An Online Signature Identification System Utilizing ReliefF Feature Selection Technique and Multi Class Optimizable SVM Classifier

Automatic signature identification is a trending research area and has importance. So, this research presents an online signature identification method that utilizes "ReliefF" feature selection algorithm and an optimizable multi-class SVM classifier. The objective of our work is to identify a person to give him/her access to a confidential area/system. Feature selection is a critical area. So, to choose the most suitable features describing a particular class is critical. Therefore, we used the "ReliefF" feature selection algorithm to minimize classification error and remove the least relevant features. Moreover, Signature verification still is a challenging research problem because of large intra-class variations and small interclass variations while considering forgeries. Therefore skilled forgery signatures of each user are used here as a separate class to enhance the identification accuracy of the proposed system resulting into multi-class classification problem. Use of K fold cross validation has enhanced the efficiency of machine learning model on unseen data. We used an optimizable multi-class SVM classifier with a Bayesian optimizer to tune the hyper parameters. The use of "ReliefF" selected features has enhanced testing accuracy of the model. The proposed system is trained, validated and tested on the SVC2004 Database (Task 1) for five and all forty users having different types of signature information yielding promising identification results.