A Comparative Analysis of Machine Learning Algorithms for Online Signature Recognition

Abstract

Biometrics recognition plays a vital role in modern human recognition and verification systems. An extensive latest research by the research community has rendered the field of biometrics inevitable for real-life applications. This research study focuses on online signature recognition. The research study is performed to identify if an online signature is genuine or forged. A novel online signature dataset, based on 1000 online signatures, has been collected from 200 participants, wherein every participant provided 5 instances of the online signature. An Android-based mobile application was developed to collect the online signature data. Moreover, a data augmentation technique was used to increase the training samples of the online signature dataset. Some common features such as the width and height of the signature, x and y coordinate values, pressure, pen ups and pen downs, total duration of the signature, etc. were extracted. The dataset has been trained and tested using machine-learning techniques. The performance of the five existing classifiers on the newly collected database has been compared. The classifiers used for training and testing included a Support Vector Machine (SVM), a Random Forest Classifier (RFC), a variant of RFC called an Extra Tree Classifier (ETC), a Decision Tree Classifier, and K-Nearest Neighbors. The performance of each classifier was evaluated in terms of precision score, recall score, and f-1 score. The RFC, and ETC classifiers gave an overall classification accuracy of 96%.