1. offline signature recognition using machine learning

8/10/2019 1. Offline Signature Recognition Using Machine Learning

1/9

Offline Signature Recognition Using M

Mohammad Ikhsan Bin Zakaria,

Engineering and Information Technologies, Inte

Sarajevo, Bosnia and Herze

E-mails:[email protected]

Abstract
mailto:[email protected]:[email protected]


2/9

3rd

International Symposium on Sustainable Development

persons. Offline signature recognition and verification

proposed. In [4] they used stroke angle and stroke speed a

This paper is organized into five sessions. The following

session 1, session 2 describes the proposed method, in

preprocessing and feature extraction, in session 4 descri

session describes conclusion.

2.SIGNATURE IMAGE PREPROCESSING

In this paper signature image preprocessing can be

Hi t E li ti (2) F i T f (3) Bi


3/9

3rd


Figure 2: Basic format CN for P

Where Pi is the bi-level pixel value in the neighborhood o


4/9

3rd


were collected from 30 students at International Burch

signatures samples. After converting 300 signatures into

into 300 single signature images. The file was analyzed ffollowing session describes ANN classification and testin

3.2. Feed-forward Backpropagation Network (newff)

In this experiment we used feed-forward backpropagatio

error as the measurement for performance on the neurinfluence of training algorithm and transfer function whi

recognized signatures. In figure 4 (a) shows the examp


5/9

3rd


20 20-10-10-1 traingdmlogsig, tan

pur

20 20-10-10-1 traingdxlogsig, tan

pur

The performance of training is influenced by number o

learning methods. Generally,mseis calculated in MATL

(2) it is just additional description of calculating mse us

used the logic to compare between target output and actu

in target output that are larger or equal to actual output an


6/9

3rd


20 20-10-10-1 trainbfg tansig, logsig, log

20 20-10-10-1 trainbfg logsig, tansig, tan

20 20-10-10-1 trainlm logsig, tansig, tan

Our attributes in table 2 are training algorithm trainlm

session trainbfg spent more time than trainlm to find ou

inputs with two hidden layers and tansig as transfer fuaccuracy rate and 0 in mse error. Thus we concluded t

produced the highest accuracy we got. However, mse do


7/9

3rd


20 20-10-10-1 trainlmlogsig, tans

pure

20 20-10-10-1 trainbfglogsig, tans

pure

In this experiment the lowest mse is 0.0476 and the hi

table 3 shows that there are two highest accuracy rates bu

output is the one that has lower mse error, even though

inputs, hidden layer but different training algorithms. Tr

As a sample of testing session in this network figure 4


8/9

3rd


using learnlvq1 or learnlvq2 and hidden neurons. Even

values. Thus learning vector quantization gave the highe

mse.

(a) (b)


9/9

3rd


Khuwaja, G. A. &Laghari, M. S. (2011). Offline Handw

Academy of Science, Engineering and Technology 59.

Basavaraj, L. &Sudhaker Samuel, R.D. (2009). Offli

Recognition: An Approach Based on Four Speed Stro

Recent Trends in Engineering, Vol 2.

Zhao, F., & Tang, X. (2006). Preprocessing and postproc

minutiae extraction, Pattern Recognition 40 (2007) 12

Recognition Society.

Zhili, W. (2002). Fingerprint Recognition. UnpublishBaptist University.

1. offline signature recognition using machine learning

Documents