Comparison of ANN and SVM to Identify Children Handwriting Difficulties

doi:10.4236/eng.2013.55B001

Paper Menu >>

Journal Menu >>

Engineering, 2013, 5, 1-5

doi:10.4236/eng.2013.55B001 Published Online May 2013 (http://www.scirp.org/journal/eng)

Comparison of ANN and SVM to Identify Children

Handwriting Difficulties

Anith Adibah Hasseim, Rubita Sudirman, Puspa Inayat Khalid, Narges Tabatabaey-Mashadi

Faculty of Electrical Engineering, Universiti Teknologi Malaysia, Johor Bahru, Johor, Malaysia

Email: [email protected]

Received 2013

ABSTRACT

This paper compares two classification methods to determine pupils who have difficulties in writing. Classification ex-

periments are made with neural network and support vector machine method separately. The samples are divided into

two groups of writers, below average printers (test group) and above average printers (control group) are applied. The

aim of this paper is to demonstrate that neural network and support vector machine can be successfully used in classi-

fying pupils with or without handwriting difficulties. Our results showed that support vector machine classifier yield

slightly better percentage than neural network classifier and it has a much stable result.

Keywords: Neural Network; Support Vector Machine; Handwriting Difficulties

1. Introduction

Handwriting is the primary form of written expression

for young elementary school students. Handwriting has

long been an effective means to record information,

transmit message and project feelings [1] for communi-

cation among people. There is evidence to indicate that at

least 10% - 30% of children who have difficulty with

handwriting and need to be resolved with the right inter-

vention [2]. Most of the studies that involve in handwrit-

ing movement only focused on children with known psy-

chological or physical problem. Yet, not all problems can

be categorized as clear cut disease and condition [3].

Various softwares have been presented for handwriting

recognition and movement analysis, but softwares di-

rectly related to child handwriting analysis with the pro-

spective of screening children in general, and addressing

difficulties are rare and the research is in its early stage.

The development of children’s writing ability is im-

portant in building self-esteem among children. For ex-

ample, in 1998 Graham and Weintraub reported that stu-

dents with poor handwriting needed twice as much time

to copy a written passage as those with good handwriting

[4]. Hence, difficulty in writing for young student can

lead to a dislike of writing, frustration with writing, and

development of a negative mind-set about writing ability.

As a result, this will truly limit their further writing de-

velopment and subsequently retarded to academic suc-

cess. Therefore, early analysis of those children will

benefit the educational system in order to provide an in-

structional handwriting program suited to their strengths

and weaknesses. It is evident that educators and mental

health experts are in need of empirically-based assess-

ment and intervention procedures to help identify and

treat children with writing disorders.

1.1. Evaluation of Handwriting Difficulties

Several studies have been done to evaluate early writing

skill in primary grade. These research findings were in-

vestigated and describes in term of legibility and speed,

see [4,5]. However, the scarcity of valid and reliable

handwriting evaluation tools, the complexity in the scor-

ing tools and the long processing time by the evaluator

who needs to judge the writing product for each of the

legibility criteria, limit the application of standardized

assessments in the evaluation of handwriting difficulties

in clinical and classroom setting [2]. In addition, in most

of the mild cases, the symptoms of handwriting difficulty

are present but normally are not recognised by the teach-

ers or certified evaluators.

In this paper we describe experiments carried out us-

ing two classification methods in classifying children

with and without handwriting problem based on drawing

tasks. In contrast with similar method known by Khalid

in [6] and other related studies [7,8] we tested each dif-

ferent feature individually and describe experiments car-

ried out using Support Vector Machine (SVM) in addi-

tion to those classification methods used in previous re-

searches. SVM is a supervised learning method that has

proven it’s efficiently over classic Neural Networks and

its subset [9]. The advantages of SVM are good gener-

A. A. HASSEIM ET AL.

alization performance, able to handle high dimensional

data and able to map the data into new high dimensional

feature space for better classification using kernel func-

tions. The aim of this paper is to present that SVM and

ANN can be effectively used as an automated system in

seeking out pupils with handwriting problem.

1.2. Artificial Neural Network (ANN)

Probably neural network methods are most widely known.

An ANN can be define as information processing con-

cept that is inspired by the way biological nervous sys-

tems, such as the brain, process information. This system

can be seen in an architecture inspired by the structure of

the cerebral cortex of the brain [10]. These processing

elements are usually organized into a sequence of layers

with entire or random connections between the layers.

Basically, ANN is divided to three layers which are an

input layer, at least one hidden layer and an output layer.

Multilayer feedforward network is the simplest of ANN

devise. It can be used to model some mapping between

sets of input and output variable with appropriate pattern

of weights. Figure 1 shows a basic diagram of feedfor-

ward neural network which can be trained using back

propagation method, supervised learning network.

Back propagation learning uses the gradient descent

procedure to modify the connection weights which is

derived from the consideration of minimizing some error

function. This error function is needed to change the

network parameter, which is advantage in improving the

network performance.

1.3. Support Vector Machine(SVM)

We now describe the basic idea of Support Vector Ma-

chine, more explanation can be found in [9,11,12]. SVMs

are new technique suitable for binary classification tasks.

In addition, SVM is one of the excellent tools for classi-

fication and regression problems with a good generaliza-

tion performance.

SVM constructs a hyperplain or a set of hyperplains to

separate the two sets of data in a feature space. The key

approach of SVM is to try finding the best hyperplain by

maximizing the minimum margin between the two sets.

An optimal separating hyperplane is shown in Figure 2.

In SVM, training vectors are mapped into higher dimen-

sional space by the function of φ which given a training

set of instance-label pairs (x i, yi); i = 1,…, l where x i є Rn

and y є {1, -1}l. All operations in learning and testing

modes are done using an appropriate kernel functions

which is define as K (x i, x) = φT(x i) φ(x) [12]. For exam-

ple, polynomial kernel with 2 orders K(x, x) = (x. xT +1)2

map the 2-dimensional space {(x1, x2) | x1, x2∈ R} into

6-dimensional space {(x1, x2, 2x1x2, 2x1, 2x2, 1) |

x1, x2∈ R} Furthermore, kernel function has an

important effect on the functional efficiency of SVM.

The popular kernel functions include Gaussian radial

basis function, polynomial and sigmoidal functions.

2. Method

2.1. Datasets

Our sample target populations are general students who

are beginning to write. The data for this research was

obtained from Khalid et al in [13]. This sample consisted

of 120 first grade children who assigned to two groups of

writers, below average printers (test group) and above

average printers (control group). There were 60 pupils

in the control group and 60 pupils in the test group. Each

participants were required to complete 8 drawing task;

vertical downward (VD), vertical upward (VU), horizon-

tal rightward (HR), horizontal leftward (HL), right

oblique downward (RD), right oblique upward (RU), left

oblique downward (LD), and left oblique upward (LU) as

shown in Figure 3. These drawings are the most basic

drawing and the most common and effective means of

communication that have been applied in various appli-

cation for more than a decade. In a simple sense, these

line drawings are a picture that convey to their viewer

information through the shape, size and manner

HIDDEN LAYER

(there may be several

hidden

layer

INPUT

LAYER

OUTPUT

LAYER

Figure 1. A simple neural network diagram.

Figure 2. Optimal separating hyperplain.

A. A. HASSEIM ET AL. 3



Figure 3. A notion of eight directions.

of interconnection of thin lines on a contrasting back-

ground [14].

Dynamic data (such as velocity and pressure values) of

drawing performance have been rigorously studied in

recent researches e.g., [15,16]. Those studies have shown

that dynamic data also affect the performance character-

istic of drawing tasks besides the used of static data. For

our experiments, 2 features were selected to be used in

the classification process:

1) Feature 1: The standard deviation of pen pressure

when drawing RU, p-value < 0.0001 and z-value = minus

4.319 and,

2) Feature 2: Ratio of time taken to draw HR and HL,

p-value < 0.0001 and z-value = minus 5.205.

2.2. Architecture

2.2.1. Artificial Neural Network Classification

Artificial Neural Network training was developed using

MATLAB 7.6 software. The network chosen for the pre-

diction neural network had one input layer, 2 hidden lay-

ers and one output. For the hidden layers, the number of

neurons is obtained by trial and error. The most compact

network is chosen and presented. The network training

parameters are:

 Training algorithm : Gradient descent with mo-

mentum training

 Perform function : Mean Square Error

 Training goal achieved : 0.04

 Training epochs : 10000

 Training momentum constant : 0.95

 Learning rate: 0.2

 Ratio to increase learning rate: 1.05

 Ratio to decrease learning rate: 0.7

The hidden layers and the output layer used log sig-

moid activation function which it calculate a layer's out-

put from its net input. This function generates outputs

between 0 and 1 as the neuron's net input could be any

values from negative to positive infinity. The threshold,



for the output was set to 0.5. Therefore if the final testing

value exceeds the threshold value, the function will takes

the value 1 and 0 otherwise. The flow chart of training

network using BP algorithm is shown in Figure 4.

Since this dataset is large, we used cross validation

method to test our classifiers. The data is randomly por-

tioned into 10 equally size folds. In each folds, we se-

lected 6 samples from the control group and 6 samples

from the test group. Next, one fold is used for testing

while the remaining 9 folds are used for validation. This

process is continued 10 times such that within each iter-

tion a different fold of the data is held-out for validation

and the rest folds are used for learning.

As the initial weight of the each training process is not

fixed, the network could give slightly different results.

Hence, we trained each algorithm with 10 trials and get

the average performance of the model.

2.2.2. Support Vector Machine Classification

The SVM was also run by using program MATLAB 7.6.

Total 120 samples are used as input signals. Linear SVM

is used as kernel function for training SVM. Usually

among popular kernels, the linear kernel is much faster in

training and testing speed. An important advantage of

linear classification is that training and testing procedures

are much more efficient. Therefore, linear classification

can be very useful for some large-scale applications [17].

Yes

Output

Control

group

Start

Identify the parameters

of the network

Splitting the data set

randoml

Training data with

ack

ation Testing data

Compare

Set threshold value

= 0.5

If < 0.5

Target

Test

group

Figure 4. Flowchart for identification of the handwriting

problem using artificial neural network model.

A. A. HASSEIM ET AL.

The 12 random signals are selected from the data used

to test SVM and the remaining data used for training.

Just like neural network, the experiment was done with

10 iterations of 10-fold cross validation and the final av-

erage performance of the classifier was taken out.

3. Results and Discussions

Table 1 shows the classification results obtained for each

attributes using ANN classifier and SVM classifier. The

classification performance was divided into 2 parts: con-

trol and test. Control performance was based on the clas-

sification rate of the samples from the control group in

the testing set while test performance was based on the

classification rate of the samples from the test group in

the testing set. The classification was considered correct

if the output from the model was similar to the one that

had been judged by the teachers (using Handwriting Pro-

ficiency Screening Questionnaire (HPSQ)).

In this paper, we used the classification error (rejection

of genuine category) as the metric. From Table 1, we can

see that the classification rate for SVM is better than

ANN in both cases; control group and test group with the

average accuracy of classificatory testing based on SVM

algorithm reaches more than 83%.

Based on the observation, neural network not only

provide weakly performance in regression problem but

also require more computational time than SVM tech-

nique. The recognition rates for SVM also show a sig-

nificant improvement compared to ANN. Moreover the

classificatory result of the SVM algorithm is more stable

since it is not easily influenced by primal weighting

value like neural network. In addition, the advantage of

this approach clearly lies in its simplicity since no pa-

rameter has to be tuned. Overall, SVM is more con-

venient and superior when it comes to identify and assess

“poor” writers.

Furthermore, between the two features, the results in-

dicated that feature 1 which is the standard deviation of

pen pressure when drawing RU is better than feature 2

(ratio of time taken to draw HR and HL) if we measure

the percentage of correctly classified performance of

control group. However, if the measure of performance

was percentage of test group correctly classified, the

feature 2 outperformed the feature 1.

Table 1. Classification result.

Classifiers Feature 1 Feature 2

Control group (%) 86.67 75.00

ANN Classifier

Test group (%) 58.88 63.33

Control group (%) 91.67 83.33

SVM classifier

Test group (%) 83.33 83.33

4. Conclusions

An experiment study of classification performance with

the aim of identifying children with and without hand-

writing difficulties has been presented. It is based on the

handwriting proficiency screening questionnaire (HPSQ).

However, the collected data from the questionnaire is

normally subjective and imprecise. Thus, the experiment

can be further improve by using dynamic data that is

both sensitive and specific is suggested for the screening

process to be effective.

Two techniques; ANN and SVM have been used in

this study to select those who are at risk of handwriting

difficulty due to the improper use of graphic rules. Here,

it can be concluded that both methods are able to classify

students with or without handwriting difficulties. How-

ever, the performances of the two classifiers are different,

SVM technique is more effective and doable way than

ANN method which it gives the average classification

rate more than 83%.

5. Acknowledgements

This work was supported by the Malaysia Ministry of

Higher Education and Universiti Teknologi Malaysia

under Vote Q.J130000.2623.09J28.

REFERENCES

[1] S. Chu, “Occupational Therapy for Children with Hand-

writing Difficulties: A Framework for Evaluation and

Treatment,” British Journal of Occupational Therapy,

Vol. 60, 1997, pp. 514-520.

[2] K. P. Feder and A. Majnemer, “Handwriting Develop-

ment, Competency, and Intervention,” Developmental

Medicine Child Neurology, Vol. 49, No. 4, 2007, pp.

312-317. doi.org/10.1111/j.1469-8749.2007.00312.x

[3] R. Sudirman, N. T. Mashadi and I. Ariffin, “Aspects of A

Standardized Automated System for Screening Children's

Handwriting,” First International Conference on Infor-

matics and Computational Intelligence, Bandung, 12-14

December 2011, pp. 49-54. doi:10.1109/ICI.2011.19

[4] S. Graham, V. W. Berninger, N. Weintraub and W.

Schafer, “Development of Handwriting Speed and Legi-

bility in Grades 1-9,” Journal of Educational Research,

Vol. 92, No. 1, 1998, pp. 42-52.

doi:10.1080/00220679809597574

[5] S. M. Koziatek and N. J. Powell, “Pencil Grips, Legibility,

and Speed of Fourth-Graders’ Writing in Cursive,”

American Journal of Occupational Therapy, Vol. 57, No.

3, 2003, pp. 284-288. doi:10.5014/ajot.57.3.284

[6] P. I. Khalid, “Handwriting Ability Assessment Model

Using Dynamic Characteristics of Drawing Process,”

PhD Thesis, Universiti Teknologi Malaysia, 2012.

[7] R. M. Guest, S. Chindaro, M. C. Fairhurst and J. M. Pot-

ter, “Automatic Classification of Hand Drawn Geometric

Shapes Using Constructional Sequence Analysis,” Sev-

A. A. HASSEIM ET AL.

enth International Conference on Document Analysis and

Recognition (ICDAR'03) icdar, Vol. 2, 2003, p. 990.

[8] S. Chindaro, R. M. Guest, M. C. Fairhurst and J. M. Pot-

ter, “Assessing Visuo-Spatial Neglect Through Feature

Selection and Combination from Geometric Shape

Drawing Performance and Sequence Analysis,” Interna-

tional Journal of Pattern Recognition & Artificial Intelli-

gence, Vol. 18, 2004, pp. 1253-1266.

doi:10.1142/S0218001404003757

[9] C. J. C. Burges, “A Tutorial on Support Vector Machines

for Pattern Recognition,” Data Mining and Knowledge

Discovery, Vol. 2, 1998, pp. 121-167.

doi:10.1023/A:1009715923555

[10] L. H. Tsoukalas and R. E. Uhrig, “Fuzzy and Neural Ap-

proaches in Engineering,” John Wiley and Sons, 1997.

[11] V. N. Vapnik, “The Nature of Statistical Learning The-

ory,” Springer, New York, 1995.

doi:10.1007/978-1-4757-2440-0

[12] C. W. Hsu, C. C. Chang and C. J. Lin, “A Practical

Guide to Support Vector Classification,” Technical Re-

port, Department of Computer Science and Information

Engineering, National Taiwan University, 2003.

[13] P. I. Khalid, J. Yunus, R. Adnan, M. Harun, R. Sudirman

and N. H. Mahmood, “The Use of Graphic Rules in

Grade One to Help Identify Children At Risk of Hand-

writing Difficulties,” Research in Developmental Dis-

abilities, Vol. 31, 2010, pp. 1685-1693.

doi:10.1016/j.ridd.2010.04.005

[14] H. Freeman, “Computer Processing of Line-Drawing

Images,” ACM Computing Surveys, Vol. 6, 1974, pp.

57-97. doi:10.1145/356625.356627

[15] R. M. Guest, M. C. Fairhust and J. M. Potter, “Diagnosis

of Visuo-Spatial Neglect Using Dynamic Sequence Fea-

tures from A Cancellation Task,” Pattern Recognition

Analysis & Applications, Vol. 5, 2012, pp. 261-270.

doi:10.1007/s100440200023

[16] P. I. Khalid, J. Yunus and R. Adnan, “Extraction of Dy-

namic Features from Hand Drawn Data for the Identifica-

tion of Children with Handwriting Difficulty,” Research

in Developmental Disabilities, Vol. 31, 2010, pp. 256-262.

doi:10.1016/j.ridd.2009.09.009

[17] G. X. Yuan, C. H. Ho and C. J. Lin, “Recent Advances of

Large-Scale Linear Classification,” Proceedings of the

IEEE, Vol. 100, No. 9, 2012, pp. 2584-2603.

doi:10.1109/JPROC.2012.2188013