Global feature selection for on-line signature verification

A large number of features can be used to represent on-line handwritten signatures in verification tasks. Depending on the signature database and acquisition conditions, some features will not help in separating writers in the feature space so that an appropriate decision boundary will be hard to estimate. Other features will provide good separability between legitimate system users and their forgers. This paper proposes a signature feature selection algorithm combining a modified Fisher ratio cost function and a sub-optimal but fast search method to explore an initial feature space of candidate global features. A large number of candidate feature subsets of various sizes is evaluated, and it is shown that our modified Fisher ratio correlates highly with experimental verification error rates. The need for forgery data in the feature selection phase of signature verification systems development is also investigated, and we postulate that user-to-user separation is a good indication of user-to-forger separation.