Sunteți pe pagina 1din 23

Signature Verification Using DWT and Neural Network

BY

DIVYA SWAPNA

ECE II/IV MRCEW

INTRODUCTION
The signature verification is an established and accepted form of verification. It is an important source of authentication about ones identity. However, with the increasing use of computer network , it is difficult to confirm whether information had been truly sent by the person whose name is mentioned or someone else in unsecured data networks.

INTRODUCTION contd

Moreover the present approach of using passwords for accessing restricted database is also cumbersome.

So it is proposed to use users mouseentered signature to access the restricted database for moderate security category.

Aim of the presentation


Processing and feature extraction of the signature

Verification by using Neural Network

Start

SYSTEM DESCRIPTION

Enter signature using Mouse

Pre-processing

Slope and Wavelet feature extraction

Verification using feed forward neural network

No

Is signature verified?

Display error

Yes

Stop

MOUSE PROGRAMMING
Workspace of 200 x 100 pixels was provided for user to enter signature. Functions were developed to restrict the mouse pointer within this specified workspace during signature insertion process. The x-y co-ordinates of present mouse position were continuously captured in an array.

PRE-PROCESSING
The number of points captured for mousewritten signature varies w.r.t size and speed of writing. It is therefore, desired to resample the signature contour data to reduce the noise. Re-sampled signature contour would contain fixed number of points. 100 points were chosen for re-sampling.

PRE-PROCESSING contd FILTERING


Filtering is done to avoid spurious points in signature due to noise. Any point having value beyond the workspace region was treated as spurious point and was replaced by filtered point, which is the average of two neighboring points.

FEATURE EXTRACTION
In this application slope and DWT coefficients are extracted for verification. SLOPE: Slope is calculated for each two successive points obtained from re-sampling. Each individual has a unique way of running signature and also slope parameter.

Features Extraction contd

WAVELET:
The discrete wavelet transform (DWT) for a signal, f(n) with period N is given by,
N 1

DWT f [n,a ]=

f [m]a

 j /2

= ((m  n) / a )

m !0

Where m and n are integers, a=2 for a dynamic transform And (n) is mother wavelet with zero mean value.

Features Extraction contd


The signal representation is not complete if wavelet decomposition is computed to a scale. The information corresponding to the scales larger than is computed by a scaling filter and is given by,
N 1

SF f[n, ]= a

m!0

f [ m]a

 j /2

** (( m  n) / a j )

Where *(n) is discrete scaling filter

Features Extraction contd


The wavelet feature of the signature was extracted by applying DWT to the vectors X and Y separately. The approximated coefficients and the detailed coefficients of DWT of X and Y were considered as the wavelet features for signature verification. Daubechies wavelet of the order 4 was used for DWT.

NEURAL NETWORK BASED VERIFICATION PROCESS


The multi layer feed-forward neural network was used for verification process. Processing elements (PEs), which compute a non-linear function of scalar product of input vector and weight vector. PEs in a layer are connected to all the neurons by links and connecting weights.

Neural network based contd

Initially activation functions and the thresholds are defined by a recursive optimization technique.

Connection weights are then determined by means of back-propagation learning algorithm.

Neural network based contd

Learning Phase:

Neural network is trained separately for each target model with 10 samples of mouse-entered signatures of the target and 10 signatures of the impostor. Zero mean random Gaussian noise of 5% variance was added during the training to check robustness of the system.

Neural network based contd

Testing Phase:
Testing phase was used to verify the inserted signature by comparing their features.

Fig. Normalized Signatures

Fig. Approximation DWT coefficients of X and Y against number of points N

Fig. Detailed DWT coefficients for X against number of points N

Fig. Approximation for Y against number of points N

Fig. Detailed DWT coefficients for Y against number of points N

APPLICATIONS
Business and financial transactions Checking forgeries in high security applications Banks ,telecom companies ,insurance companies.

Results and Conclusions


The results show that success rate was very good (95% to 98.5%) using wavelet feature. Features like slope resulted only in marginal improvement in the accuracy (96% to 99%). It has been observed that the output of neural network shows robustness of system against unauthorized usage.

Conclusion contd

For person having skewed signature with variable inclination, alignment can be done along X-axis using histogram technique before per-processing phase. This approach can also be easily applied to enter non-English alphabets (for eg. Devnagri script letters) without the need to use cumbersome encoding of keyboard keystrokes and their combinations.

Thank You.. Any Queries ????

S-ar putea să vă placă și