In this paper, a novel Bayesian network representation is proposed for signature verification. It is different from the traditional Bayesian network, in that its nodes are divided into two classes: common hypotheses and alternative hypotheses, to ensure the constructed network is a tree structure. The network not only captures the conditional probability associated with each node, but also the topological relations among components associated with the network nodes, so that the uncertainty in structure description and the dependencies among components are encoded. Results based on eight persons' signatures indicate that the method offers a considerable improvement in performance over some other popular techniques for signature verification. Since the uncertainty and interaction of pattern components are common phenomena in pattern representation and matching, the approach may also be enlightening for other pattern classification problems. (C) 2002 Pattern Recognition Society. Published by Elsevier Science Ltd. All rights reserved. [References: 22]
展开▼
