Meta-heuristic optimization methods on SPD Manifolds and applications to computer vision

Abstract

Verifying whether or not a questioned handwritten signature belongs to a claimed identity is a quite challenging task, usually requiring careful examination, skills and knowledge. An accurate and efficient signature verification system, can provide considerable assistance by reducing the time it takes to complete a manual task, usually performed by Forensic Handwriting Examiners. Towards this purpose, this thesis proposes an offline handwritten signature verification framework in the space of Symmetric Positive Definite matrices, by utilizing a metric learning approach which enables the use of a special purpose neural network, acting as a meta-heuristic optimization procedure. Specifically, the static signature images are mapped into points of the Symmetric Positive Definite Riemannian manifold and then, a metric learning approach is applied in order to learn a Mahalanobis based distance which will minimize the distance of similar samples and maximize the distance of dissimilar ones. The experimental results with the use of two popular signature datasets of western origin, shows that the proposed framework is comparable at least to State-of-the-Art models, typically realized under a framework of Euclidean nature.