Improving performance of HMM-based off-line signature verification systems through a multi-hypothesis approach

The neural and statistical classifiers employed in off-line signature verification (SV) systems are often designed from limited and unbalanced training data. In this article, an approach based on the combination of discrete Hidden Markov Models (HMMs) in the ROC space is proposed to improve the performance of these systems. Inspired by the multiple-hypothesis principle, this approach allows the system to select, from a set of different HMMs, the most suitable solution for a given input sample. By training an ensemble of user-specific HMMs with different number of states and different codebook sizes, and then combining these models in the ROC space, it is possible to construct a composite ROC curve that provides a more accurate estimation of system performance. Moreover, in testing mode, the corresponding operating points—which may be selected dynamically according to the risk associated with input samples—can significantly reduce the error rates. Experiments performed by using a real-world off-line SV database, with random, simple and skilled forgeries, indicate that the multi-hypothesis approach can reduce the average error rates by more than 17%, as well as the number of HMM states by 48%.