Incorporating user individuality into writer-independent on-line signature verification

In this paper we present a novel writer-independent on-line signature verification system. The proposed system obtains a universal signature representation in dissimilarity space, but remains able to compensate for the personal variability of an individual's handwriting by subsequently employing a writer-specific dissimilarity normalisation strategy. Signature modelling is achieved by utilising either quadratic discriminant analysis or a support vector machine with a radial basis function kernel. We also present an experimental comparison between the aforementioned dissimilarity normalisation strategy and an existing dissimilarity individualisation technique, that is the so-called user-specific global-parameter fusion protocol. The feasibility of incorporating either of these techniques into an arbitrary signature modelling framework, as well as the resulting impact on system proficiency, is investigated. When evaluated on 4530 samples collected from 51 writers in the Philips signature database, the most proficient system proposed in this paper achieves equal error rates of 0.65%-1.19% when 5-15 reference signatures are available per writer. We show that the aforementioned system outperforms previous systems that were evaluated on the same data set. We also show that, when evaluated under identical experimental conditions, the dissimilarity normalisation strategy outperforms the fusion protocol in terms of system proficiency. Furthermore, we assert that, unlike the fusion protocol, the dissimilarity normalisation strategy is suitable for incorporation into any existing writer-independent on-line signature verification system, regardless of the feature extraction and/or classification technique(s) utilised.