Automatic Handwritten Signature Verification – Which features should be looked at?

The increasing need for personal authentication in many daily applications has made biometrics a fundamental research area. In particular, handwritten signatures have long been considered one of the most valuable biometric traits. Signatures are the most popular method for identity verification all over the world, and people are familiar with the use of signatures for identity verification purposes in their everyday life. In fact, signatures are widely used in several daily transactions, being recognized as a legal means of verifying an individual?s identity by financial and administrative institutions. In addition, signature verification has the advantage of being a non-invasive biometric technique. Two categories of signature verification systems can be distinguished taking into account the acquisition device, namely, offline systems, where only the static image of the signature is available, and online systems, where dynamic information acquired during the signing process, such as pen coordinates, pen pressure and pen inclination angles, is available. In this Thesis both, the offline and the online modalities, are addressed. For the offline signature verification case, a new feature extraction approach based on a circular grid scheme is proposed. Graphometric features are adapted to be extracted resorting to this new grid geometry. In addition, the property of rotation invariance of the Fast Fourier Transform (FFT) is used in order to achieve robustness against rotation of the signatures, which is an important issue for offline signature verification. For the case of online signature verification, contributions towards feature selection and extraction are done. In addition, the designed systems are tested on a challenging recently available public database containing signatures from different cultural origins, namely Western (Dutch) and Chinese signatures. Two different approximations based on orthogonal series expansions of the time functions associated to the signing process are proposed for online signature feature extraction. One of them is based on Legendre polynomials, and the other one is based on wavelet decomposition. The coefficients in these orthogonal series expansions of the time functions are used as features to model the signatures. In addition, an in depth analysis of different combinations of several time functions is carried out in order to provide some insight on their actual discriminative power. Moreover, a novel consistency factor is proposed in order to quantify this discriminative power. On a subsequent step towards improving the performance of the online signature verification systems, a pre-classification stage based on global features is incorporated to the system. The idea is to quickly recognize and discard gross forgeries based on the pre-classification approach. It is expected that incorporating the pre-classification step would speed up and simplify the verification process. To bridge the gap between Forensic Handwriting Expert (FHE) and Pattern Recognition (PR) communities, is currently one of the most important challenges in the field. In an attempt to make some contribution to this issue, the discriminative power of a set of features which seems to be relevant to signature analysis by FHEs is analysed and particularly compared to the discriminative power of automatically selected feature sets. This analysis is intended to help FHEs to further understand the signatures and the writer behaviour. Even the feasibility of developing a system which could complement the FHEs work is evaluated. For this purpose, a fusion between automatic selected features and FHE based features is proposed. To conclude this analysis, the feasibility of using only FHE based features for automatic online signature verification is evaluated. Finally, an attempt to give an answer to the question of whether is it possible to combine all the proposed online features in order to achieve better verification results, is provided. The obtained experimental results are technically sound, since the verification performances are comparable and, in some cases, even better than the ones in the state-of-the-art. In addition, interesting observations can be made on the basis of the analysis and discussions carried out which can make contributions towards some of the most important actual challenges in the field.