A Two-Stage Siamese Network Model for Offline Handwritten Signature Verification

Offline handwritten signature verification is one of the most prevalent and prominent biometric methods in many application fields. Siamese neural network, which can extract compare writers’ style features, proves to be efficient verifying offline signature. However, traditional network fails represent writing fully suffers from low performance when distribution positive negative samples unbalanced. To address this issue, study proposes a two-stage model for accurate with two main ideas: (a) adopting verify original enhanced signatures simultaneously, (b) utilizing Focal Loss deal extreme imbalance between signatures. Experimental results on four challenging datasets different languages demonstrate that compared state-of-the-art models, our proposed achieves better performance. Furthermore, tries extend Chinese dataset real environment, significant attempt field identification.