Wavelet-based compared to function-based on-line signature verification

We implemented two direct methods for on-line signature verification. First, data produced by a graphics tablet describing a signature to be tested are treated with wavelet transforms to generate features to be nonlinearly confronted with a reference signature chosen among 10 previously stored from the same writer. In order to recover the time dependence lost in the wavelet treatment, we included the level of departure from the diagonal line in the warping function as a complementary measure of distance. In a second approach, the functions x(t) and y(t) describing position in time of each pixel of the same test signature were directly (though nonlinearly) compared to their counterparts from the reference. We concluded that both approaches showed good fidelity to all details in the signatures, with acceptable false rejection rates (we obtained around 30% FRR) to this kind of biometry. On the other hand, the inclusion of the wavelet transform turned out to be an essential step for the achievement of low false acceptance rates. It was only with the inclusion of the wavelet transform, at the right level of resolution, that we managed to completely prevent trained forgeries to be accepted (0% FAR) in the cases studied.