An Improved Siamese Network Model for Handwritten Signature Verification

Abstract

Handwritten signature verification is one of the most prominent and prevalent biometric methods in many real applications. A siamese neural network, which can extract stylistic features of handwriting writers, proves to be efficient in verifying handwritten signature. However, a traditional siamese neural network fails to fully represent an author’s writing style and suffers from low performance when the distribution of positive and negative handwritten samples is extremely unbalanced. To address this issue, this paper proposes an improved siamese network model with two main ideas: a) adopting a two-stage convolutional neural network to verify original and enhanced handwriting images simultaneously, and b) utilizing the Focal loss to handle the extreme imbalance between positive and negative handwritten samples. Experimental results on three challenging signature datasets of different languages demonstrate that compared with state-of-the-art models, the proposed model achieves a higher prediction accuracy.