JOURNAL OF COMPUTER SCIENCE AND ENGINEERING, VOLUME 3, ISSUE 2, OCTOBER 2010

56

A comparison among Features Used in Offline Signature Verification Systems Nabil A. Lashin, Hanaa M. Hamza

Abstract—Many features have been used for signature verification systems in the last decades. A comparison among the most commonly used sets of features (Global, Moments, Grid, and Texture) has been presented in this paper. The proposed system combines the results of using global, moments, grid, and texture features, then compares among the effectiveness of using each feature individually and using the combined results. For each set of features a multi-layer perceptrons (MLP) neural network is used as a first and preliminary stage classifier. Then taking the average of these individual outputs represents the final decision. The system is tested and proved experimentally that combining various feature sets in verification process achieves better results than using individual features. Moreover, the proposed system can detect the different types of forgeries in low false acceptance rate (FAR). Index Terms— Global features, Grid features, moment invariants, neural network, offline signature verification, texture features.

——————————
——————————

1 INTRODUCTION andwritten signature reflects unique characteristics for each signer. Signature can be composed of clear written letters or symbols or undefined patterns. Automatic signature based applications have a wide spread domain because it is easier for human to write his signature when he / she is asked to do, and the devices used to automatically capture the signature is cheaper than other devices used to capture other biometrics. These applications may aim to recognize the signer through his / her signature, or aim to verify if the queried signature is signed by the claimed signer. This paper presents a model system for automatic signature verification. A signature verification system should be able to differentiate between the genuine signature and the forgery one. By forgery is meant copying, falsifying, or altering any kind of written or printed matter for the purpose of defrauding others [1]. There are three different types of forgeries to be recorded: first one is random forgery which is written by the person who doesn’t know the shape of genuine signature, and just knows the signer's name. The second, called simple forgery, which is represented by a signature sample, written by the person who knows the shape of original signature without much practice. The last type is skilled or simulated forgery, represented by a trained imitation of the genuine signature. Signature verification systems exist in two different categories based on the way the signatures are captured. They can be either online or offline systems. In online systems, an electronic tablet with a special pen is used to capture the signature during the signing process. Dynamic information can be stored online such as: pen pressure, timing and speed. In offline systems, a scanner is used to capture the signature after the signing process is completed. Online systems are more accurate than offline because of using dynamic data which is difficult to be forged. The forger always tries to imitate the shape of the signature not how the signature was made. But

H

————————————————

Nabil A. Lashin, Lecturer - Faculty of Computers and Informatics, Zagazig University. Hanaa M. Hamza, Demonstrator - Faculty of Computers and Informatics, Zagazig University.

using these systems requires the presence of the signer during the capturing process, so it may be undesirable in some applications. To authenticate documents which are signed every day in official issues, using offline systems is more suitable and quicker than online ones. Other systems try to combine the benefits of both online and offline systems.[2, 3] It captures the online signature dynamically at the registration phase to capture more than one reference signature. After that at each transaction, the signature can be put on any paper and captured offline. This paper interests in offline signature verification system as a trial to enhance its results because of the needing for such systems in many security based applications. Such systems are composed of two phases: registration phase, which is done only once at which the signer is asked to give many reference signatures to let the system learns its characteristics. The second phase is a testing phase at which the signer signs to test the system which is already learnt, and this is done at each time the signer needs to query the system. During each phase basic steps are done as follow: preprocessing, feature extraction, classification, and performance evaluation. The architecture of the proposed model is exhibited in Fig. 1. In which four features sets are extracted. Each set's results are fed into an artificial neural network as a first stage classifier, then try to combine the output results by taking the average to give the final decision. A complete classifier is being implemented for each signer to allow for the system to be dynamically add / remove signers to system database [4].

2. SYSTEM ARCHITECTURE The following figure presents the architecture of the proposed system for offline signature verification based on four different sets of features. The system consists as shown of four phases. These phases are preprocessing, feature extraction, Multi-layer perceptrons (MLP) neural network classifier built for each feature set, and the final decision tool which combines the decisions of the previous stage to verify the queried signature and decide if it is genuine or forged. The following sections will discuss each phase in more details.

JOURNAL OF COMPUTER SCIENCE AND ENGINEERING, VOLUME 3, ISSUE 2, OCTOBER 2010

57 Captured queried Signature Preprocessing

Feature Extraction

Global Features 14

Classification

MLP

Moments Features 84

Grid Features 96

MLP

Texture Features 48

MLP

MLP

Taking the average

Output the final decision (Genuine, or Forgery)

Fig. 1. Structure of the proposed system in the testing phase.

2.1 Preprocessing This stage is considered as the first treatment with the signatures or preprocessing phase. At which the captured signatures are converted into binary data images in the aim of preparing the signature for the next phase.

Where size X and size Y represent both the horizontal and vertical dimensions of the signature, xi and yi the coordinates of each image pixel, CmImgX and CmImgY the center of gravity in both directions, and totPixImg the total number of pixels in the image.

2.2 Feature extraction Each set of features provides a specific kind of information about the signature topography. The extracted global features are concerned with the structure of the whole signature. The extracted moments features form shape descriptors which are invariant with respect to translation, scale, and rotation. The extracted grid features divide the captured signature into grids to extract each grid local features. While the extracted texture features collect the information about the transitions of black and white pixels.

– Out Ratio [2]

•

Global Features -

The ratio of pixels (4) localized outside the rectangle (totPixOut) formed by one standard deviation of the pixels previously determined around the mean and the total number of pixels in the image (totPixImg):

totPixOut OutR =

(4)

totPix Im g

Where both SD_X and SD_Y [2] are defined as the distribution of the pixels around the X axis and Y axis respectively as exhibited in (1),(2).

Aspect Ratio [2]

The aspect ratio (3) of the signature is calculated by the ratio of the standard deviation of the pixels in relation to the center of gravity of the signature on both the horizontal and vertical directions. sizeX 2 (1) ∑ ( x i − Cm Im gX ) / totPix Im g SD x = i =1

– Vertical and Horizontal Center [4],[5],[6] Both the vertical center Cy and horizontal center Cx are given by (5) and (6) respectively. y max

∑

C

y

=

∑

y

y =1 x max x =1

x max

sizeY

SD y = Pr op =

2

∑ ( y − Cm Im gY ) / totPix Im g i =1 i SD x SD y

(2)

(3)

∑

C

x

=

∑

x =1

x =1

∑

∑ x∑ x =1 x max

x max

y max y =1

y max y =1

∑

y max y =1

b[ x, y ]

(5)

b[ x, y ]

b[ x , y ] b[ x , y ]

Where b[x, y] are the foreground pixel.

(6)

JOURNAL OF COMPUTER SCIENCE AND ENGINEERING, VOLUME 3, ISSUE 2, OCTOBER 2010

58 – Base line shift [4] It is defined as the difference between the vertical centers of gravity (5) of the left and the right part of the image. It is considered as a measure for the orientation of the signature. - Maximum Vertical and Horizontal Projection [4,5,6] The vertical projection of the skeletonized signature image is calculated. The highest value of the projection histogram is considered as the maximum vertical projection. Also, the horizontal projection histogram is calculated and the highest value of it and considered as the maximum horizontal projection. - Vertical and Horizontal Projection Peaks [4]

•

The number of the local maxima of the vertical projection histogram is considered as a vertical projection peaks while the number of the local maxima of the horizontal projection histogram is considered as horizontal projection peak. - Number of Edge and Cross Points [4,5,6] Edge points are defined as signature points that have only one 8-neighbor. While, cross points are defined as signature points that have at least three 8-neighbors. - Number of closed loops [4] The number of closed loops (CL) can be defined as given by (7) where EP represents the number of edge points and EL represents the number of extra departures and can be determined by (8).

CL = 1 +

EL =

∑

EL − EP 2

[( Number of 8−nighbors ) − 2]

(7)

(8)

It is noticed that if the skeletonized signature image is not compact, that is, the signature is divided into two or more non overlapping segments, the number of closed loops, will have no physical meaning. But still, this number can describe the amount of complexity that the signature lines involve. Moments Features [7]

Moments are used in image analysis where they are considered the best descriptor for the image feature. Hu (1962) defines seven values, computed by normalizing central moments through order three, that are invariant to object scale, position, and orientation. The signature is rotated from 0o to 360o in 30o increments, the seven moments' values are extracted correspond to twelve different rotation angle. So, for each signature 84 features are extracted. •

Texture Features [4]

The texture feature values are calculated using the cooccurrence matrices. As shown in Equation 9 the co-occurrence matrix is a 2 * 2 matrix which describes the transition of black and white pixels. p p  P d [ i , j ] =  p 00 p 01  , 11   10 (9) where p00 is the number of times that two white pixels occur, separated by d. p01 is the number of times that a combination of a white and a black pixel occurs, separated by d. p10 is the same as p01. p11 is the number of times that two black pixels occur, separated by d. The image is divided into six rectangular segments (3 * 2). For each segment the P(1,0), P(1,1), P(0,1) and P(-1,1) matrices are calculated and the P01 and P11 elements of these matrices are used as texture features of the signature. So, totally 48 features are calculated. 2.3 Classifier

All cross Po int s

•

Fig. 2. The grid feature vector of a signature.

Grid Features [4]

The grid feature vector has 96 values calculated from dividing the signature's image into 96 rectangular segments (12 * 8), and for each segment, the sum of foreground pixels is calculated. A signature image and the corresponding grid feature vector are shown in Fig. 2. A black rectangle indicates the maximum number of black pixels, and a white rectangle indicates the smallest number of black pixels.

For each signer an individual two stage classifier is being implemented as cited before to easily add / remove signers to the proposed system. An artificial neural network classifier for each set of features forms the first stage classifier, and then in the second stage a combining classifier is built by calculating the average of the previous stage’s outputs. As shown in Fig. 3, MLP neural network for each feature set consists of 2 layers (2 neurons in a single hidden layer, 1 neuron in the output layer). The learning phase of the first stage classifier is done using 16 reference signatures collected from each signer in the registration phase. Through the testing phase presented by Fig. 1, the verification scenario will be as follow: the queried signature's features are extracted, and then these values will be presented to the classifier of the claimed signer to decide if this signature is genuine or forged one.

JOURNAL OF COMPUTER SCIENCE AND ENGINEERING, VOLUME 3, ISSUE 2, OCTOBER 2010

1 2 3 . . . n

Output

Inputs Where n, is the number of features in the feature’s set. (n = 14 in the Global Feature set ANN, n = 84 in the Moments Feature set ANN, n = 96 in the Grid Feature set ANN, and n = 48 in the Texture Feature set ANN)

59 There is no unique database for Automatic Signature Verification systems. Previous researches considered the attributes of their used databases during the system's development. To have a precise automatic signature verification system the behavioral features of the signer should be considered and the system should be customized based on the signers' attributes [8]. In the proposed system, signatures are taken from different signers with various signatures' styles which vary among clear Arabic letters; English letters, and patterns which vary in its writing direction. The forged samples are signed by different volunteers. Sample signatures from the database are shown in Fig. 4.

Fig. 3. MLP for each feature set.

3. SIGNATURE DATABASE SIGNER NAME

ORIGIN

RANDOM FORGED

SIMPLE FORGED

SKILLED FORGED

Mohamed Hamza Manar Sabry Fatma Hamed Sabry Morsy Samar Sabry Mohamed Hassan Abd-El rahman Mohamed Hosni Shabana

Nora Zaki

Akmal sleem Fig. 4. A sample from database.

As indicated in Table 1, samples are collected from 55 signers. For training, 16 references genuine, 8 random forgeries, 8 simple forgeries, and 8 skilled forgeries are used as commonly developed [9]. For testing, 8 genuine, and 24 forged ones represents three types of forgeries.

Random

TABLE 1 : VERIFICATION DATABASE

Forged

For 55 Signers, Number of Signatures

Genuine

8 * 55

Random

8 * 55

Training Set

Genuine

16 * 55

Forged Simple Forged Skilled

Testing Set

8 * 55 8 * 55 8 * 55

JOURNAL OF COMPUTER SCIENCE AND ENGINEERING, VOLUME 3, ISSUE 2, OCTOBER 2010 (Unseen

Forged

during

Simple

training

Forged

process)

Skilled

8 * 55 8 * 55

Forged

4. SYSTEM'S PERFORMANCE In the feature extraction phase, a comparison among the times consumed to extract each set of features from on signature is recorded. As shown in Fig. 5, it is clear that Grid features are extracted in the least time, while moments features take a longest time to be extracted. Time to extract features from one signature 1.4 1.2

Minutes

1 0.8

60 - False Acceptance Rate (FAR): if the forged signature is accepted falsely as genuine one. - False Reject Rate (FRR): if the genuine signature is rejected falsely as forged one Minimizing FAR rate is considered better than FRR because the person can always be given a second chance to verify his/her identity [7]. - Correct Acceptance Rate (CAR): if the genuine signature is accepted correctly as genuine one. - Correct Rejection Rate (CRR): if the forged signature is rejected correctly as forged one. 4.1 Using feature sets individually The first stage classifier is consisted of 4 MLP, one network for each feature set. In testing phase, a queried signature is presented to the system, given a candidate signer name. The output of each one of these classifiers should be close to ``1'' if the signature is genuine or close to ``0'' if it is a forged signature. The features are ranked according to the minimum FAR as follows: global, grid, moments, and texture. So, the global features achieve the minimum error rate in forgery detection.

0.6

4.2 Combining the feature set's outputs

0.4 0.2 0

Global

Minutes 0.333333333

Moments

Grid

Texture

1.25

0.125

0.208333333

Feature Sets

Fig. 5. Time Complexity.

In the second stage, the outputs of the previous stage are averaged to constitute the final decision. The experimental results are summarized at Table 2. It is noticed that the second stage in the classifier which combines the outputs of the individual features has the minimum FAR that is reached nearly to zero. So, the system won't accept falsely the queried signature as genuine where it is forged one. Moreover, it is clear that the maximum CRR is reached; hence the system will reject correctly the forged signatures.

The classifier system is evaluated using the following statistical evaluation parameters: TABLE 2 : THE EXPERIMENTAL RESULTS. Origin Signatures

Global Moments Grid Texture

Simple Forgeries

Skilled Forgeries

FRR

CAR

FAR

Random Forgeries CRR

FAR

CRR

FAR

CRR

23.86 3.86 17.05 58.41

76.14 96.14 82.95 41.59

2.73 12.95 3.18 18.86

97.27 87.05 96.82 81.14

2.05 11.59 4.32 11.59

97.95 88.41 95.68 88.41

2.05 11.82 3.64 16.82

97.95 88.18 96.36 83.18

11.14

88.86

0.45

99.55

0.45

99.55

0.68

99.32

Combining classifier

5. CONCLUSION The paper presents a model that compares the effectiveness of using various sets of features in signature verification systems. The tested features are classified to four sets: global, moments, grid, and texture features. Each set is adequate to discriminate among signatures, and provides a specific kind of information. This paper compares among these features, and then the proposed system's final decision is based on the average function which combines the outputs of the four feature sets' classifiers. The proposed model achieve the least FAR in

forgery detection, so these results are satisfactory compared to the results of the previous presented systems. The proposed comparison may be considered as a guide for the coming automatic signature verification systems; according to the sensitivity of the automatic signature verification system's application, the suitable set of features can be chosen and extracted. Further work could be repeating these experiments using larger data sets, and by adding more new features.

JOURNAL OF COMPUTER SCIENCE AND ENGINEERING, VOLUME 3, ISSUE 2, OCTOBER 2010

61

REFERENCES [1] Rasha Abbas, B.E., “Backpropagation Networks Prototypes for Off-line Signature Verification". Master Thesis, Department of Computer Science, RMIT, 1994. [2] A. Zimmer and L.L. Ling, "Offline Signature Verification System Based on the Online Data". EURASIP Journal on Advances in Signal Processing, 2007. 2008(Article ID 492910): p. 16 pages. [3] Y. Qiao, J. Liu, and X.Tang, "Offline Signature Verification Using Online Handwriting Registration". Computer Vision and Pattern Recognition, CVPR '07. IEEE Conference 2007: p. 1-8. [4] H. Baltzakis and N. Papamarkos, "A new signature verification technique based on a two-stage neural network classifier". Engineering Applications of Artificial Intelligence, 2001: p. 95-103. [5] E. Özgündüz, T. Şentürk, and M. Elif Karslıgil, "Off-Line Signature Verification and Recognition by Support Vector Machine". 13th European Signal Processing Conference, 2005. [6] I. Barbantan, C. Vidrighin, and R. Borca, "An Offline System for Handwritten Signature Recognition". IEEE Xplore, 2009. [7] Oz, C., "Signature Recognition and Verification with Artificial Neural Network Using Moment Invariant Method". Advances in Neural Networks - Second International Symposium on Neural Networks. Proceedings, Part II., 2005. 3497: p. 195-202. [8] S. Ghandali and M. Moghaddam, "Off-Line Persian Signature Identification and Verification Based on Image Registration and Fusion". Journal of Multimedia, North America, 4, 2009. [9] M. K. Kalera, S. Sriharly, and A. Xu, "Offline Signature Verification and Identification Using Distance Statistics". International Journal of Pattern Recognition, and Artificial Intelligence, 2004. 18(7): p. 1339 – 1360.