On-Line Signature Verification and Feature Extraction
Rakesh Deivachilai · Digital Repository at the University of Maryland (University of Maryland College Park) · 2015
This thesis considers the possibility of discretizing visually acquired signatures to represent them as an unordered collection of words and then use Sequential Minimal Optimization (SMO) for training Support Vector Machine (SVM) classifiers to classify the signatures as either legitimate or forged. Signature discretization is done using Symbolic Aggregate Approximation (SAX). SAX reduces the dimensions of the signature and produces a list of SAX words that are then represented as a Bag-of-Patterns (BoP) model for classification purposes. Each SAX word will act as a separate feature and can be evaluated by using techniques like Chi-squared for informativeness. This will lead us to the set of SAX words that were used by the classifier to make the decision. These SAX words can then be mapped to their corresponding original signature coordinates. The approach was tested on a dataset provided by California Institute of Technology that contains 3960 signatures of 106 subjects distributed across two sets. This dataset includes a set of original signature samples as well as forged signature samples for each subject. The classification accuracy achieved is promising and the extracted features provide insight into the unique patterns present in each of the subject's signature.