8/11/2019 my EUVIP 13 Poster

1/1RESEARCH POSTER PRESENTATION DESIGN 2012www.PosterPresentations.com

Handwritten signatures are one of the most widely used biometrics, particularly in

financial and legal transactions. Offline Signature verification is still one of the most

challenging problems in biometrics. In this study, we have evaluated the performanceof different classifiers for offline signature verification based upon local binary

patterns feature set. The feature vector is formed by dividing the signature images

into twelve local regions and forming a code matrix by their LBPs. The histogram ofeach code matrix is formulated and concatenated. The dimensionality of feature

vector is subsequently reduced by keeping the 256 DCT coefficients of the

concatenated vector. We have investigated the performance of seven classifiers onFUM-PHSDB dataset comprising 20 classes of genuine and forged signatures of depth

20 and 10 respectively. Experimental findings depict that LS-SVM performs the best

among the seven classifiers, achieving the Equal Error Rate (EER) of 13%.

Index Terms Of fline signature verification, local binary patterns, classification, LS-SVM.

ABSTRACT

EXPERIMENTATION

RESULTS

The paper presents a comparative analysis of seven classifiers for offline signature

verification using gray level features based upon local binary patterns. SVM based

classifiers outperform the others, whereby LS-SVM demonstrated the bestclassification accuracy with EER of 13%. The results were also compared with

different published approaches, depicting promising result of the proposed approach.

In future, we intend to extend the approach to multi-modal offline signatureverification.

CONTACT

rameezw@gmail.com, 120808@student.au.edu.pk

atif.mansoor@mail.au.edu.pk

The database

The FUM-Persian Handwritten Signature Database (FUM-PHSDB 2006) is used inexperiments. The database contains 20 signature classes, consisting of 20 genuine

signatures and 10 expert forgery signatures for each individual. The signatures were

obtained by signing on a 1010 cm white paper using a black pen and later scanned ata resolution of 300 DPI (dot per inch). Fig. 1 shows genuine and forged signature

samples from a class of the database.

Fig.1 Genuine (left) and forged (right) signature sample from FUM-PHSDB

We divided the database in two equal sets. The training and validation set contains 10

original signatures and 5 skilled forgeries for each class. Thus, a total of 300 samplesare used for training the classifier and a further 300 for verification.

Gray Level Feature Extraction

Local binary pattern (LBP) is an efficient gray-level feature commonly used in texture

classification. The popularity of the LBP operator, especially in real-time applications,is mainly due to its simple computation and its robustness to illumination variations

and other monotonic gray-scale changes.

,(v) = v v 21

=0where, vis the value of original pixel, vthe value of neighbor, Pis the number ofneighbors and Ris radius over which LBP is applied, andis a step function. The gray-level image from a signature class was divided into 12 equal blocks for feature

extraction. The selected blocks were four in vertical and three in horizontal direction,with a 60% overlap. We have applied the rotation invariant LBP to these blocks

forming a code matrix. The histogram of each code matrix is further formulated,

having a dimension of 255. Fig. 2 shows the histograms of code matrices of twoblocks, clearly demonstrating their variations.

Subsequently, the 12 histograms were concatenated to form the feature vector of3060 dimension. The dimensionality of feature vector was reduced by the Discrete

Cosine Transform (DCT). The DCT was calculated as follows:

= () () (21)(1)2

=1k= 1,2,N

Where, () is the nth value of LBP histogram, N is the dimension of input vector(3060 in this case). The first 256 coefficients of the DCT form the final feature vectorfor classification. The histogram of a reduced feature vector is shown in Fig.2.

Fig.2 Histograms of a code matrix of individual blocks to form the reduced feature vector

Rank

Classifier Performance Evaluation

Classifier EER

1 LS-SVM 13 %

2 SVM 22 %

3 DLRT 24 %

4 ANN 32 %

5 FLD 36 %

6 Log Discriminant 36 %

7 Naive Bayes 47 %

S

No

Performance Comparison

Technique / Feature Classifier Performance

1 Stroke distributions SVM EER = 13.12%

2

Gray level features

with Discrete Cosine

Transform

LS-SVM EER = 11.05 %

3

Pixel distributions

using Chebyshev

distances.

Combination

of similarity

based

classifier and

SVM

EER = 7%

4

Discrete Wavelet

Transform and

Image Fusion

Difference

based

Classifier

FRR = 8.9%

FAR = 10%

5Gray-Level Stoke

VariationsSVM EER = 14.32%

6Contourlet

Transform

Difference

based

Classifier

EER = 14.00%

7 PROPOSED LBP LS-SVM EER = 13%

Institute of Avionics & Aeronautics, Air University, Islamabad, Pakistan

Rameez Wajid, Atif Bin Mansoor

CLASSIFIER PERFORMANCE EVALUATION FOR OFFLINE

SIGNATURE VERIFICATION USING LOCAL BINARY

PATTERNS

The Receiver Operating Characteristic (ROC) curves obtained from the classificationsare displayed in Fig 3. The classifier performance is evaluated in Table I, using Equal

Error Rate (EER) as the performance parameter for each classifier.

Fig.3 Receiver Operating Characteristic (ROC) Curves for different classifiers

CONCLUSIONS

Table I clearly display the suitability of SVM based classification methods over others

for offline signature verification. The best results are obtained using the LS-SVM

classifier, achieving an EER of 13%.

The EER of LS-SVM classifier is compared with different approaches reported in

literature in Table II. The comparative analysis shows promising results by employingLS-SVM based classification using LBP features.

TABLE I CLASSIFIERPERFORMANCE EVALUATION TABLE II PERFORMANCECOMPARISON