Multiple Classifier System for Writer Independent Offline Handwritten Signature Verification using Hybrid Features

Abstract

Offline handwritten signature verification is a very challenging area of research as the handwriting of two people may bear similarity whereas handwriting of a person may vary at different times. The accuracy of handwritten signature verification system depends on the classifier system and the way of feature extraction. Keeping this point of view, four types of hybrid feature sets and three types of classifiers specifically support vector machine with polynomial kernel, support vector machine with quadratic kernel and decision tree are investigated for writer-independent offline handwritten signature verification in the present work. To obtain hybrid feature sets, local oriented statistical information booster, discrete wavelet transform, and histogram of oriented gradient feature descriptors are extracted and are coupled with each other. To create multiple classifier system, the training set is partitioned into subsets and these training subsets are used to train the classifiers of multiple classifier system using same training algorithm for all classifiers. The performance analysis is carried out using two scenarios. In the first scenario, genuine and random forgery signatures are used to train the classifiers whereas genuine, random, unskilled and simulated forgery signatures are used to train the classifiers in the second scenario. False rejection rate 8.00 and false acceptance rate 0.00 for all types of forgeries are reported as the best result of the experiments.