Profound impact of artificial neural networks and Gaussian SVM kernel on distinctive feature set for offline signature verification

Abstract

Signature (Latin - signare) is a handwritten stylized form of identification of its owner. Often handwritten signatures are generally used in secured identity preservation. An ideal signature recognition system handles image noise as well as that of learning unique patterns in an individual's signature. This paper analyzes the performance of artificial neural network (ANN) architectures and Gaussian support vector machine (SVM) kernel for offline signature recognition scheme that is trained on a distinct feature set extracted from signature images. We investigated the impact of using ANN and SVM on specialized feature set and present comparative analysis of the two. Three distinct features - gradient histogram, dot density and slices were used - yielding testing accuracies of 93.1%, 98% and 85.1% respectively. Using ANN and SVM on this set, a maximum accuracy of 96.57% was achieved over a group of 30 individuals, covering an entire data set of 3000 signatures.