Attention Based Multiple Siamese Network for Offline Signature Verification

Abstract

Offline handwritten signatures play an important role in biometrics and document forensics, and it has been widely used in the fields of finance, judiciary and commerce. However, the skilled signature forgeries bring challenges and difficulties to personal privacy protection. Thus it is vital to discover micro but critical details between genuine signatures and corresponding skilled forgeries in signature verification tasks. In this paper, we propose an attention based Multiple Siamese Network (MSN) to extract discriminative information from offline handwritten signatures. MSN receives the reference and query signature images and their corresponding inverse images. The received images are fed to four parallel branches. We develop an effective attention module to transfer the information from original branches to inverse branches, which attempts to explore prominent features of handwriting. The weight-shared branches are concatenated in a particular way and formed into four contrastive pairs, which contribute to learn useful representations by comparisons of these branches. The preliminary decisions are generated from each contrastive pair independently. Then, the final verification result is voted from these preliminary decisions. In order to evaluate the effectiveness of proposed method, we conduct experiments on three publicly available signature datasets: CEDAR, BHSig-B and BHSig-H. The experimental results demonstrate the proposed method outperforms that of other previous approaches.