A Bayesian approach to off-line signature verification

A new approach to off-line signature verification using non parametric Bayesian approach is discussed. Given sample(s) of Genuine signatures of an individual, the task of signature verification is a problem of classifying a questioned(test) signature as Genuine or Forgery(skilled). The verification problem is a two step approach (i)Enrollment: Genuine signature samples of an individual are provided. The method presented here maps from features space to distance space by comparing all the available genuine signature samples amongst themselves to obtain a distribution in distance space “within person distribution”. This distribution captures the variation and similarities that exist within a particular person’s signature; (ii)Classification: The questioned signature to be classified, is then compared to each of the genuine signatures to obtain another distribution in distance space “Questioned vs Known distribution”. The two distributions are then compared using a new Bayesian similarity measure to test whether the samples in the distribution are from the same distribution(Genuine) or not(Forgery). Noisy distance distributions that could be formed due to presence of multiple types of signature are avoided by detecting such samples in advance by using a normalized cut algorithm. The approach yields improved performance over other non-parametric non Bayesian approaches.