Analysis of Intra-Person Variability of Features for Off-line Signature Verification BENCE KOVARI, HASSAN CHARAF Department of Automation and Applied Informatics Budapest University of Technology and Economics 1111. Budapest, Magyar Tudosok Korutja 2, Q HUNGARY bence.kovari@aut.bme.hu http://www.aut.bme.hu/signature Abstract: - One of the major challenges in off-line signature verification is the fact that a person’s own signature is influenced by a number of external and internal factors. This influence results in a high variability even between signatures written by the same signer. This paper proposes a method which is able to model the intra-person variability of a signature feature and also to identify and eliminate the effects of external factors. To demonstrate the efficiency of the algorithm, a sample signature verifier is constructed and evaluated on the Signature Verification Competition 2004 database. Experiments have shown that by using 3 features (endings, loops and skew vectors) an average error rate of 12% can be achieved by the system. These results may be further improved by increasing the number of features, used during the comparison of signatures. Key-Words: - signature verification; off-line; classification, normal distribution 1 Introduction In the past century several studies [1][2][3][4][5] have confirmed that signatures can be used with a high success rates for biometrical identification. There are several methodological guides like [6] which formalize the process of verification. However, as all human experts, even opinions of forensic document examiners (FDE’s) are subjective and prone to human errors. Also considering the huge numbers of signed documents created each day, and the limited number of FDE’s it is obvious why automated signature verification has been in a focus of researchers for the past few decades. Computer based signature verification can be divided into two main approaches, the on-line and the off-line approach. In online signature recognition the whole process of signing is captured using some kind of acquisition device (camera, digital tablet etc.), then analyzed and used to make a decision. The aim of off- line signature verification is to decide, whether a signature originates from a given signer merely based on the scanned image of the signature and a few images of the original signatures of the signer. Unlike on-line signature verification, which requires special acquisition hardware and setup, off-line signature verification can be performed independently from the normal signing process, and is thereby less intrusive and more user friendly. On the other hand, important information like velocity, pressure and the difference between up- and down strokes is partially lost. When evaluating verification approaches we also have to differentiate between them based on the signature database used and the way it was used. A typical signature database is a collection of signatures from several signers, containing some (10-20) original signatures from a given signer and usually also containing several forged signatures (forgeries) for the same signer. We focus on the scenario, where verification systems are trained only using original signatures and tested against both original signatures and skilled forgeries and the verification is performed off- line, as this approach suits the most real world scenarios. The performance of signature recognition systems is usually measured in terms of equal error rate (EER), which is the point where Type I and Type II errors are equal. One also has to take into consideration, that (although usually created on a lower level) signatures are the results of conscious behavior and can thereby be influenced by a huge amount of factors [1]. In the lack of a common signature corpus and a well defined evaluation methodology, the results of different studies may only be hardly comparable; therefore, the values mentioned later should be only taken as approximations. As of today, when tested against skilled forgeries, even the best off-line verification systems deliver worse or equal error rates than 5-10% [7] [8], in contrast with a human expert, who is able to do the distinction with an error rate of 1% [8]. In the past decade a bunch of solutions (like [9] or [10]) have been introduced, to overcome the limitations of off-line signature verification and to compensate for the loss of accuracy compared to on-line systems. To break the 5% barrier it is essential to identify, understand and compensate for the different sources of error in the algorithms. This paper presents a solution to address the problem of improvement and thereby possibly break the 5% barrier. WSEAS TRANSACTIONS on COMPUTERS Bence Kovari, Hassan Charaf ISSN: 1109-2750 1359 Issue 11, Volume 9, November 2010