WRITER VERIFICATION BASED ON GRAPHOMETRIC

FEATURES USING FEED-FORWARD NEURAL NETWORK

Carlos F. Romero, Carlos M. Travieso, Jesús B. Alonso and Miguel A. Ferrer

Signals and Communications Department, Technological Centre for Innovation on Communication (CeTIC)

University of Las Palmas de Gran Canaria,Campus of Tafira, Ed. Telecomunicación

Pabellón B. 35017, Las Palmas de G.C., Spain

Keywords: Biometric systems, Writer verification, Handwriting analysis, Neural networks, Pattern recognition.

Abstract: This paper shows a writer verification automatic system based on a set of graphometric characteristics

extracted from handwritten words. That dataset has been tested with our off-line handwritten database,

which consists of 110 writers with 10 samples per writer, where a sample is a dataset of 34 words. After our

experiments, we have got a verification success rate of 95.63% and Equal Error Rate (EER) of 3.90% is

achieved. For previous results, we have used as classifiers a Neural Network, for each writer.

1 INTRODUCTION

The Biometric Recognition Systems have

proliferated greatly in the last decade. Its objective is

to substitute or to change the conventional key

(passwords, cards, PINs, etc.) for the individual's

innate keys. Nowadays, the writing remains of great

importance to society because of its wide and

extended use in various activities of persons. For this

reason, the writer verification on the basis of

scanned handwriting images is a very useful

biometric modality.

Writer verification involves a one-to-one

comparison with a decision as to whether or not the

two samples are written by the same person. In this

paper, we use Neural Network based handwriting

recognition systems for the purpose of writer

verification. Thus for N different writers, we obtain

N different Neural Networks (NNs) (Bishop, 1995),

(Juang and Rabiner, 1992), each NN can be

understood as an expert specialized in recognizing

the handwriting of one particular person.

The verification is possible because our proposal

extracts information biometric of the writing. The

scientific bases for this idea come from the brain

human. If we try to do writing with the less skilful

hand, there will be some parts or forms very similar

to the writing with the skilful hand, due to those

orders are sent by the brain, and each brain is

intrinsic of each person (Romero et al., 2007).

The act to write is a phenomenon governed by

the brain and integrated in the psychomotricity of

the individual; in contrast to mimic movements, the

handwriting movements are fixed toward a plane

that allows its study and measurement. The writing,

like codified and dynamic message that reflects

certain biometric information of the individual in

his/her communication with the others, is

fundamentally individual, recognizable, univocal

and unique; that makes possible the user

identification.

Recently, different approaches to writer

verification have been proposed in (Cha and Srihari,

2000), they addressed the problem of writer

verification, i.e., the problem of determining

whether two documents are written by the same

person or not. In order to identify the writer of a

given document, they model the problem as a

classification problem with two classes, authorship

and non-authorship. Given two handwriting

samples, one of known and the other of unknown

identity, the distance between two documents is

computed. Then, the distance value is used to

classify the data as positive or negative.

(Leedham and Chachra, 2003) present a set of

eleven features which can be extracted and used for

the identification and verification of documents

containing handwritten digits. These features are

represented as vectors and using the Hamming

distance measure and determining a threshold value

353

F. Romero C., Travieso C., Alonso J. and Ferrer M. (2010).

WRITER VERIFICATION BASED ON GRAPHOMETRIC FEATURES USING FEED-FORWARD NEURAL NETWORK .

In Proceedings of the Third International Conference on Bio-inspired Systems and Signal Processing, pages 353-358

DOI: 10.5220/0002589003530358

 SciTePress

for the intra-author variation a high degree of

accuracy in authorship detection is achieved.

(Schalapbach and Bunke, 2004) used HMM

based recognizers for the identification and

verification, a text line of unknown origin is

presented to each of these recognizers and each one

returns a transcription that includes the log-

likelihood score for the considered input. These

scores are sorted and the resulting ranking is used

for both identification and verification. The system

must decide, based on a verification criterion,

whether a text line with a claimed identity is in fact

from this writer or if it is an impostor attempt. Using

a confidence measures to define the following

verification criterion: if the confidence measure is

minor than a certain threshold, then it assumes that

the text line is from certain writer; otherwise the

input is classified as not being of the claimed

identity.

(Bensefia et al., 2005) proposed a writer

verification building a hypothesis test, based on the

mutual information between the grapheme

distributions in the two handwritings that are

compared. Mutual information allows measuring the

independence between the two writers, for low

values of mutual information indicate a strong

independence showing that the set of features is

distributed in the same way on the two documents

and reflecting the same identity for the two writers.

While the high values denote a strong dependence,

showing different identities for the two writers.

(Bucalu and Shomaker, 2007) perform writer

verification using a distance measure between the

feature vectors to compute the similarity in

individual handwriting style, between any two

chosen samples. Distances up to a predefined

decision threshold are deemed sufficiently low in

order to consider that two samples have been written

by the same person. Otherwise, samples are

considered written by different persons.

The goal of this paper is to introduce a writer

verification system from isolated words. The rest of

the paper is organized with the following sections. In

sections 2 and 3 are shown the methodology of this

work and the building of our database, respectively.

In section 4 is briefly described the image pre-

processing and the segmentation of the words. In the

following section, the procedure for the extraction of

the characteristics is explained. Section 6 includes

the classification system based on Neural Networks.

In section 7 is shown verification tasks, and the

following section can be observed conclusions. And

finally, this present work ends with

Acknowledgement and references.

2 DESCRIPTION OF THE

SYSTEM PROPOSED

As the majority of the works proposed up to now

(see Figure 1), on biometric recognition, the

framework of the system depends on the following

basic steps.

Figure 1: Proposal of our writer verification system.

1. Image pre-processing and segmentation.

Preparation and modification of images, so that

the module of segmentation produce the results

desired. The segmentation separates the zones

of interest (lines, words or characters), and it is

key for the success or error of the following

analysis (Feature Extraction).

2. Feature Extraction. This is the calculation of

qualitative and quantitative measures that

permit to obtain a significant geometrical

characterization of the style of writing, in order

to differentiate writers among themselves.

Pressure, speed, direction, inclination, cohesion,

continuity, opening of ovals constitutes some of

the graphologist features. The aspect of the

letters, its form and dimension is some of the

classical features. The graphologist features

help to be more discriminating. In this present

work, the most of our graphologist features are

Acquire Images

Pre-

rocessing

Segmentation

Features Extraction

Verification

(Neural Networks)

TRAINING MODE

-line

Database

Models

Decision

Accept /

Reject

TEST MODE

BIOSIGNALS 2010 - International Conference on Bio-inspired Systems and Signal Processing

354

automatically got. In future works, we hope to

reach it.

3. Classification: A statistical supervised analysis

of the extracted characteristics is carried out,

which will permit the comparison with the

samples of our database, if the query sample

possess similarities is accepted, otherwise is

rejected.

3 DATABASE

For the building of our database, we have used a

paragraph of 15 lines,the language used is Spanish.

That text is from “Don Quijote de la Mancha” from

Miguel de Cervantes, and we have used the same

text for each writer. With this size of text, writers

can show their personal characteristics, because they

keep their writing habits. This database has been

built with 110 writers, and each one has made 10

times this template (paragraph of 15 lines). The size

of paper was DIN-A4 format (297 mm. x 210 mm).

The sheet was written with a pen of black ink. Each

writer of our database had one week for doing the

writing, and therefore, it is considered like an effect

of temporal invariance on this database.

The creation conditions of our database were the

normalized with the same type of paper (80 gr/m2),

ballpoint pen, and similar place of support (for doing

the writing). Of this way, our work is centred on the

writing and the efficiency of proposed parameters.

In the future work, we are going to change the rest

of variables.

The samples are scanned with 200 dpi, obtaining

images on grey scale, with 8 bit of quantification

(see figure 2).

4 IMAGE PRE-PROCESSING

AND SEGMENTATION

The first step of the image pre-processing consists of

utilizing Otsu’s the method, in order to get the

binarization of the samples (Otsu, 1979).

As a result of the binarization, in most cases, the

line of writing remains with irregular aspect. For that

reason, we have implemented other image

preprocessing for skew elimination. We developed

an algorithm in order to detect the maximum

projection of the word, using histogram tool. This

permit to smooth out the baseline, so, the baseline

remains well defined. Besides, it is eliminated the

existing noise in the images after scanning process,

by morphological mathematics.

As previous step to the separation of words or

components connected (by labelling), the detection

and elimination of the punctuation marks (points,

accents and comma) is carried out, by a size

threshold, In particular, we have removed

components connected if they were minor than 60

pixels.

Finally, it is segmented words, which compose

the lines of writings (baselines) and for it, it is must

establish limits of each one of the words. For this

estimation, the method of the “Enclosed Boxes”

(Jaekyu and Haralick, 1995) was used, which

provides us the coordinates that will permit segment

the words. The enclosed boxes are defined as the

most minimum rectangle that contains to the

component connected.

Figure 2: Example of the words extraction for a sample.

5 FEATURE EXTRACTION

Writer verification and identification methods fall

into two broad categories: text-dependent versus

text-independent methods. If any text may be used to

establish the identity of the writer the system is text

independent. Otherwise, if a writer has to write a

particular predefined text to identity the verification

task is text dependent.

WRITER VERIFICATION BASED ON GRAPHOMETRIC FEATURES USING FEED-FORWARD NEURAL

NETWORK

355

This present work has used a selection of the

most discriminative features from (Romero et al.

2009), (Hertel and Bunke. 2003), (Srihari et al.,

2001), which are depending on a set of 34 specific

words. These features were separated into two

groups, one set of dependent text feature and other

set of independent text feature.

Handwriting cohesion is called to the percentage

of unions that appear between the letters of the same

one; when saying unions, we are talking about the

final strokes of the letters, continuing with the

initials of the following letters without ballpoint pen

is risen of the paper.

In order to make an estimation of the cohesion in

the handwriting, the images of the 34 words of each

sample are selected and binarized and their

components connected with connectivity-8 are

labelled to them. As soon as the quantity of

components connected of each word is obtained, it is

proceeded to calculate the average and the variance

of the components connected. Those words have

been selected by their size, the largest.

The estimation of the width of letters is carried

out, seeking the row with greater quantity of black to

white transition (0 to 1). It is counted the number of

white pixels between each transition, this result is

averaged.

In order to measure the height of the medium

body of the words, the goal is to determine the upper

and lower baseline through maximums and

minimum values and to measure the distance among

them.

In order to approach baselines of each word, it

was decided to use the adjustment of minimum

mean square error that is based on find the equation

(see expression 1) that better be adjusted to an set of

points "n" (Morita et al., 2001). The equation is the

following:

baxy +=

(1)

where the coefficients “a” and “b” determine the

lineal polynomial regression by means of the

following expressions:

∑∑

∑∑∑

===

⎟

⎠

⎞

⎜

⎝

⎛

−

⎟

⎠

⎞

⎜

⎝

⎛

⎟

⎠

⎞

⎜

⎝

⎛

−

xxn

yxyxn

111

(2)

xay

∑∑

−

(3)

Those values of “a” and “b”, based on the

coordinates of minimums or maximums detected in

the contour of the word, are different baselines.

Minimums are to approach the lower baseline and

the maximums for the superior baseline (see figure

3).

Figure 3: Ovals and loops with its respective major and

minor axes.

For the estimation of pressure by means of the

width of the strokes, the run lengths are determined

on the binary image taking into consideration black

pixels corresponding to the ink trace. There two

basic scanning methods: horizontal along the rows

of the image and vertical along the columns of the

image.

The area of thinning is defined as the area

remaining in the resulting image of the word after it

undergoes a process of three morphological stages:

thinning, dilation and thinning again.

The following features are dependent text: length

of the words, quantity of pixels in black, heights of

the ascending and descending, height relation

between descending and ascending, height relation

between descending/ascending and medium body.

As for the analysis of the ovals and loops of the

words, is done the measurement of the minor and

major axes of each ovals and loops from 34 words.

Axes are calculated by maximum projection using

histograms. Finally, it is calculated the average size

of the above mentioned axes of the handwriting

sample in analysis.

6 NEURAL NETWORKS

As classifier, we have used a Feed-Forward Neural

Network (NN) with a Back-propagation algorithm

for training (Bishop, 1995) (Juang et al., 1992),

where the number of input neurons is given by the

dimension of the vector of features with 374

parameters (11 parameter x 34 words). And the

BIOSIGNALS 2010 - International Conference on Bio-inspired Systems and Signal Processing

356

number of output neurons is given by the number of

writers to identify.

Too, we have researched and checked this

system varying the number of neurons in the hidden

layer in order to improve the success rate.

Finally, we have used the method of the ‘more

voted’ algorithm, where we have built a schedule

with 10 neural networks (see figure 4), and we have

reached a improvement between 3 and 4 points on

success rates

Figure 4: Classification System with ‘more voted’

algorithm, based on NN.

7 WRITER VERIFICATION

EXPERIMENTS

The verification can be seen as a problem of

classification of two classes. When are compared the

samples: the variation of a writer among their own

samples should be smaller than the variation among

samples of two different writers.

In order to give solution to this problem, the

methodology of the used verification was

independent supervised classification. Therefore, we

have a system with two modes, training and test

modes (see Figure 1).

For the training, we have used the 50% of our

database, and the remainder to carry out the test

mode. That is, five samples have been chosen to

training and other five for the test, since we have 10

samples for each writer. Besides, a total of 34 words

have been extracted from paragraphs, and there will

be 34 words by sample.

Like our parameters depend on words used and

its writer, we have used the same word for this

process; in particular, we have used a set of 34

words from the paragraph of 15 lines. But, the

samples set for training and test mode of these 34

words are different, being obtained from 10 different

samples of each writer. Therefore, this system works

with a close set of words (34 words).

In order to calculate the characteristics, we have

used 170 words (34 x 5 samples/writer) on the

training process. The criterion of selection to choose

the previous 34 words was their length, upper than 5

letters, because with this length, they offer

information more general than a word with a shorter

size.

For writer verification, we employed a threshold

to decide if the samples query is accept or reject. By

varying the threshold, a Receiver Operating

Characteristics (ROC) curve (see Figure 5) is

obtained that illustrates the inevitable trade-off

between the False Acceptance Rate (FAR) and False

Rejection Rate (FRR). The Equal Error Rate (EER)

corresponds to the point on the ROC curve where

FAR=FRR and it quantifies in a single number the

writer verification performance.

To analyze the performance of our system, we

used dependent text feature, independent text feature

and combination of both. The results of the

verification experiments are given in Table 1.

The best ROC curve is produced using

combination features with a threshold of 0.027

yielding the following results: nearly 96% of correct

verification with an EER of about 3.90%, FAR of

4.23% and FRR of 4.37%.

Table 1: Writer verification performances.

Independent

Text

Dependent

Text

Combination

Threshold 0.030 0.027 0.027

EER 8.10 % 6.10 % 3.90 %

FAR 8.35 % 6.17 % 4.23 %

FRR 8.19 % 6.19 % 4.37 %

Success

Rate

92.23 % 93.82 % 95.70 %

8 CONCLUSIONS

In this paper, we have presented a system that uses

NNs for the task of off-line writer verification. The

basic input units presented to the system are

handwritten words. We have tested our system on

the verification task using independent and

dependent text features are extracted from each

word.

Neural Network

‘More

Voted’

Algorithm

Neural Network

WRITER VERIFICATION BASED ON GRAPHOMETRIC FEATURES USING FEED-FORWARD NEURAL

NETWORK

357

0 5 10 15 20 25 30 35 40

FRR

FAR

Combination Feature

Dependent Text Feature

Independet Text Feature

Figure 5: ROC curve of the verification experiment.

Currently, our better performances have been

combining independent and dependent text features.

The system performs very well on both tasks of

accept and reject. An Equal Error Rate (EER) of

about 3.90% is achieved.

In the future work, we plan to address at

improving the performance independent text feature.

We also plan to expand our database and test our

system with other databases such as the IAM

database.

ACKNOWLEDGMENTS

This present work has been supported by private

funds from Spanish Company, Telefónica. This

research project is called “Cátedra Teléfonica-

ULPGC 2009”; and by funds from Research Action

from Excellent Networks on Biomedicine and

Environment belong to ULPGC.

REFERENCES

Bishop, C.B., 1995. Neural Networks for Pattern

Recognition, Oxford University Press.

Juang, B.H., Rabiner, L.R., 1992. Spectral representations

for speech recognition by neural networks-a tutorial.

In Proceedings of the Workshop Neural Networks for

Signal Processing, pp. 214 – 222.

Romero, C.F., Travieso, C. M., Alonso, J. B., Ferrer, M.

A., 2007. Using Off-line Handwritten Text for Writer

Identification. In WSEAS Transactions on Signal

Processing. Issue 1, Vol. 3, pp. 56-61.

Cha, S. H., Srihari, S., 2000. Multiple feature integration

for writer verification. In Proceedings Seventh

Inernational. Workshop on Frontiers in Handwriting

Recognition, pages 333-342.

Leedham, G., Chachra, S., 2003. Writer Identification

using Innovate Binarised Features of Handwritten

Numerals. In Proceeding of the 7th International

Conference on Document Analysis and Recognition.

Vol. 1, pp. 413-416.

Schlapbach, A., Bunke, H., 2004. Using HMM Based

Recognizers for Writer Identification and Verification.

In Proceeding of the 9th Int’l Workshop on Frontiers

in Handwriting Recognition

Bensefia, A., Paquet, T., Heutte, L., 2005. Handwritten

Document Analysis for Automatic Writer

Recognition. Electronic Letter on Computer Vision

and Image Analysis, pp. 72-86.

Bucalu, M., Shomaker, L., 2007. Text-Independent Writer

Identification and Verification Using Textual and

Allographic Features. In IEEE Transactions on

Pattern Analysis and Machine Intelligence, Vol. 29,

No. 4.

Otsu, N., 1979. A threshold selection method from

graylevel histograms. In IEEE Transaction on

Systems, Man and Cybernetics, Vol. 9, Is. 1, pp 62-66.

Jaekyu Ha, R. M., Haralick, I. T., 1995. Document page

decomposition by the bounding-box project, In

Proceedings International Conference on Document

Analysis and Recognition, Vol. 02, No. 2, p. 1119.

Romero, C.F., Travieso, C. M., Alonso, J. B., Ferrer, M.

A., 2009. Hybrid Parameterization System for Writer

Identification. In Second International Conference on

Bio-inspired Systems and Signal Processing, pp. 449-

454.

Hertel, C., Bunke, H., 2003. A Set of Novel Features for

Writer Identification. In Proceedings of the Audio and

Video Based Biometric Person Authentication. pp.

679-687.

13. Srihari, S., Cha, S.H., Arora, H., Lee, S., 2001.

Individuality of Handwriting: A Validity Study. In

Proceedings of International Conference on Document

Analysis and Recognition pp 106-109.

14. Morita, M.E. Facon, J. Bortolozzi, F. Garnes,

S.J.A. Sabourin, R., 1999. Mathematical Morphology

and Weighted Least Squares to Correcthandwriting

Baseline Skew. In Proccedings of the 5th

International Conference on Document Analysis and

Recognition, pp. 430-433.

BIOSIGNALS 2010 - International Conference on Bio-inspired Systems and Signal Processing

358