IAETSD JOURNAL FOR ADVANCED RESEARCH IN APPLIED SCIENCES ISSN NO: 2394-8442

Handwritten Signature Verification using Hybrid Wavelet

Transform (HWT) – 2
1
Manoj Chavan, 2 Ravish R. Singh, 3 Vinayak Bharadi
1 Facultyof Engineering, Pacific Academy of Higher Education and Research University, Udaipur, India,
2 Thakur Educational Trust, Mumbai, India,

3 Information Technology Department, Finolex Academy of Management and Technology, Ratnagiri (MH), India

1 prof.manoj@gmail.com, 2 ravishrsingh@yahoo.com , 3 vinayak.bharadi@famt.ac.in

Abstract— Offline signatures contains two dimensional image of the signatures. Online signature contains additional
dynamic features such as pressure applied by user, speed of writing, method of holding the pen etc. along with two
dimensional image of the signatures. The HWT which are formed by using Kronecker product of two transform, has the
ability to analyse the signal at global as well as local level. HWT - 2 was applied on the first 128 samples of the pressure
parameter and first 1- 16 samples of the output were used as feature vector for signature verification. Using Hidden
Markov Model (HMM) based classifier with ergodic model, the performance of the proposed system was analysed.
KEKRE 128 offers best performance with FRR & FAR of 13%. Orthogonal DCT transform, DCT HADAMARD, DHT
DCT and HAAR HADAMARD transform offers best performance of 4 training samples. KEKRE 128 offers best
performance for 2 to 4 states and DCT KEKRE 128 offers best performance for state 5. The best performance is at 500
symbols by KEKRE 128 and DCT KEKRE.

Keywords— HMM, HWT - 2, Online Signature Verification, FRR, FAR, EER.

I. INTRODUCTION
Any individual can be uniquely identified by using biometric features. [1] These features are of two types, namely,
Physical and behavioural. Handwritten Signatures fall in the category of behavioural features which also include voice,
typing rhythm etc. Signatures have been used for a long time for authenticating an individual as it is very easy to obtain. [2]
As a biometric feature, signatures are of two type, offline and online signatures. Offline signature is a two dimensional
image of the signature, done on the paper whereas the online signature contains additional dynamic features such as
pressure applied by user, speed of writing , method of holding the pen etc. along with two dimensional image of the
signatures. [3] For computerizing the process of signature verification, online signatures offers more advantages than the
offline signatures due to its dynamic features.

II. PROPOSED TECHNIQUE

In general biometric system consist of data acquisition, pre-processing, feature extraction, classification and
performance evaluation. In this system, the user will sign some sample signatures on pressure sensitive writing pads. One
of such writing pads, Wacom Intuos Pro is shown in the fig. 1. The proposed method is as shown fig. 2. Every signature
sample will have X-coordinate - scaled cursor position along the x-axis, Y-coordinate - scaled cursor position along the y-
axis, Time stamp - system time at which the event was posted, Button status - current button status (0 for pen-up and 1
for pen-down), Azimuth - clockwise rotation of cursor about the z-axis, Altitude - angle upward toward the positive z-axis,
Pressure - adjusted state of the normal pressure. We used signature database provided by The First International Signature
Verification Competition (SVC 2004) consisting of 1600 signatures. It has signature samples of 40 users. Every user has 40
signature samples, out of which, 20 are genuine and 20 are skilled forgeries. [4]

VOLUME 4, ISSUE 7, DEC/2017 174 http://iaetsdjaras.org/

IAETSD JOURNAL FOR ADVANCED RESEARCH IN APPLIED SCIENCES ISSN NO: 2394-8442

It is observed that out of the seven features of online signatures, Pressure applied by the tip of the pen on the pressure
sensitive pad has the highest level of variation and has therefore been chosen for the feature vector. Various orthogonal
transforms have been used to analyse the overall signal thereby giving its global properties. Wavelet Transform can analyse
the signal at various resolution level thereby extracting its global as well as local properties. The orthogonal transforms are
used to analyse the global properties of a signal. The HWT which are formed by using Kronecker product of two
transform, has the ability to analyse the signal at global as well as local level [5].

Fig. 1: Wacom Intuos Pro Pressure Sensitive Pad

Fig. 2: Proposed System

III. METHODOLOGY
HWT-2 matrix ‘TXY’ of size (NxN) can be formed by the Kronecker product of two orthogonal transform matrices X
and Y respectively, with sizes (a x a) and (b x b), such that N=ab. For HWT-2 matrix, first N/2 rows of the matrix are
formed by product of each element of first a/2 rows of the matrix X with each of the columns of the matrix Y. For next
‘b’ number of rows of matrix, the ‘a/2+1’th row of the orthogonal transform matrix X is shift rotated after being
appended with zeros. Next N/2 rows are generated as set of b rows each time for each of the ‘a/2’ rows of orthogonal
transform matrix X starting from ‘a/2+1’th row up to last row. [6] We have used Discrete Cosine transform (DCT),
Discrete Hartley transform (DHT), Discrete Walsh transform (DWT) and Discrete Kekre transform (DKT) to form the
Wavelet and HWT-2. [7] The first 128 samples of every signature is used to find HWT. The first 16 samples and the
samples 33 to 64 of the output are used as feature vector.

VOLUME 4, ISSUE 7, DEC/2017 175 http://iaetsdjaras.org/

IAETSD JOURNAL FOR ADVANCED RESEARCH IN APPLIED SCIENCES ISSN NO: 2394-8442

There are many topologies of HMM such as Left to Right, Ergodic and Ring. Ergodic model as shown in fig 2.

Fig 2: Ergodic model

Classifiers are trained by using the sample training signatures. After training they are used to verify, the feature of the
test signature with those stored in the database to decide whether the signature is genuine or forged. Various classifiers
used are classifiers based on Dynamic Time warping [8], Neural Networks (NN) [9], Hidden Markov Model (HMM) [10]
[11] and Support Vector Machine (SVM). [12]

We have used HMM based classifier. Consider a system which is in a distinct state (S1, S2, …., SN) at any point of time.
In Markov Models the future state is always dependent on the current state and is independent of past state P (St=j / St-
1=i, St-2=k …) = P (St=j / St-1=i). The states and associated transition probabilities are known to the user. In case of
Hidden Markov Model states will not be known but the output which is dependent on the state, will be known to user. [13]
HMM is represented by the transition probability matrix (A), Observation matrix (B) and initial probability distribution
matrix (π). Various parameters of HMM are as follows:

Number of states (N) in the model: In this experiment the number of states are varied from 2 to 5. As the number of
states increase, the time needed for training increases.

Number of observations (M) corresponding to each state: In this experiment the number of observations are varied
from 200 to 750 in the increments of 25. The output of HWT is a matrix of dimension [1 × 128]. The matrix elements
from 1 to 16 and 33 to 64 are chosen as a feature vector. Feature vectors are scaled into M number of observations.

Initial Probability Distribution (π): πi = P (q1= Si); 1 ≤ i ≤ N. For the Ergodic HMM, the initial probability of the first
state is 1 and the others are 0. This implies that at the beginning, the HMM is always in state 1.

State transition probability (aij): aij = P (St=j / St-1=i ). For the left-to-right HMM, aij=0 when i>j. we are using the
HMM of first order so that aij=0 when j>i+1. Initially, the state transition matrix is generated using the random numbers
such that ; 1 ≤ i ≤ N.

Observation probability (bj): bj (k) = P (Vk at t / qt = Sj); 1 ≤ j ≤ N; 1 ≤ k ≤ M; the probability of generating a symbol
Vk in state j.

We have used statistics and machine learning toolbox of the MATLAB 13. Initially we assumed a randomly generated
transition probability Matrix (A) generated using MATLAB, Observation probability matrix (B) with equal probability for
every symbols and HMM to be in state 1. HMM is trained using hmmtrain for 3 to 20 genuine training signature samples,
number of states from 2 to 5 and symbols from 200 to 750. After HMM is trained, it is used to test 20 genuine and 20
forged signatures of 40 users.

VOLUME 4, ISSUE 7, DEC/2017 176 http://iaetsdjaras.org/

IAETSD JOURNAL FOR ADVANCED RESEARCH IN APPLIED SCIENCES ISSN NO: 2394-8442

IV. RESULT AND DISCUSSION

Performance of the system will be measured on the basis of False Rejection Ratio (FRR) and False Acceptance Ratio
(FAR). FRR refers to false rejection of genuine signature and FAR refers to false acceptance of forged signature. [14] FRR
will be computed as ratio of the number of signatures detected as forged to the total number of genuine signatures tested.
FAR will be computed as ratio of the number of signatures detected as genuine to the total forged signatures tested.
Testing has been carried out for 40 users and then the average FRR and FAR are calculated. In FRR-FAR plot, the point
where two graphs cross each other is referred as Equal Error Rate (EER). At this point the value of FRR and FAR is
minimum. We choose the EER such that it is the nearest number of training samples such that FRR and FAR are
minimum.

Fig. 3: FRR – FAR plot

The performance of various DCT combinations is shown in the Table II.

Best FRR – FAR: FRR – FAR should be as low as possible. DCT KEKRE offers best performance with FRR 22% &
FAR of 31 %. All combinations of DCT offer better performance than Orthogonal DCT transform. Best EER: EER in
terms of training samples should be as low as possible. Orthogonal DCT transform and DCT HADAMARD offers best
performance of 4 training samples compared to all combinations of DCT HWT.

VOLUME 4, ISSUE 7, DEC/2017 177 http://iaetsdjaras.org/

IAETSD JOURNAL FOR ADVANCED RESEARCH IN APPLIED SCIENCES ISSN NO: 2394-8442

Best state wise FRR – FAR: FRR – FAR should be as low as possible for the given state from 2 to 5. DCT KEKRE offers
best performance for 2 to 5 states compared to orthogonal DCT transform Best Number of Symbol: It should be as low
as possible. Testing was carried out for number of symbols from 200 to 750. It evident that the best performance in terms
of FRR – FAR is offered by 450 – 500 symbols.

The performance of various DHT combinations is shown in the Table II.

Best FRR – FAR: DHT KEKRE offers best performance with FRR & FAR of 24%. Only DHT KEKRE offers better
performance than Orthogonal DHT transform.

Best EER: DHT DCT offers best performance of 4 training samples compared to 5 training samples for Orthogonal
DHT transform.

Best state wise FRR – FAR: DHT KEKRE offers best performance for 2 to 5 states compared to orthogonal DHT
transform.

Best Number of Symbol: The best performance in terms of FRR – FAR is offered by 450 - 500 symbols

The performance of various HAAR combinations is shown in the Table II.

Best FRR – FAR: HAAR KEKRE offers best performance with FRR 20% & FAR of 21%. HAAR DCT and HAAR
KEKRE offers better performance than Orthogonal HAAR transform.

VOLUME 4, ISSUE 7, DEC/2017 178 http://iaetsdjaras.org/

IAETSD JOURNAL FOR ADVANCED RESEARCH IN APPLIED SCIENCES ISSN NO: 2394-8442

Best EER: HAAR HADAMARD offers best performance of 4 training samples compared to 6 training samples for
Orthogonal HAAR transform.

Best state wise FRR – FAR: HAAR KEKRE offers best performance for 2 to 5 states compared to orthogonal HAAR
transform. Best Number of Symbol: The best performance in terms of FRR – FAR is offered by 400 – 500 symbols.

The performance of various HADAMARD combinations is shown in the Table III.

Best FRR – FAR: HADAMARD KEKRE offers best performance with FRR 30% & FAR of 32%. The performance
offered by all combinations of HADAMARD is better than Orthogonal HADAMARD transform.

Best EER: Orthogonal HADAMARD transform and HADAMARD DCT offer best performance of 5 training samples
compared to all combinations of HADAMARD HWT

Best state wise FRR – FAR: HADAMARD DHT offers best performance for 2 and 4 states and HADAMARD KEKRE
offers best performance for 3 and 5 states compared to orthogonal HADAMARD transform.
Best Number of Symbol: The best performance in terms of FRR – FAR is offered by 475-500 symbols

The performance of various KEKRE combinations is shown in the Table III.

Best FRR – FAR: KEKRE 128 offers best performance with FRR 5% & FAR of 2%. The performance offered by
KEKRE 128 is better than all combinations of KEKRE HWT.

VOLUME 4, ISSUE 7, DEC/2017 179 http://iaetsdjaras.org/

IAETSD JOURNAL FOR ADVANCED RESEARCH IN APPLIED SCIENCES ISSN NO: 2394-8442

Best EER: KEKRE HADAMARD offers best performance of 6 training samples compared to 16 training samples for
Orthogonal KEKRE transform.

Best state wise FRR – FAR: KEKRE 128 offers best performance for 2 to 5 states compared to combinations of KEKRE
HWT.

Best Number of Symbol: The best performance in terms of FRR – FAR is offered by 450 - 500 symbols.

V. CONCLUSION
Comparing the various combinations of DCT, DHT, HAAR, HADAMARD and KEKRE for 1 – 16 samples of HWT
1 and 2 for pressure as feature vector we come to following conclusions:
1) FRR – FAR: KEKRE 128 offers best performance with FRR & FAR of 13%.
2) EER: Orthogonal DCT transform, DCT HADAMARD, DHT DCT and HAAR HADAMARD transform offers
best performance of 4 training samples.
3) State wise FRR – FAR: KEKRE 128 offers best performance for 2 to 4 states and DCT KEKRE 128 offers best
performance for state 5.
4) Number of Symbol: The best performance is at 500 symbols by KEKRE 128 and DCT KEKRE.
5) In terms of FRR-FAR, all combinations of DCT offer better performance than DCT orthogonal transform. DHT
KEKRE offers better performance than DHT orthogonal transform. HAAR DCT and HAAR KEKRE offer better
performance than HAAR orthogonal transform. All HWT of HADAMARD offer better performance than
HADAMARD orthogonal transform. KEKRE orthogonal transform offers better performance than all HWT of KEKRE.

REFERENCES
[1] A. K. Jain, A. Ross, S. Prabhakar, “An Introduction to Biometric Recognition”, IEEE Transactions on Circuits and Systems for
Video Technology, Vol. 14, No. 1, January 2004
[2] S. Nanavati, M. Thieme, and R. Nanavati, Biometrics: Identity Verification in a Networked World. New York: Wiley, 2002, pp.
123–131.
[3] T. Ohishi, Y. Komiya, H. Morita, and T. Matsumoto, “Pen-input online signature verification with position, pressure, inclination
trajectories,” in Proc. 15th Int. Parallel Distrib. Process. Symp. (IPDPS-15), San Francisco, CA, Apr. 2001, p. 170.
[4] [Online]. Available: www.cse.ust.hk/svc2004/download.html. [Accessed 1 January 2016].
[5] H B Kekre, T K Sarode, S D Thepade, “Inception of HWT using Two Orthogonal Transforms and It’s use for Image Compression”,
(IJCSIS) International Journal of Computer Science and Information Security, Vol. 9, No. 6, 2011
[6] Vinayak Ashok Bharadi, Vikas I. Singh and Bhushan Nemahe, “Hybrid Wavelets based Feature Vector Generation from
Multidimensional Data set for On-line Handwritten Signature Recognition” Proc.57th Int. Conf. Confluence The next Generation Information
Technology Sumit (Confluence) Noida, India, Nov. 2014.
[7] Dr. H.B.Kekre, Archana Athawale, Dipali Sadavarti, “Algorithm to generate Wavelet Transform from an orthogonal transform”,
International Journal of Image Processing (IJIP), vol. 4, no. 4, pp. 444–455, Oct. 2010
[8] L. Bovino, S. Impedovo, G. Pirlo, and L. Sarcinella, “Multi-expert verification of hand-written signatures,” in Proc. 7th Int. Conf. Doc.
Anal. Recognit. (ICDAR-7), IEEE Comput. Soc., Edinburgh, U.K., Aug. 2003, pp. 932–936.
[9] H. D. Chang, J. F. Wang, and H. M. Suen, “Dynamic handwritten Chinese signature verification,” in Proc. 2nd Int. Conf. Doc. Anal.
Recognit. (ICDAR-2), Tsukuba, Japan, Oct. 1993, pp. 258–261.
[10] Manoj Chavan, Ravish R Singh, Vinayak Bharadi, “Online Signature Verification using HWT with Hidden Markov Model” in Proc
4th International Conference on Computing, Communication, Control And Automation (ICCUBEA), Pune, India Aug 2017.
[11] Manoj Chavan, Ravish R Singh, Vinayak Bharadi, “Handwritten Signature Verification using Hidden Markov Model with HWT”
in Proc 4th International Conference on Computing, Communication, Control And Automation (ICCUBEA), Pune, India Aug 2017.
[12] E. J. R. Justino, F. Bortolozzi, and R. Sabourin, “A comparison of SVM and HMM classifiers in the offline signature verification,”
Pattern Recog- nit. Lett. vol. 26, pp. 1377–1385, 2005.

VOLUME 4, ISSUE 7, DEC/2017 180 http://iaetsdjaras.org/

IAETSD JOURNAL FOR ADVANCED RESEARCH IN APPLIED SCIENCES ISSN NO: 2394-8442

[13] A. P. Thad Starner, "Real-Time American Sign Language Visual Recognition From Video Using Hidden Markov Models," Master's
Thesis, MIT, Program in Media Art, Feb 1995.
[14] Donato Impedovo and Giuseppe Pirlo, "Automatic Signature Verification: The State of the Art," IEEE Transaction on Systems,
MAN and Cybernatics Part C: Application and Reviews, vol. Vol 28, no. No 5, September 2008.

VOLUME 4, ISSUE 7, DEC/2017 181 http://iaetsdjaras.org/