Publication
Evaluation of Novel Features and Different Models for Online Signature Verification in a Real-World Scenario
Marcus Liwicki
In: Proceedings of the 14th Conference of the International Graphonomics Society. International Graphonomics Society Conference (IGS-09), September 13-16, Dijon, France, Pages 22-25, 2009.
Abstract
In this paper we describe signature verification experiments on a recently collected dataset which is publicly available. We investigate a novel set of local features and compare it to a reference set often used in the literature. Furthermore we compare the use of Gaussian mixture models (GMMs) and hidden Markov models (HMMs) for classification. We optimize all the standard meta-parameters on a validation set and measure the final performance on a separate test set. The task considered in our experiments is the most challenging in automatic signature verification, i.e., to verify a questioned signature when only one reference signature by the claimed author is given. We found out that the system with the novel feature set outperforms the reference system. Furthermore, HMMs perform better than GMMs if we do not restrict the number of model parameters. Despite the difficulty of the task, we could finally achieve an equal error rate of about 3% without optimizing any meta-parameters on the test set.