Wet-Ink Signature Forgery Detection Using Siamese ResNeXt

Abstract

In cases of wet-signed documents, signature forgery refers to the unauthorized imitation of an individual's signature by another person without their consent, posing a serious threat to the reliability of important documents. In this study, Siamese networks have been employed as a solution to address the issue of signature forgery in wet-signed documents. Siamese networks prove to be effective tools in detecting signature forgery due to their capability to compare signature pairs belonging to similar and different classes during the learning process. As inputs to the Siamese networks, not only signature images but also the frequency spectra and Histogram of Oriented Gradients (HOG) features of these images have been utilized. Feature vectors generated by Siamese ResNeXt-50 are combined, and linear layers are employed for classification. Binary Cross Entropy Loss has been chosen as the loss function, in accordance with the binary classification problem. The results indicate the successful application of Siamese networks in detecting signature forgery, emphasizing their potential to offer a reliable and effective solution to counter signature forgery.