Verification of static signature pattern based on random subspace, REP tree and bagging

Abstract

In this paper we propose an effective method for static signature recognition from spontaneous handwritten text images. Our method relies on different aspects of writing: the presence of redundant patterns in the writing and its features. Signatures are analyzed at small fragments in which we seek to extract the patterns that an individual employs frequently as he writes. We exploit different features of writing like orientation, centroid and contour by computing a set of features from writing samples at different levels of observations. Orientation like intersecting point, edge point and gradient change of signature achieve great success in feature description. These features are extracted from the standard signature database and extracted features are trained and tested by machine learning (ML) approach. The machine learning approaches like Bagging, Random subspace (RS) and REP tree are used for classification purpose. Bagging with 10 iterations and base learner achieved efficiency upto 88 %. RS randomly selects features from feature set and creates new feature set. It uses decision tree as base classifier with different tree size. RS achieved efficiency same as Bagging but has more statistical errors. However, in case of REP tree we have achieved efficiency upto 75 %. The experimental results show that the Bagging and RS achieves promising results on publicly available data set.