS   807 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Development of models of new porous coatings obtained on biomaterials by Plasma Electrolytic Oxidation View project

A method for grading the quality of a mechanical construction design View project

All content following this page was uploaded by Wojciech Kacalak on 27 April 2015.

The user has requested enhancement of the downloaded file.

 HANDWRITING RECOGNITION METHODS
USING ARTIFICIAL NEURAL NETWORKS

WOJCIECH KACALAK MACIEJ MAJEWSKI

Technical University of Koszalin Technical University of Koszalin
Department of Mechanical Engineering Department of Mechanical Engineering
Raclawicka 15-17, 75-620 Koszalin Raclawicka 15-17, 75-620 Koszalin
Poland Poland

ABSTRACT
We propose a new handwriting recognition method that utilizes geometric
features of letters. The paper deals with recognition of isolated handwritten
characters using an artificial neural network. The characters are written on a
regular sheet of paper using a pen, and then they are captured optically by a
scanner and processed to a binary image which is analyzed by a computer. In
this paper we present that new method for off-line handwriting recognition and
also describe our research and tests performed on the neural network.

INTRODUCTION
Handwriting recognition has been studied for nearly forty years and there
are great many proposed approaches. The problem is quite complex, and even
now there is no single approach that solves it both efficiently and completely in
all settings. In the handwriting recognition process (fig. 1), an image containing
text must be appropriately supplied and preprocessed. Next, the text must either
undergo segmentation or feature extraction. Small processed pieces of the text
will be the result, and these must undergo recognition by the system. Finally,
contextual information should be applied to the recognized symbols to verify the
result. Artificial neural networks, applied in handwriting recognition, allow for
high generalization ability and do not require deep background knowledge and
formalization to be able to solve the written language recognition problem.
PREPROCESSING FEATURE
MATCHING
EXTRACTION
digitization (by character, word Recognition of
UNDERSTANDING
binarization part, word) characters using
artificial neural and
noise elimination Segmentation (explicit
networks ACTION
or implicit)
thinning Context verification
Detection of major
normalizing from knowledge base
features (top-down
approach)

Fig. 1. Steps involved in handwriting recognition.

Handwriting recognition can be divided by its input method into two

categories: off-line handwriting recognition and on-line handwriting recognition.
For off-line recognition, the writing is usually captured optically by a scanner.
For on-line recognition, a digitizer samples the handwriting to time-sequenced
pixels as it is being written. Hence, the on-line handwriting signal contains
additional time information which is not presented in the off-line signal.

THE STATE OF THE ART

The state of the art of automatic recognition of handwriting at the beginning
of the new millennium is that as a field it is no longer an esoteric topic on the
fringes of information technology, but a mature discipline that has found many
commercial uses. On-line systems for handwriting recognition are available in
hand-held computers such as personal digital assistants. Their performance is
acceptable for processing handprinted symbols, and, when combined with
keyboard entry, a powerful method for data entry has been created.
Off-line systems are less accurate than on-line systems. However, they are
now good enough that they have a significant economic impact on for
specialized domains such as interpreting handwritten postal addresses on
envelopes and reading courtesy amounts on bank checks (Kavallieratou et al.
2002, Li and Suen 2000, Mitiche and Lebidoff 2001, Zhou et al. 2002).
The success of on-line systems makes it attractive to consider developing
off-line systems that first estimate the trajectory of the writing from off-line data
and then use on-line recognition algorithms (Nishida 1995). However, the
difficulty of recreating the temporal data (Boccignone 1993, Doermann 1995)
has led to few such feature extraction systems so far (Artieres et al. 2003).
Research on automated written language recognition dates back several
decades. Today, cleanly machine-printed text documents with simple layouts
can be recognized reliably by OCR software. There is also some success with
handwriting recognition, particularly for isolated handprinted characters and
words. For example, in the on-line case, the recently introduced personal digital
assistants have practical value. Similarly, some online signature verification
systems have been marketed over the last few years and instructional tools to
help children learn to write are beginning to emerge. Most of the off-line
successes have come in constrained domains, such as postal addresses [31], bank
checks, and census forms. The analysis of documents with complex layouts,
recognition of degraded printed text, and the recognition of running handwriting
continue to remain largely in the research arena. Some of the major research
challenges in on-line or off-line processing of handwriting are in word and line
separation, segmentation of words into characters, recognition of words when
lexicons are large, and the use of language models in aiding preprocessing and
recognition. In most applications, the machine performances are far from being
acceptable, although potential users often forget that human subjects generally
make reading mistakes (Besner and Humphreys 1991, Dori 1995, Bishop 2004).
The design of human-computer interfaces (Majewski and Kacalak 2005ab)
based on handwriting is part of a tremendous research effort together with
speech recognition (Majewski, Kacalak and Stuart 2006), language processing
and translation to facilitate communication of people with computers. From this
perspective, any successes or failures in these fields will have a great impact on
the evolution of languages (Bradford 1992, Mori 1999).
METHOD DESCRIPTION
The proposed system attempts to combine two methods for handwriting
recognition, neural networks and preprocessing for geometric features
extraction. The motivation behind that preprocessor is to reduce the
dimensionality of the neural network input. However, another benefit given by
the preprocessor is immunity against image translation, because all the
information is relative to the image’s center of mass. The system consists of the
preprocessing subsystem and the neural network subsystem, as shown in (fig. 2).

Fig. 2. The proposed system of handwriting recognition.

The handwritten text is made subject to the following preprocessing:

digitization, binarization, noise elimination, thinning, normalizing and
segmentation. The next step is to find the center of mass of the character image.
With the center of mass as a reference point, the vectors are drawn, creating a set
of points describing the contour of the character so that its pattern description is
made. The neural network training patterns are based on the geometric analysis
of letters (fig. 4). The description patterns of each isolated character are inputs
for an artificial neural network (fig. 5).

Fig. 3. a) Determining the center of mass of a character, b) determining the

approximate position of the center of an optional letter.
 The letter can be considered as a set of points which mass mk is constant,
and have gravity forces Gk=mk*g, and also vectors rk (1).
n

∑ r ∆G k k
rc = k =1 (1)
G
The center of a letter i.e. the center of mass of the character image (fig. 3) is
found by determining gravity center coordinates (2):
n n

∑ x ∆G k k ∑ y ∆G k k
xc = k =1
yc = k =1 (2)
G G
In (1) and (2) gravity force G is total weight of the set of points (3):
n
G = ∑ ∆Gk
k =1 (3)
Fig. 4. Geometric analysis of letters for geometric feature extraction.
Fig. 5. Scheme of the preprocessing method for geometric feature extraction of
characters: a) geometric analysis of letters a, b, c; b) pattern description:
histograms for a, b, c; c) analysis of letters d, e, f; d) pattern description:
histograms for d, e, f.
RESEARCH RESULTS
In the proposed system, the description patterns of each isolated character
are the input signals for the neural network as shown in (fig. 6).
all points have the state
greater than 0 Input signals of the neural network:
letter ‘a’
x1 y1
points have the state letter ‘b’ -all points have the state greater than 0
greater than 1 y2 -points have the state greater than 1
letter ‘c’
x2 y3 -group has the state greater than 1
letter ‘d’ -group is vertical
y4
letter ‘e’ -group is horizontal
groups are skew (45’) y5
letter ‘f’ -2 groups are skew (45’)
xn ym -2 groups are skew (45’)

Fig. 6. Input signals for the neural network.

The data set for this system (fig. 7) can be the “Optical Recognition of
Handwritten Digits” database, which was originally assembled by the National
Institute of Standards and Technology (NIST) in 1994. It consists of handwritten
numerals from a total of 43 people, 30 contributed to the training set database
and 13 different people to the test set database.

Fig. 7. The data set for this system.

The simulation set of the recognition of isolated handwritten characters,
built for creating and training artificial neural networks is shown in (fig. 8a).
The neural networks are trained with the model of isolated written language
characters. The applied neural network architecture is presented in (fig. 8b). The
networks consist of two layers of neurons with the competitive mechanism.
The ability of the neural network to learn to recognize specific letters
depends on the number of learning epochs. The specified time of learning
enables the network to minimize the error so that it could work more efficiently.
Based on the research, the following conclusion has been reached as shown in
(fig. 8c).
Error rate is about 20% at learning time equals 50 epochs and 5% at 100
epochs. The error rate dropped by about 90% after training with 60 series of all
patterns.
 a) Oscilloscope of
weight values
during
learning process

Activation
Power Control Control Function:
meter of system system Coefficients of Network linear,
neuron of network of network learning and threshold sigmoid
weights weights learning forgetting value (lambda factor)

Input signal character

recognized
pattern
characters value description Preprocessing
Neural network change
coefficient
isolated handwritten character models determine
system
subsystem

Input Report Report to Competition Learning

signal message file: mechanism pattern
polarization TXT WTA creation
RTF
HTML

b) c)

n
y j = f (∑ xi wij )
i =1

Fig. 8. Neural network simulations of isolated handwritten characters models,

neural network architecture and error rate.

CONCLUSIONS AND PERSPECTIVES

Many advances and changes have occurred in the field of automated written
language recognition, over the last decade. The different sources of variability of
various psychophysical aspects of the generation and perception of written
language make handwriting processing so difficult.
Considerable progress has been made in handwriting recognition
technology particularly over the last few years. Handwriting recognition systems
have been limited to small and medium vocabulary applications, since most of
them often rely on a lexicon during the recognition process. The capability of
dealing with large lexicons, however, opens up many more applications.

REFERENCES
Artieres T., Gauthier N., Gallinari P., Dorizzi B., 2003, "A Hidden Markov Models combination
framework for handwriting recognition", International Journal on Document Analysis and
Recognition, Vol. 5, No. 4, July 2003, Springer Berlin Heidelberg, pp. 233-243.
Besner D., Humphreys G. W., 1991, "Basic Processes in Reading: Visual Word Recognition",
Lawrence Erlbaum Associates, Hillsdale, New Jersey 1991.
Bishop C. M., 2004, "Neural Networks for Pattern Recognition", Oxford University Press Inc., New
York 2004.
 Bradford R., Bradford R. B., 1992, "An Introduction to Handwriting Examination and
Identification", Nelson-Hall Publishers, Chicago 1992.
Boccignone G., Chianese A., Cordella L. P., Marcelli A., 1993, "Recovering Dynamic Information
from Static Handwriting," Pattern Recognition, Vol. 26, Issue 3, March 1993, Elsevier Science,
pp. 409-418.
Doermann D. S., Rosenfeld A., 1995, "Recovery of Temporal Information from Static Images of
Handwriting", International Journal of Computer Vision, vol. 15, no. 1-2, June 1995, Springer
Netherlands, pp. 143-164.
Dori D., Bruckstein A., 1995, "Shape, Structure and Pattern Recognition", World Scientific
Publishing Co., New Jersey 1995.
Kavallieratou E., Fakotakis N., Kokkinakis G., 2002, "An unconstrained handwriting recognition
system", International Journal on Document Analysis and Recognition, Vol. 4, No. 4, July
2002, Springer Berlin Heidelberg, pp. 226-242.
Li Z. C., Suen C. Y., 2000, "The partition–combination method for recognition of handwritten
characters", Pattern Recognition Letters, Vol. 21, Issue 8, July 2000, Elsevier Science, pp. 701-
720.
Majewski M., Kacalak W., Intelligent Two-Way Speech Communication System between the
Technological Device and the Operator, Proceedings of the ANNIE 2005 Conference,
Artificial Neural Networks in Engineering ANNIE 2005, Vol. 15: Smart Engineering System
Design, St. Louis 6-9 November 2005, ASME Press, New York 2005, p. 841-850.
Majewski M., Kacalak W., Intelligent Layer of Two-Way Voice Communication of the
Technological Device with the Operator, Knowledge-Based Intelligent Information and
Engineering Systems, 9th International Conference, KES 2005, Melbourne, Australia, 14-16
September 2005, Lectures Notes in Artificial Intelligence 3683, Springer-Verlag Berlin
Heidelberg New York 2005, p. 930-936.
Majewski M., Kacalak W., Douglas Stuart K., Selected Problems of Intelligent Natural Language
Processing, International Journal of Research on Computing Science, Mexico 2006.
Mitiche A., Lebidoff M., 2001, "Pattern Classification by a Condensed Neural Network", Neural
Networks, vol. 14, Issues 4-5, May 2001, Elsevier Science, pp. 575-580.
Mori S., Nishida H., Yamada H., 1999, "Optical Character Recognition", John Wiley & Sons, Inc.,
New York, New Jersey 1999.
Nishida H., 1995, "An Approach to Integration of Off-Line and On-Line Recognition of
Handwriting", Pattern Recognition Letters, Vol. 16, No. 11, November 1995, Elsevier Science,
pp. 1213-1219.
Zhou J., Krzyzak A., Suen C. Y., 2002, "Verification-a Method of Enhancing the Recognizers of
Isolated and Touching Handwritten Numerals", Pattern Recognition, Vol. 35, Issue 5, May
2002, Elsevier Science, pp. 1179-1189.

View publication stats