Writer-Independent Offline Signature Verification Using Deep Siamese Neural Network

Abstract

Several researchers have worked on signature verification problems from different aspects utilizing insights from signal processing and computer vision, since the last few decades. Despite the advancement in technology, signature model building with an appropriate classifier to distinguish between skilled forgeries and genuine is still a critical problem. This paper presents Siamese neural network-based signature verification system which consists of twin convolutional neural networks with shared weights which maximize the distance between dissimilar pairs while simultaneously minimizing the distance between similar pairs. The signatures are paired among similar (genuine, genuine) and dissimilar (genuine, forged) pairs. This research achieves higher classification accuracy compared to state-of-the-art methods. The results are validated on our created dataset and CEDAR, UTSig, BHSig260, GDPS300, GDPS synthetic signature benchmark datasets.