Offline signature verification with generated training samples

Abstract

It is often difficult to obtain sufficient signature samples to train up a signature verification system. An elastic matching method to generate additional samples is proposed to expand the limited training set so that a better estimate of the statistical variations can be obtained. The method differs from existing ones in that it is more suitable for the generation of signature samples. Besides this, a set of peripheral features, which is useful in describing both the internal and external structures of signatures, is employed to represent the signatures in the verification process. Results showed that the additional samples generated by the proposed method could reduce the error rate from 15.6% to 11.4%. It also outperformed another existing method which estimates the class covariance matrix through optimisation techniques. Results also demonstrated that the peripheral features are useful for signature verification.