Multi-classifier systems for off-line signature verification

摘要

Handwritten signatures are behavioural biometric traits that are known to incorporate a considerable amount of intra-class variability. The Hidden Markov Model (HMM) has been successfully employed in many off-line signature verification (SV) systems due to the sequential nature and variable size of the signature data. In particular, the left-to-right topology of HMMs is well adapted to the dynamic characteristics of occidental handwriting, in which the hand movements are always from left to right. As with most generative classifiers, HMMs require a considerable amount of training data to achieve a high level of generalization performance. Unfortunately, the number of signature samples available to train an off-line SV system is very limited in practice. Moreover, only random forgeries are employed to train the system, which must in turn to discriminate between genuine samples and random, simple and skilled forgeries during operations. These last two forgery types are not available during the training phase.ududThe approaches proposed in this Thesis employ the concept of multi-classifier systems (MCS) based on HMMs to learn signatures at several levels of perception. By extracting a high number of features, a pool of diversified classifiers can be generated using random subspaces, which overcomes the problem of having a limited amount of training data.ududBased on the multi-hypotheses principle, a new approach for combining classifiers in the ROC space is proposed. A technique to repair concavities in ROC curves allows for overcoming the problem of having a limited amount of genuine samples, and, especially, for evaluating performance of biometric systems more accurately. A second important contribution is the proposal of a hybrid generative-discriminative classification architecture. The use of HMMs as feature extractors in the generative stage followed by Support Vector Machines (SVMs) as classifiers in the discriminative stage allows for a better design not only of the genuine class, but also of the impostor class. Moreover, this approach provides a more robust learning than a traditional HMM-based approach when a limited amount of training data is available. The last contribution of this Thesis is the proposal of two new strategies for the dynamic selection (DS) of ensemble of classifiers. Experiments performed with the PUCPR and GPDS signature databases indicate that the proposed DS strategies achieve a higher level of performance in off-line SV than other reference DS and static selection (SS) strategies from literature.