signature=127df89cc6d60119127f71d7084ecdc0,TIET Digital Repository

Abstract:Biometrics is defined as recognizing an individual on the basis of his/her distinctive characteristics. Several biometric verification techniques have been implemented to enhance the security. Amongst all the possible biometric modalities, the handwritten signature has been used for the longest period of time as a means of identification. Its application is mostly present in banking transactions, cheque authentication, online shopping, legal documents etc. These applications increase the demand for the use of signature for the identification purpose. This leads to the development of documents like electronic passports. Accurate verification can be done by using online signatures to achieve better results. Several verification techniques used are Dynamic Time Warping (DTW), Bayesian Learning, Hidden Markov Model (HMM), Neural Networks, Support Vector Machine (SVM), etc.

Many techniques have been applied on online signatures till now. One such technique employed Discrete Cosine Transform (DCT) which considered six features for verification process. However, only 7.04% of Equal Error Rate (EER) was achieved when DCT was applied. When the number of features got increased from six to ten, the EER further reduced to 5%. In order to attain much better results, a new methodology has been proposed in this dissertation which makes use of DCT with

Principal Component Analysis (PCA) as a feature selector. PCA reduces the large dimensionality of data space to the small intrinsic dimensionality of the feature space, so as to capture the maximum variability in the data. The advantages of PCA include smaller representation of data along with the reduction in the noise. As a result, 3.5% of EER is obtained when eight features are considered, which clearly proves the superiority of DCT with PCA over DCT only.

Another effort has been put up to improve the signature verification which uses

Discrete Fractional Cosine Transform (DFrCT). TheDFrCT is a general form of DCT which has an additional free parameter, its fraction that gives an extra degree of freedom. But calculations done in the case of DFrCT are more than DCT, which increases the computational complexity of the signature verification system. When ten features of the test signature are extracted and verified using this technique, an EER of 3.25% is attained. Further, another method is proposed in this dissertation that combines DFrCT with PCA for achieving more appropriate results. 3% of EER is obtained using eight features which is by far the least EER achieved.

The experiments described above are performed over a number of features ranging from five to ten. SVC2004 database is used for carrying out the work. In case of DCT and DFrCT alone, it is verified that the EER is minimum when ten features are used independently. When these techniques are combined with PCA, minimum EER is obtained at eight features only, thus reducing the computations.