Learning discriminative representations by a Canonical Correlation Analysis-based Siamese Network for offline signature verification

Abstract

In offline signature verification tasks, capturing different writing behaviors between genuine and forged signatures is a crucial and challenging step. In this paper, a novel writer independent Canonical Correlation Analysis-based Siamese Network (CCASigNet) is proposed to learn discriminative representations between different signature pairs. Specifically, we first construct signature pairs with three types: genuine-genuine, genuine-forged, and forged-forged. Then, different signature pairs are fed into CCASigNet for training with the Canonical Correlation Analysis (CCA) and classification-based losses. After network training, we extract the feature of signatures by CCASigNet and use writer-dependent classifiers to construct a comprehensive verification system. Extensive experiments on four benchmark signature datasets demonstrate that the proposed CCASigNet learns discriminative representations between different signature pairs and achieves state-of-the-art or competitive performance compared with advanced verification systems. In addition, the proposed CCASigNet has good generalization ability and is easy to transfer to different datasets with different language scripts within the realm of offline signature verification tasks.