Search

Abstract: Verification of identity based on the analysis of dynamic signatures is an important problem of biometrics. The effectiveness of the verification significantly increases when the dynamic characteristics of signature (e.g. velocity of the pen) are considered. These characteristics are individual for each user and difficult to imitate. The effectiveness of the verification can be further improved by using partitioning. In this paper we propose a new method which uses partitioning. In our method partitions represent areas of high and low speed of signature and high and low pen׳s pressure. All selected partitions are used by our algorithm, but more important in the classification process are these partitions, in which the signatures of the user acquired during training phase are more stable. Moreover, final classification is interpretable. In this paper we present the simulation results of the proposed method for the two databases: SVC2004 and BioSecure Database. [Copyright &y& Elsevier]

Highlights: [•] We propose a new dynamic signature verification system using vertical partitioning. [•] In our algorithm partitions represent moments of signing process. [•] Our algorithm takes advantage of fuzzy sets and systems theory. [•] Our algorithm allows to extract knowledge about the signature verification. [•] Our algorithm takes into account hierarchy of partitions individually for each user. [Copyright &y& Elsevier]

In this paper, we propose a new class of neuro-fuzzy systems. Moreover, we develop a novel method for reduction of such systems without the deterioration of their accuracy. The reduction algorithm gradually eliminates inputs, rules, antecedents, and the number of discretization points of integrals in the center of area defuzzification method. It then automatically detects and merges similar input and output fuzzy sets. Through computer simulations it is shown that accuracy of the system after reduction and merging has not deteriorated despite the fact that in some cases up to 54% of various parameters and 74% of inputs were eliminated. The reduction algorithm has been tested using well-known classification benchmarks. [ABSTRACT FROM AUTHOR]