Siamese Network for Learning Genuine and Forged Offline Signature Verification

Abstract

This work aims to use a Siamese network to verify between genuine and forged signatures by making signature embeddings more robust. Currently, the Siamese network is most widely used in many applications such as Dimensionality reduction, Learning image descriptor, Face recognition, Image ranking, etc. This network is termed as twin network since it consists of two similar neural networks which take two input images and shares same weights. The critical task in signature verification is to discriminate between genuine and skilled forger since forged signature differs by some precise kind of deformation. Embedding vector is generated by Siamese network and to make embedding vector more robust we propose to add statistical measures to it, which are calculated on the embedding vector itself. The contrastive loss function is then applied on the resultant embedding vector.