SigAuthUrdu: Signature Verification for Urdu Language

In this paper, we introduce SigAuthUrdu, a convolutional Siamese network designed to address offline writer-independent signature verification for the Urdu language. Siamese networks, consisting of twin networks with shared weights, are trained to map signatures into a feature space where dissimilar signatures are farther apart and similar signatures are closer together. This is achieved by presenting the network with pairs of similar and different signatures and optimizing the Euclidean distance between them. Our experiments on an Urdu dataset demonstrate the network's capability to accurately detect forgeries across various scripts and handwriting styles using dissimilarities. Additionally, we created and utilized our own dataset, 98UrduSignatureCorpus, for this purpose, as no such dataset previously existed. This dataset is a significant contribution, facilitating the development and evaluation of our Siamese network model for Urdu signature verification.