Nonparametric estimation of a topological invariant from static signature images

Abstract:

Given an image of a signature, its classification as genuine or a forged one has been and remains an important challenge of Forensic Sciences and in particular of forensic documents examiners. The static image contains only information about the color components of the pixels and no information about the speed nor other dynamic components of the writing are available. We present some statistical aspects of the problem and focus on nonparametric estimation of contour components of the signatures. To avoid bias resulting from changing the scale, rotation or translation of the considered shapes we use a topological invariant for curves: the curvature presented in the normalized arc length coordinates. We explore the power of nonparametric curve estimation in detecting the forgery and test it on an unusual sample. In the sample a number of 'forged' signatures are available coming from an adept penman mimicking a signature of another man. We compare results obtained by applying the topological invariant and the Geenens statistics.