Abstract— this paper proposes two systems for offline signature verification based on a global and on a local approach respectively. The features used consist of different kinds of geometrical, statistical and structural features. For comparison purposes, we used two baseline systems (global and local), both based on a larger number of features encoding the orientations of the strokes using mathematical morphology. Experiments are performed on two offline signature databases, namely DS2-50 and GPDS-104. The obtained results show that we may obtain similar performances even when using a much smaller but more discriminant set of features and that stability of the performance across different databases can be a real challenge. I. INTRODUCTION ANDWRITTEN signature verification is one of the most important modalities in biometrics. This can be explained by the use of signatures as an official mean to verify the identity of the authors of social and legal documents such as checks, credit cards, contracts, certificates… A handwritten signature depends on the physical and psychological state of the signer as well as on the acquisition device and conditions. Thus, the signature acquired from a person is susceptible to changes leading to high intra-class variability. This variability makes signature verification a difficult discrimination problem. Depending on the acquisition process, automatic signature verification systems can be classified into two categories: 1) on-line signature verification [1, 2] where signature is captured during the writing process, and which makes available dynamic information like writing speed and pressure as well as static information, and 2) offline signature verification [2, 3, 4, 5,6,7,8,9,10] where the static image of a signature is captured once the writing process is over, so only the signature geometry is available. Robust offline systems are, therefore, more difficult to design. A recent review of offline and online signature verification approaches is proposed in [2]. Although offline signature verification systems are less accurate than online verifications systems, they are still important owing to the reasons mentioned above. This paper deals with offline signature verification, and our aim i is to study different approaches for a better discrimination between genuine signatures and skilled forgeries. Offline verification systems can be classified into Global [2,3;4,5,6,7] and Local [2;3;7,8] systems. The first are based on global feature extraction, which describe the signature as a whole. The latter are based on local feature extraction, which represent the signature as a sequence of feature vectors or observations by an appropriate segmentation or scanning (windowing) along a specific direction. Global systems are generally fast but have lower performance than local systems since the order information (order of features) is not taken into account [2]. Moreover, the choice of features is very important for the system to correctly discriminate between authentic and forged signatures. It is worth noting that a good level of performance for a verification system does not depend only on the number of features but also on the discriminating power of these features and on the signature image quality. The aim of this work is, on one hand, to study the influence of the number and nature of features on performance, and, on the other hand, to assess the stability of performance across different signature databases. To this end, we propose two systems for offline signature verification: the first one is based on a global approach while the second one is based on a local approach. The features used by the global approach consist of a set of geometrical, statistical and structural features while the local approach employs directional and curvature features after an explicit segmentation of the hand-drawn signature into strokes each with a roughly uniform direction. For comparison purposes, we design two baseline systems (global and local) inspired by [3], which will serve as a benchmark. Both baseline systems are based on a larger number of features encoding the orientations of the strokes using mathematical morphology. Experiments are performed on two offline signature databases, namely the offline BioSecure DS2 database containing data of 50 persons, and the GPDS database containing data of 104 persons. The results obtained show that we may obtain similar performances even when using a much smaller but more discriminant set of features and that stability of the performance across different databases can be a real challenge. This paper is organized as follows: in Section 2, we describe the two databases that we use (DS2-50 and GPDS- 104). Section 3 describes the pre-processing phase. In Section 4, we present the two baseline systems based on morphological features, and the two proposed systems based on geometrical, statistical and structural features. The experimental framework including results and analysis is H A Comparison of Feature Extraction Approaches for Offline Signature Verification Y. Rekik, N. Houmani, M.A. El Yacoubi, S. Garcia-Salicetti, and B. Dorizzi. Intermedia, Dept. EPH Institut Telecom; Telecom SudParis Evry, France 978-1-61284-732-0/11/$26.00 ©2010 IEEE