Identification of Handwritten Signature Forgery using Deep Learning

Abstract

Recently, deep convolutional neural networks are successfully applied in several fields of computer vision and pattern recognition. Handwritten signature is one of the most important biometrics applied in banking systems, administrative and financial applications. A signature is a representation of a person’s name that is used to prove that person’s identity, but it is frequently falsified. Consequently, it is necessary to confirm a signature’s legitimacy. One of the widely used biometric traits for confirming a person is their signature. In this paper, a tiny 3-layer deep convolutional neural network (CNN) followed by a fully connected layer is suggested to validate the off-line signature. In two separate configurations, this network has been used: first as a feature extractor in a hybrid classifier, and then as an end-to-end classifier in a Siamese network. Support vector machines are used in a hybrid classifier strategy to check the validity of the signature. Two or more identical subnetworks connected by one or more completely connected layers make up a siamese network. The Siamese network employs the suggested neural network as a subnetwork. On three datasets—Signature Set-1, Signature Set-2, and BHSig260—the hybrid classifier and Siamese network are tested for both writer independent (WI) and writer dependent (WD) verification. The verification accuracy for HINDI in WI beats state-of-the-art methods, according to experimental results.