Multiple Classifier System for Writer Independent Offline Signature Verification using Genetic Algorithm and SVM

Abstract

An efficient multiple classifier system for writer independent offline handwritten signature verification is proposed. Local oriented statistical information booster and histogram of oriented gradients feature descriptors are extracted from the signature image and a genetic algorithm is used to reduce the dimension of feature descriptors. To create multiple classifier system, two scenarios are generated. In scenario I, training set is divided into subsets using k-fold cross validation method and these training subsets are used to train the classifiers of multiple classifier system using same training algorithm for all classifiers. In scenario II, the classifiers of multiple classifier system are trained with multiple training algorithms. Unskilled and skilled forgeries are used to test the performance of multiple classifier systems. The performance analysis is carried out using support vector machine with polynomial kernel, quadratic kernel and gaussian radial basis function kernel. For the study, the signature database of 160 writers is used. The experiments of scenario I and scenario II reveal average error rate 2.08 and 2.50, respectively.