Study on the effect of number of training samples on HMM based offline and online signature verification systems

Abstract

This paper reports on the effect that the number of samples used in training a signature verification system has on the system's accuracy which is describes by the pair combination of the system False Acceptance Rate (FAR) and False Rejection Rate (FRR). This paper also describes such an effect on the system Failure to Capture Rate (FCR), which is often neglected by biometrics study. The paper covers both online and offline signature verification systems developed using Hidden Markov Models (HMMs). Experimental results are based on Sigma - database of Malaysian signatures with over 6,000 genuine samples and 2,000 skilled forgeries collected in a real life environment.