Off-line signature verification using ensembles of local Radon transform-based HMMs

[摘要] ENGLISH ABSTRACT: An off-line signature verification system attempts to authenticate the identityof an individual by examining his/her handwritten signature, after it hasbeen successfully extracted from, for example, a cheque, a debit or credit cardtransaction slip, or any other legal document. The questioned signature is typicallycompared to a model trained from known positive samples, after whichthe system attempts to label said signature as genuine or fraudulent.Classifier fusion is the process of combining individual classifiers, in order toconstruct a single classifier that is more accurate, albeit computationally morecomplex, than its constituent parts. A combined classifier therefore consistsof an ensemble of base classifiers that are combined using a specific fusionstrategy.In this dissertation a novel off-line signature verification system, using amulti-hypothesis approach and classifier fusion, is proposed. Each base classifieris constructed from a hidden Markov model (HMM) that is trained fromfeatures extracted from local regions of the signature (local features), as well asfrom the signature as a whole (global features). To achieve this, each signatureis zoned into a number of overlapping circular retinas, from which said featuresare extracted by implementing the discrete Radon transform. A global retina,that encompasses the entire signature, is also considered.Since the proposed system attempts to detect high-quality (skilled) forgeries,it is unreasonable to assume that samples of these forgeries will be availablefor each new writer (client) enrolled into the system. The system is thereforeconstrained in the sense that only positive training samples, obtainedfrom each writer during enrolment, are available. It is however reasonable toassume that both positive and negative samples are available for a representativesubset of so-called guinea-pig writers (for example, bank employees).These signatures constitute a convenient optimisation set that is used to selectthe most proficient ensemble. A signature, that is claimed to belong toa legitimate client (member of the general public), is therefore rejected or acceptedbased on the majority vote decision of the base classifiers within themost proficient ensemble.When evaluated on a data set containing high-quality imitations, the inclusionof local features, together with classifier combination, significantly increasessystem performance. An equal error rate of 8.6% is achieved, whichcompares favorably to an achieved equal error rate of 12.9% (an improvementof 33.3%) when only global features are considered.Since there is no standard international off-line signature verification dataset available, most systems proposed in the literature are evaluated on datasets that differ from the one employed in this dissertation. A direct comparisonof results is therefore not possible. However, since the proposed systemutilises significantly different features and/or modelling techniques than thoseemployed in the above-mentioned systems, it is very likely that a superior combinedsystem can be obtained by combining the proposed system with any ofthe aforementioned systems. Furthermore, when evaluated on the same dataset, the proposed system is shown to be significantly superior to three othersystems recently proposed in the literature.

[发布机构] Stellenbosch University