Offline Signature Recognition and Verification System ... - ScienceDirect

Available online at www.sciencedirect.com Available online at www.sciencedirect.com

Available online at www.sciencedirect.com

ScienceDirect

Procedia Computer Science 00 (2017) 000–000 Procedia Computer Science (2017) 000–000 Procedia Computer Science 11600 (2017) 621–628

www.elsevier.com/locate/procedia www.elsevier.com/locate/procedia

2nd International Conference on Computer Science and Computational Intelligence 2017, 2nd International Conference on Computer Science andBali, Computational ICCSCI 2017, 13-14 October 2017, Indonesia Intelligence 2017, ICCSCI 2017, 13-14 October 2017, Bali, Indonesia

Offline Offline Signature Signature Recognition Recognition and and Verification Verification System System using using Efficient Fuzzy Kohonen Clustering Network (EFKCN) Algorithm Efficient Fuzzy Kohonen Clustering Network (EFKCN) Algorithm Dewi Suryani, Edy Irwansyah∗∗, Ricki Chindra Dewi Suryani, Edy Irwansyah , Ricki Chindra

Computer Science Department, School of Computer Science, Bina Nusantara University, Jl. K. H. Syahdan No. 9, DKI Jakarta, 11480, Indonesia Computer Science Department, School of Computer Science, Bina Nusantara University, Jl. K. H. Syahdan No. 9, DKI Jakarta, 11480, Indonesia

Abstract Abstract Research on offline signature recognition still has not shown satisfactory results as the results of recent research. Therefore this Research signature recognition has not shown satisfactory results as theemployed results ofanrecent research. Thereforeclusthis study aimsontooffline proposed an offline signaturestill recognition and verification system which efficient fuzzy Kohonen study aims to proposed an 1offline signature recognition and verification system which employed an efficient fuzzy Kohonen clustering networks (EFKCN) 1 algorithm. The proposed recognition system and signature verification system consist of five stages algorithm. The proposed recognition clustering, system andand signature verification system consist of fivepatterns stages tering networks (EFKCN) image including data acquisition, processing, data normalization, evaluation. The recognition of signature including data acquisition, processing, data normalization, clustering, evaluation. The recognition of signature patterns using the clustering methodimage with the EFKCN algorithm shows relatively betterand result with 70% accuracy compared to the accuracy using the clustering the EFKCN shows relatively better result accuracy compared to the accuracy 2 which is 53%, algorithm and a good signature recognition resultwith can70% be developed to assist the verification of previous researchmethod resultswith 2 which is 53%, and a good signature recognition result can be developed to assist the verification of previous research results system as well as the personal data verification system as made in this study. system as well as the personal data verification system as made in this study. c 2017 2017 The The Authors. Authors. Published Published by by Elsevier Elsevier B.V. © B.V. c 2017 The Authors. Published by Elsevier B.V. Peer-review Computer Science Science and and ComPeer-review under under responsibility responsibility of of the scientific committee of the 2nd International Conference on Computer Peer-review under responsibility putational Intelligence 2017.2017.of the scientific committee of the 2nd International Conference on Computer Science and ComComputational Intelligence putational Intelligence 2017. Keywords: Signature Recognition, Verification System, EFKCN Keywords: Signature Recognition, Verification System, EFKCN

1. Introduction 1. Introduction In the present, there are several ways to check the validity of one’s personal data, starting from using signature to In the present, thereisare several to check validity of by one’s starting from usingassignature to fingerprint. Signature a sign as aways symbol of the the name written thepersonal hand anddata, by the person himself a personal fingerprint. Signature is a sign as a symbol of the name written by the hand and by the person himself as a personal marker. Signatures are often used in data verification either in schools, banks, corporations, hospitals, government, and marker. Signatures used inofdata verification either in schools, banks, corporations, government, and much more. Due to are the often importance signature function, there are many parties who want tohospitals, manipulate the signatures much more. Due to the importance of signature function, there are many parties who want to manipulate the signatures of others. Duplicate signatures can be detrimental and included in the criminal realm. Identifying signatures can be of others. Duplicate signatures can Online be detrimental included is inused the criminal realm. Identifying be ascertained both online and offline. signatureand recognition by putting a signature on thesignatures pen tabletcan while ascertained both online and offline. Online signature recognition is used by putting a signature on the pen tablet while offline signature recognition is done by using a scanner. Until now, research on offline signature recognition still has offline signature recognition by using scanner. Untilconducted now, research on offline signature recognition still has 3 not shown satisfactory resultsisasdone the results of arecent research by Ahmed et al. 3 . In the other hand, Bezdek not shown satisfactory results as the results 4 1 5of recent research conducted by Ahmed et al. . In the other hand,4 Bezdek et al. 4 , Yang et al. 1 , and Irwansyah et al. 5 are developing a fuzzy Kohonen clustering networks (FKCN) 4 and an et al. , Yang al. , andclustering Irwansyahnetworks et al. are developing a fuzzy Kohonen clusteringwhich networks and an 1 algorithms for data clustering in its(FKCN) implementation efficient fuzzyetKohonen (EFKCN) efficient fuzzy Kohonen clustering networks (EFKCN) 1 algorithms for data clustering which in its implementation ∗ ∗

Corresponding author. Tel.: +62-21-534-5830 ext 2188; fax: +62-21-530-0244 Corresponding Tel.: +62-21-534-5830 ext 2188; fax: +62-21-530-0244 E-mail address:author. [email protected] E-mail address: [email protected]

c 2017 The Authors. Published by Elsevier B.V. 1877-0509 cunder 1877-0509 2017 The Authors. of Published by Elsevier B.V. the 2nd International Conference on Computer Science and Computational IntelliPeer-review the scientific committee 1877-0509 © 2017responsibility The Authors. Published Elsevier of B.V. Peer-review responsibility of the scientificbycommittee of the 2nd International Conference on Computer Science and Computational Intelligence 2017. under Peer-review under responsibility of the scientific committee of the 2nd International Conference on Computer Science and gence 2017. Computational Intelligence 2017. 10.1016/j.procs.2017.10.025

622 2

Dewi Suryani et al. / Procedia Computer Science 116 (2017) 621–628 Suryani et al. / Procedia Computer Science 00 (2017) 000–000

can produce a correct rate up to 92.67%. In this paper, we proposed an offline signature recognition and verification system which employed the EFKCN algorithm. Like Chaudhari et al. 2 , Salambue 6 , and Ainun et al. 7 , we also applied Hu’s seven moment variant 8 as our preprocessing method. Unlike the others, we focus on EFKCN algorithm. 2. Related Work Signature is an important part of human life. Several tasks need a signature to be completed, especially for personal authentication and verification. Due to its significant function, sometimes is misused by other people. In order to prevent that problem, many researches were done in developing a system to automatically recognize the human signatures 2,3,6,7,9 . Chaudhari et al. 2 are being part of it who proposed a signature recognition system by implementing fuzzy min-max algorithm 10 in a neural network framework. In their approach, the construction of the system is begun with preprocessing data and extracting its information using data acquisition and Hu’s seven moment invariant. They reported a significant improvement in the accuracy which is nearly 53% and up to 92% when increasing the signatures per class. Moreover, the study of Ismail and team 9 successfully built another recognition and verification system using principal components analysis (PCA) 11 which focuses on offline signature. They separated the process into two different tasks, i.e., the recognition that used k-nearest neighbors (k-NN) classifier and the verification that used artificial neural network (ANN). However, before both tasks running, the image preprocessing and feature extractions were applied by PCA. Their result using PCA is rising approximately 5% in false recognition rate (FRR) compared to the experiment without PCA. Furthermore, the approach of Ahmed et al. 3 concentrated in feature extraction methods for offline signature recognition and verification system. Here, they employed the projection-based of Discrete Radon Transform (DRT) as the method that consists of horizontal, vertical, and the combination of both projections. However, they need a signature preprocessing, which is done by Otsu’s threshold method 12 . For training the signatures data, their work implemented dynamic time warping (DWT) algorithm using Euclidean distance. By taking the advantages of the projection-based DRT, they achieved optimal performances in processing time, memory storage, and results, particularly the combination of horizontal and vertical projections. The results are measured by false rejection rate (FRR), false acceptance rate (FAR), total error rate (TER) and equal error rate (EER), i.e., 8.49, 5.60, 14.09, and 7.60 respectively, which is impressive results as a new approach. Similar to Chaudhari et al. 2 approach, Salambue 6 and Ainun et al. 7 proposed the signature recognition system which preprocessed the data using Hu’s moment invariant. However, their recognition methods are different. For Salambue 6 , he used the Euclidean distance algorithm as the classification and recognition method while Ainun et al. 7 utilized the radial basis function (RBF) neural network. Based on their conducted experiments, Ainun and the team gain 12% of error rate for the classification task and 20% for the recognition task. 3. Methodology Five stages of the study were carried out consisting of (1) data acquisition, (2) image processing, (3) data normalization, (4) clustering, and (5) evaluation. Data acquisition was conducted for 80 samples of signatures obtained from 8 persons with 10 different conditions in both sitting, standing, silent, and fast conditions. 50 signatures were used as training data and another 30 were used as test data. Image processing stages consist of preprocessing and feature extraction sub-stages. The preprocessing sub-stages are a series of data scan processes to Red Green Blue (RGB) format, conversion to grayscale, convert to binary image, inverted binary image, border elimination, and bounding box extraction. Grayscale image is a blend of gray color variations. From black at the lowest intensity to white at the highest intensity. To get the intensity value by using the following formula: grayscale = 0.2989 ∗ R + 0.5870 ∗ G + 0.1140 ∗ B

(1)

Border elimination function to make the signatures look more unified and the bounding box extract works to remove unused backgrounds, so the signature pattern looks more clear. At the feature extraction stage will be used moment


623 3

invariant. This method is useful for normalizing an image that changes due to scale and rotation. The respondent’s signature will be normalized to: 1. Regular moment: moments obtained from boundary shapes and interior areas, converted into functions to be implemented for digital imagery. x yx p yq f (x, y) (2) M pq = 2. Center moment: centroid image is used to normalize invariant translations in the image field. M10 ¯ M01 X¯ = ,Y = M00 M00 Furthermore, the central moment can be determined discretely as follows: ¯ q ¯ p (Y − Y) x y(X − X) M pq = 3. Normalization of the central moment: to normalize the change. µ pq n pq = µ00 γ

(3)

(4)

(5)

4. Seven non-linear function: normalize the signature of scale change, translation, and rotation. Φ1 = n20 + n02 Φ2 = (n20 + n02 )2 + 4n20 2 11 Φ3 = (n30 + 3n12 )2 + (n30 + 3n21 )2 Φ4 = (n30 + n12 )2 − (n30 + 3n21 )2

Φ5 = (3n30 − 3n12 )(n30 + n12 )[(n30 + n12 )2 3(n21 + 3n03 )2 ](n21 − n03 )(n21 + n03 )x[3(n30 + n12 )2 − (n21 + n03 )2 ]

Φ6 = (n20 − n02 )[(n30 + n12 )2 − (n21 + n03 )2 ]4n11 (n30 − n12 )(n21 + n03 )

Φ7 = (3n21 − n03 )(n30 + n12 )[(n30 + n12 )2 3(n21 + n03 )2 ](3n12 − n30 )(n21 + n03 )x[3(n30 + n12 )2 − (n21 + n30 )2 ]

Obtained results: Φ1 = 2.2846; Φ2 = 0.097881; Φ3 = 1.3926; Φ4 = 0.027233; Φ5 = −0.0010864; Φ6 = 0.0020606; Φ7 = 0.0051911; The invariant moment value is then normalized using min-max normalization in order to generate values in the range of 0 to 1. Having obtained the value of normalized invariant moment, then is clustering data with FKCN algorithm. This algorithm combines the learning rate of Fuzzy C-Mean (FCM) algorithm and KCN’s all vector update. FKCN algorithm works with following stages: 1. Compute the learning rate:

mt = m0 − t ∗ ∆m, ∆m = (m0 − 1)/tmax uik,t =

||Xk − Vi,t−1 || m−1 ||Xk − V j,t−1 || 2

−1

(6) (7)

2. Determination of FKCN learning rate: αik,t = (uik,t )mt

(8)

624 4


3. Update all weight vector: vi,t = vi,t−1 +

n

k=1

αik,t (xk − vi,t−1 ) n j=1 αi j,t

(9)

4. Compute: Et = ||Vt − Vt−1 ||

(10)

5. If Et < E, else t = t + 1 then back to first step. After obtaining cluster center data invariant moment invited using Euclidean distance by using cluster center FKCN. In this study used EFKCN which update the learning rate can be adjusted based on fuzzy membership to produce optimal learning rate 1,5 . EFKCN improved learning rate: tu = fuzzy membership upper limit and td = the lower limit of fuzzy membership.    (uik,t )mu , uik,t > tu     αik,t =  (uik,t )mt , td ≤ uik,t ≤ tu     (uik,t )md , uik,t ≤ td tu = 0.7; td = 0.3; mu = 0.4; md = 3;

The clustering results are then evaluated by testing 30 signatures. The data in the form of a signature will be identified by the range of distance to the cluster center, if it is within the specified Euclidean minimum distance, the data will be recognized as the original signature and vice versa, if the signature is beyond the specified range, it is considered not as an original signature. After believed the test results show good results, then developed an application that can be used to verify access to personal data.

4. Experiment and Result Signature Data Training Sample signatures in the process by means of a scan so that the format of RGB, which is then converted into grayscale format with the aim to more easily converted into binary image. The converted binary image has a pixel value ‘0’ for the black signature pattern and ‘1’ for the white background. The binary image is then converted into an inverted binary image so that the black background has a pixel value of ‘1’ and the object becomes white with pixel value ‘0’. Eliminating the border is meant for a more unified signature pattern which can then be extracted bounding box so that the unnecessary background will be removed and only take the signature pattern. The extracted results data are then normalized using Hu’s moment invariant so that it is invariant to scale, rotation, and transformation changes. Data moment invariant then normalized using min-max normalization. The functionality of the training feature is intended to select the signature to be tested in order to determine the signature owner (Fig. 1). For training needs, signatures are performed in a variety of conditions whether sitting, standing, and fast conditions. The facts show that signatures under different conditions affect the shape of the signature. Training conducted on 50 signatures from 10 respondents, obtained training results in the form of a cluster center for each signature owner as Fig. 2. The signature to be tested will be processed in order to find the Euclidean distance value to determine the cluster data center of the tested signature. If the Euclidean distance can be within a predetermined distance then the signature will go into the cluster. Before entering, the process must be through the combination of data first as can be seen in Fig. 3.


625 5

Fig. 1: Screen display data and training results where the left side is the results of the right side data

Fig. 2: Signatures of 10 respondents used in the training. The label of each signature is attached on top of the image Table 1: Rule combination after normalization

Data 1 with data 2. Data 1 with data 3. Data 1 with data 4. Data 1 with data 5. Data 1 with data 6. Data 1 with data 7. Data 4 with data 5. Data 4 with data 6. Data 4 with data 7.

Data 2 with data 3. Data 2 with data 4. Data 2 with data 5. Data 2 with data 6. Data 2 with data 7.

Data 3 with data 4. Data 3 with data 5. Data 3 with data 6. Data 3 with data 7.

Data 5 with data 6. Data 5 with data 7.

Data 6 with data 7.

Seven data moment invariant that has been normalized with min-max normalization will be combined with the rules as described in Table 1. The combination rules are divided into six variants where the first rule is the merging of data 1 to data 7, then the data 2 collaborates with the data 3 to data 7 as the second rules, and so on.

626 6


Fig. 3: Screen display data clustering and training results Table 2: Accuracy signature testing data using EFKCN algorithm

Cluster Respondent 1 2 3 4 5 6 7 8 9 10

Signature of Respondent to1 2 3 6 9 6 2 2 2 3 10 5 4 4 4 5 5 8 6 6 6 7 7 7 4 4 4 9 9 9 10 10 10 Accuracy Average

Accuracy (%) 0% 100% 33.3% 100% 66.7% 100% 100% 0% 100% 100% 70%

Signature Data Clustering The results of the training data are then used to test the results of signature identifiers by the clustering method. The result of clustering test with EFKCN algorithm is as can be seen in Table 2 as follows. The test results show that EFKCN is able to recognize the signatures as much as 70% with the other 30% leading to the signatures of other respondents. In the training and test data, there are 3 data that have low accuracy of 1, 3, and 8 person signatures respectively with 33.3% and 0% accuracy (based on Table 2). Although the signatures generated by each person are different but the value of feature extraction uses the person invariant moments of 1.3 and 8 are almost identical with others causing the classification of the signature to point to different clusters. The 1st person has similarity with the invariant moment value of the 6th and 9th persons, the 3rd person is similar to the 5th and 8th persons invariant value, and the 8th person has the same invariant moment with the 4th person. In Table 2, the accuracy of the training data is smaller than the accuracy of the test data. Where should the value of test accuracy is smaller than the value of training accuracy. This is because a person’s signature may not always be exactly the same between one signature with another signature so, it does not rule out that the accuracy of the test results is greater than the accuracy of the training.


627 7

Signature Recognition for Personal Data Verification A signature recognition system with better accuracy than the previous research is then developed into an application for verification of personal data in order to limit access to personal data. The appearance of the created application is depicted in Fig. 4.

Fig. 4: The application appearance with Otnel’s signature as an example that does not match with the recognition result

Testing of forged signatures in the case of this study is actually the property of Otnel’s signature as the example. The signature does not match the personal data of each person. The signature shows the clusters of others. And the invariant moment obtained from the signature is much different. The comparison can be seen in the Fig. 4 and 5.

Fig. 5: Moment invariant of Otnel’s signature

As we can see in Fig. 5, the comparison between invariant moments of the original signatures and the forged signature of Otnel respondents is as follows: Otnel: 2.8238; 5.4619; 1.6474; 0.82552; 0.95825; 1.7105; 0.092353 Forged Otnel: 1.6117; 0.70283; 1.1087; 0.18211; 0.079151; 0.1293; 0.020746

628 8


5. Conclussion The same signature pattern of respondents may differ depending on the condition of the signature so as to allow an error in the recognition of the signature pattern. The recognition of signature patterns using the clustering method with the EFKCN algorithm shows relatively better results with 70% accuracy compared to the accuracy of previous research results 2 which is 53%. A good signature pattern recognition can be developed to assist the verification system as well as the personal data verification system made in this study. References 1. Yang, Y., Jia, Z., Chang, C., Qin, X., Li, T., Wang, H., et al. An efficient fuzzy kohonen clustering network algorithm. In: Fuzzy Systems and Knowledge Discovery, 2008. FSKD’08. Fifth International Conference on; vol. 1. IEEE; 2008, p. 510–513. 2. Chaudhari, B.M., Barhate, A.A., Bhole, A.A.. Signature recognition using fuzzy min-max neural network. In: Communication and Energy Conservation 2009 International Conference on Control, Automation. 2009, p. 1–7. 3. Ahmed, H., Shukla, S., Rai, H.M.. Static Handwritten Signature Recognition Using Discrete Random Transform and Combined Projection Based Technique. In: 2014 Fourth International Conference on Advanced Computing Communication Technologies. 2014, p. 37–41. doi: 10.1109/ACCT.2014.76. 4. Bezdek, J.C., Tsao, E.C.K., Pal, N.R.. Fuzzy Kohonen clustering networks. In: [1992 Proceedings] IEEE International Conference on Fuzzy Systems. 1992, p. 1035–1043. doi:10.1109/FUZZY.1992.258797. 5. IRWANSYAH, E., FAISAL, M., PRIMADINI, A.. Does efficient fuzzy kohonen clustering network algorithm really improves clustering data result? Journal of Theoretical & Applied Information Technology 2015;71(1). 6. Salambue, R.. Pengenalan Pola Tanda Tangan dengan Metode Momennt Invariant dan Euclidean Distance. Prosiding SEMIRATA 2013 2013; 1(1). URL http://jurnal.fmipa.unila.ac.id/index.php/semirata/article/view/918. 7. Ainun, J., Mohammad Isa, I., Imam, M.. PENGENALAN POLA TANDA TANGAN MENGGUNAKAN METODE MOMENT INVARIANT DAN JARINGAN SYARAF RADIAL BASIS FUNCTION (RBF). Pemantapan Keprofesionalan Peneliti, Pendidik, dan Praktisi MIPA Untuk Mendukung Pembangunan Karakter Bangsa 2011;URL http://www.uny.ac.id. 8. Hu, M.K.. Visual pattern recognition by moment invariants. IRE Transactions on Information Theory 1962;8(2):179–187. doi: 10.1109/TIT.1962.1057692. 9. Ismail, I.A., Ramadan, M.A., Danf, T.E., Samak, A.H.. Automatic Signature Recognition and Verification Using Principal Components Analysis. In: Imaging and Visualisation 2008 Fifth International Conference on Computer Graphics. 2008, p. 356–361. doi:10.1109/CGIV.2008.8. 10. Simpson, P.K.. Fuzzy min-max neural networks. I. Classification. IEEE Transactions on Neural Networks 1992;3(5):776–786. doi: 10.1109/72.159066. 11. Wold, S., Esbensen, K., Geladi, P.. Principal component analysis. Chemometrics and intelligent laboratory systems 1987;2(1-3):37–52. 12. Otsu, N.. A Threshold Selection Method from Gray-Level Histograms. IEEE Transactions on Systems, Man, and Cybernetics 1979;9(1):62– 66. doi:10.1109/TSMC.1979.4310076.