Multi-Kernel Approach to On-Line Signature Verification

V. Sulimova, V. Mottl, and A. Tatarchuk (Russia)


Signal recognition, on-line signature verification, time warping, kernel function, kernel fusion.


The problem of on-line signature verification is considered within the bounds of the kernel-based methodology of pattern recognition and, more specifically, SVM principle of machine learning. In accordance with this methodology, any set of on line signatures as vector signals of individual length is repre sented by a two-argument function which measures the pair wise similarity between respective signals and possesses the properties of a kernel, i.e. inner product in a hypothetical linear space. Since the SVM principle completely predefines the al gorithms of both training and recognition, it remains only to choose a kernel produced by an appropriate metric in the set of signatures, so that the genuine signatures of the same person would be much closer to each other than those of different per sons. However, different viewpoints of signature similarity lead, a priori, to different kernels. We propose a principle of fusing several kernels into an entire training and verification technique. Experiments with the data base of the World Signa ture Verification Competition 2004 have shown that multi kernel verification essentially decreases the error rate in com parison with decision rules based on single kernels.

Important Links:

Go Back